Merge pull request #2449 from mrry/branch_122885116

Branch 122885116

tensorflow/BUILD

@@ -77,7 +77,7 @@ filegroup(
         "//tensorflow/contrib/ctc:all_files",
         "//tensorflow/contrib/distributions:all_files",
         "//tensorflow/contrib/ffmpeg:all_files",
-        "//tensorflow/contrib/ffmpeg/kernels:all_files",
+        "//tensorflow/contrib/ffmpeg/default:all_files",
         "//tensorflow/contrib/framework:all_files",
         "//tensorflow/contrib/layers:all_files",
         "//tensorflow/contrib/layers/kernels:all_files",

tensorflow/contrib/distributions/python/kernel_tests/exponential_test.py

@@ -80,6 +80,49 @@ class ExponentialTest(tf.test.TestCase):
       self.assertEqual(exponential.entropy().get_shape(), (3,))
       self.assertAllClose(exponential.entropy().eval(), expected_entropy)
 
+  def testExponentialSample(self):
+    with self.test_session():
+      lam = tf.constant([3.0, 4.0])
+      lam_v = [3.0, 4.0]
+      n = tf.constant(100000)
+      exponential = tf.contrib.distributions.Exponential(lam=lam)
+
+      samples = exponential.sample(n, seed=137)
+      sample_values = samples.eval()
+      self.assertEqual(sample_values.shape, (100000, 2))
+      self.assertFalse(np.any(sample_values < 0.0))
+      for i in range(2):
+        self.assertLess(
+            stats.kstest(
+                sample_values[:, i], stats.expon(scale=1.0/lam_v[i]).cdf)[0],
+            0.01)
+
+  def testExponentialSampleMultiDimensional(self):
+    with self.test_session():
+      batch_size = 2
+      lam_v = [3.0, 22.0]
+      lam = tf.constant([lam_v] * batch_size)
+
+      exponential = tf.contrib.distributions.Exponential(lam=lam)
+
+      n_v = 100000
+      n = tf.constant(n_v)
+      samples = exponential.sample(n, seed=138)
+      self.assertEqual(samples.get_shape(), (n_v, batch_size, 2))
+
+      sample_values = samples.eval()
+
+      self.assertFalse(np.any(sample_values < 0.0))
+      for i in range(2):
+        self.assertLess(
+            stats.kstest(
+                sample_values[:, 0, i], stats.expon(scale=1.0/lam_v[i]).cdf)[0],
+            0.01)
+        self.assertLess(
+            stats.kstest(
+                sample_values[:, 1, i], stats.expon(scale=1.0/lam_v[i]).cdf)[0],
+            0.01)
+
 
 if __name__ == '__main__':
   tf.test.main()

tensorflow/contrib/distributions/python/ops/distribution.py

@@ -200,15 +200,26 @@ class ContinuousDistribution(BaseDistribution):
   """Base class for continuous probability distributions.
 
   `ContinuousDistribution` defines the API for the likelihood functions `pdf`
-  and `log_pdf` of continuous probability distributions.
+  and `log_pdf` of continuous probability distributions, and a property
+  `is_reparameterized` (returning `True` or `False`) which describes
+  whether the samples of this distribution are calculated in a differentiable
+  way from a non-parameterized distribution.  For example, the `Normal`
+  distribution with parameters `mu` and `sigma` is reparameterized as
 
-  Subclasses must override both `pdf` and `log_pdf` but one can call this base
-  class's implementation.
+  ```Normal(mu, sigma) = sigma * Normal(0, 1) + mu```
+
+  Subclasses must override `pdf` and `log_pdf` but one can call this base
+  class's implementation.  They must also override the `is_reparameterized`
+  property.
 
   See `BaseDistribution` for more information on the API for probability
   distributions.
   """
 
+  @abc.abstractproperty
+  def is_reparameterized(self):
+    pass
+
   @abc.abstractmethod
   def pdf(self, value, name="pdf"):
     """Probability density function."""

tensorflow/contrib/distributions/python/ops/exponential.py

@@ -20,7 +20,11 @@ from __future__ import print_function
 
 from tensorflow.contrib.distributions.python.ops import gamma
 from tensorflow.python.framework import ops
+from tensorflow.python.framework import tensor_shape
+from tensorflow.python.framework import tensor_util
+from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import random_ops
 
 
 class Exponential(gamma.Gamma):
@@ -45,3 +49,35 @@ class Exponential(gamma.Gamma):
   @property
   def lam(self):
     return self._lam
+
+  @property
+  def is_reparameterized(self):
+    # While the Gamma distribution is not reparameterizeable, the
+    # exponential distribution is.
+    return True
+
+  def sample(self, n, seed=None, name=None):
+    """Sample `n` observations from the Exponential Distributions.
+
+    Args:
+      n: `Scalar`, type int32, the number of observations to sample.
+      seed: Python integer, the random seed.
+      name: The name to give this op.
+
+    Returns:
+      samples: `[n, ...]`, a `Tensor` of `n` samples for each
+        of the distributions determined by the hyperparameters.
+    """
+    broadcast_shape = self._lam.get_shape()
+    with ops.op_scope([self.lam, n], name, "ExponentialSample"):
+      shape = array_ops.concat(
+          0, [array_ops.pack([n]), array_ops.shape(self._lam)])
+      sampled = random_ops.random_uniform(
+          shape, maxval=math_ops.cast(1.0, dtype=self.dtype),
+          dtype=self.dtype)
+
+      n_val = tensor_util.constant_value(n)
+      final_shape = tensor_shape.vector(n_val).concatenate(broadcast_shape)
+      sampled.set_shape(final_shape)
+
+      return -math_ops.log(sampled) / self._lam

tensorflow/contrib/distributions/python/ops/gamma.py

@@ -206,3 +206,7 @@ class Gamma(ContinuousDistribution):
         beta = self._beta
         return (alpha - math_ops.log(beta) + math_ops.lgamma(alpha) +
                 (1 - alpha) * math_ops.digamma(alpha))
+
+  @property
+  def is_reparameterized(self):
+    return False

tensorflow/contrib/distributions/python/ops/gaussian.py

@@ -229,3 +229,7 @@ class Gaussian(object):
       sampled.set_shape(final_shape)
 
       return sampled * self._sigma + self._mu
+
+  @property
+  def is_reparameterized(self):
+    return True

tensorflow/contrib/distributions/python/ops/mvn.py

@@ -454,3 +454,7 @@ class MultivariateNormal(object):
       samples.set_shape(final_shape)
 
       return samples
+
+  @property
+  def is_reparameterized(self):
+    return True

tensorflow/contrib/distributions/python/ops/student_t.py

@@ -277,6 +277,10 @@ class StudentT(ContinuousDistribution):
 
         return samples * self._sigma + self._mu
 
+  @property
+  def is_reparameterized(self):
+    return True
+
   def _ones(self):
     return array_ops.ones_like(self._df + self._mu + self._sigma)
 

tensorflow/contrib/distributions/python/ops/uniform.py

@@ -231,6 +231,10 @@ class Uniform(ContinuousDistribution):
     """`b - a`."""
     return self.b - self.a
 
+  @property
+  def is_reparameterized(self):
+    return True
+
   # TODO(rsepassi): Find a more efficient way of doing the broadcasting in_ones
   # and _zeros.
   def _ones(self):

tensorflow/contrib/ffmpeg/BUILD

@@ -24,7 +24,7 @@ cc_library(
     linkstatic = 1,
     visibility = ["//visibility:private"],
     deps = [
-        "//tensorflow/contrib/ffmpeg/kernels:ffmpeg_lib",
+        "//tensorflow/contrib/ffmpeg/default:ffmpeg_lib",
         "//tensorflow/core:framework_headers_lib",
         "//third_party/eigen3",
     ],
@@ -38,7 +38,7 @@ cc_library(
     linkstatic = 1,
     visibility = ["//visibility:private"],
     deps = [
-        "//tensorflow/contrib/ffmpeg/kernels:ffmpeg_lib",
+        "//tensorflow/contrib/ffmpeg/default:ffmpeg_lib",
         "//tensorflow/core:framework_headers_lib",
         "//third_party/eigen3",
     ],

tensorflow/contrib/ffmpeg/decode_audio_op.cc

@@ -18,7 +18,7 @@
 #include <cstdio>
 #include <set>
 
-#include "tensorflow/contrib/ffmpeg/kernels/ffmpeg_lib.h"
+#include "tensorflow/contrib/ffmpeg/default/ffmpeg_lib.h"
 #include "tensorflow/core/framework/op.h"
 #include "tensorflow/core/framework/op_kernel.h"
 #include "tensorflow/core/lib/io/path.h"

tensorflow/contrib/ffmpeg/default/BUILD

@@ -25,7 +25,6 @@ cc_test(
         "--should_ffmpeg_be_installed=true",
     ],
     data = [
-        ":testdata/test_sound1.mp3",
         "//tensorflow/contrib/ffmpeg:test_data",
     ],
     tags = [
@@ -46,9 +45,6 @@ cc_test(
     args = [
         "--should_ffmpeg_be_installed=false",
     ],
-    data = [
-        ":testdata/test_sound1.mp3",
-    ],
     tags = [
         "local",
         "manual",

tensorflow/contrib/ffmpeg/default/ffmpeg_lib.cc

@@ -13,15 +13,13 @@
 // limitations under the License.
 // =============================================================================
 
-#include "tensorflow/contrib/ffmpeg/kernels/ffmpeg_lib.h"
+#include "tensorflow/contrib/ffmpeg/default/ffmpeg_lib.h"
 
 #include <errno.h>
 #include <stdlib.h>
-#include <string>
 #include <sys/stat.h>
 #include <sys/types.h>
 #include <sys/wait.h>
-#include <tuple>
 #include <unistd.h>
 
 #include <vector>

tensorflow/contrib/ffmpeg/default/ffmpeg_lib.h

@@ -13,8 +13,8 @@
 // limitations under the License.
 // =============================================================================
 
-#ifndef THIRD_PARTY_TENSORFLOW_CONTRIB_FFMPEG_KERNELS_FFMPEG_LIB_H_
-#define THIRD_PARTY_TENSORFLOW_CONTRIB_FFMPEG_KERNELS_FFMPEG_LIB_H_
+#ifndef THIRD_PARTY_TENSORFLOW_CONTRIB_FFMPEG_DEFAULT_FFMPEG_LIB_H_
+#define THIRD_PARTY_TENSORFLOW_CONTRIB_FFMPEG_DEFAULT_FFMPEG_LIB_H_
 
 #include <string>
 
@@ -47,4 +47,4 @@ Status CreateAudioFile(const string& audio_format_id, int32 samples_per_second,
 }  // namespace ffmpeg
 }  // namespace tensorflow
 
-#endif  // THIRD_PARTY_TENSORFLOW_CONTRIB_FFMPEG_KERNELS_FFMPEG_LIB_H_
+#endif  // THIRD_PARTY_TENSORFLOW_CONTRIB_FFMPEG_DEFAULT_FFMPEG_LIB_H_

tensorflow/contrib/ffmpeg/default/ffmpeg_lib_test.cc

@@ -13,7 +13,7 @@
 // limitations under the License.
 // =============================================================================
 
-#include "tensorflow/contrib/ffmpeg/kernels/ffmpeg_lib.h"
+#include "tensorflow/contrib/ffmpeg/default/ffmpeg_lib.h"
 
 #include <stdlib.h>
 #include <vector>
@@ -35,7 +35,7 @@ namespace {
 const char kTestWavFilename[] =
     "contrib/ffmpeg/testdata/mono_10khz.wav";
 const char kTestMp3Filename[] =
-    "contrib/ffmpeg/kernels/testdata/test_sound1.mp3";
+    "contrib/ffmpeg/testdata/test_sound1.mp3";
 
 // Set to true via a command line flag iff the test is expected to have FFmpeg
 // installed.

tensorflow/contrib/ffmpeg/encode_audio_op.cc

@@ -15,7 +15,7 @@
 
 #include <limits>
 
-#include "tensorflow/contrib/ffmpeg/kernels/ffmpeg_lib.h"
+#include "tensorflow/contrib/ffmpeg/default/ffmpeg_lib.h"
 #include "tensorflow/core/framework/op.h"
 #include "tensorflow/core/framework/op_kernel.h"
 

tensorflow/contrib/layers/python/layers/summaries.py

@@ -117,8 +117,9 @@ def summarize_tensor(tensor, tag=None):
   Returns:
     The summary op created or None for string tensors.
   """
-  # Skips string tensors.
-  if tensor.dtype.is_compatible_with(dtypes.string):
+  # Skips string tensors and boolean tensors (not handled by the summaries).
+  if (tensor.dtype.is_compatible_with(dtypes.string) or
+      tensor.dtype.base_dtype == dtypes.bool):
     return None
 
   if tensor.get_shape().ndims == 0:

tensorflow/contrib/learn/BUILD

@@ -61,6 +61,20 @@ py_test(
     ],
 )
 
+py_test(
+    name = "test_feeding_queue_runner",
+    size = "small",
+    srcs = [
+        "python/learn/tests/dataframe/test_feeding_queue_runner.py",
+    ],
+    srcs_version = "PY2AND3",
+    deps = [
+        ":learn",
+        "//tensorflow:tensorflow_py",
+        "//tensorflow/python:framework_test_lib",
+    ],
+)
+
 py_test(
     name = "test_early_stopping",
     size = "medium",
@@ -108,6 +122,42 @@ py_test(
     ],
 )
 
+py_test(
+    name = "dnn_linear_combined_test",
+    size = "small",
+    srcs = ["python/learn/estimators/dnn_linear_combined_test.py"],
+    srcs_version = "PY2AND3",
+    deps = [
+        ":learn",
+        "//tensorflow:tensorflow_py",
+        "//tensorflow/python:framework_test_lib",
+    ],
+)
+
+py_test(
+    name = "dnn_test",
+    size = "small",
+    srcs = ["python/learn/estimators/dnn_test.py"],
+    srcs_version = "PY2AND3",
+    deps = [
+        ":learn",
+        "//tensorflow:tensorflow_py",
+        "//tensorflow/python:framework_test_lib",
+    ],
+)
+
+py_test(
+    name = "linear_test",
+    size = "small",
+    srcs = ["python/learn/estimators/linear_test.py"],
+    srcs_version = "PY2AND3",
+    deps = [
+        ":learn",
+        "//tensorflow:tensorflow_py",
+        "//tensorflow/python:framework_test_lib",
+    ],
+)
+
 py_test(
     name = "test_grid_search",
     size = "small",

tensorflow/contrib/learn/python/learn/__init__.py

@@ -20,6 +20,9 @@ from __future__ import print_function
 import numpy as np
 
 from tensorflow.contrib.learn.python.learn import datasets
+from tensorflow.contrib.learn.python.learn import estimators
+from tensorflow.contrib.learn.python.learn import graph_actions
+from tensorflow.contrib.learn.python.learn import io
 from tensorflow.contrib.learn.python.learn import models
 from tensorflow.contrib.learn.python.learn import monitors
 from tensorflow.contrib.learn.python.learn import ops

tensorflow/contrib/learn/python/learn/dataframe/queues/feeding_functions.py

@@ -0,0 +1,144 @@
+# Copyright 2016 Google Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Helper functions for enqueuing data from arrays and pandas `DataFrame`s."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import tensorflow as tf
+import tensorflow.contrib.learn.python.learn.dataframe.queues.feeding_queue_runner as fqr
+
+# pylint: disable=g-import-not-at-top
+try:
+  import pandas as pd
+  HAS_PANDAS = True
+except ImportError:
+  HAS_PANDAS = False
+
+
+class _ArrayFeedFn(object):
+  """Creates feed dictionaries from numpy arrays."""
+
+  def __init__(self, placeholders, array):
+    if len(placeholders) != 2:
+      raise ValueError("_array_feed_fn expects 2 placeholders; got {}.".format(
+          len(placeholders)))
+    self._placeholders = placeholders
+    self._array = array
+    self._reset()
+
+  def _reset(self):
+    self._row_iterator = enumerate(self._array)
+
+  def __call__(self):
+    try:
+      index, row = next(self._row_iterator)
+    except StopIteration:
+      self._reset()
+      index, row = next(self._row_iterator)
+    return {self._placeholders[0]: index, self._placeholders[1]: row}
+
+
+class _PandasFeedFn(object):
+  """Creates feed dictionaries from pandas `DataFrames`."""
+
+  def __init__(self, placeholders, dataframe):
+    if len(placeholders) != len(dataframe.columns) + 1:
+      raise ValueError("Expected {} placeholders; got {}.".format(
+          len(dataframe.columns), len(placeholders)))
+    self._index_placeholder = placeholders[0]
+    self._row_placeholders = placeholders[1:]
+    self._dataframe = dataframe
+    self._reset()
+
+  def _reset(self):
+    self._row_iterator = self._dataframe.iterrows()
+
+  def __call__(self):
+    try:
+      index, row = next(self._row_iterator)
+    except StopIteration:
+      self._reset()
+      index, row = next(self._row_iterator)
+    feed_dict = dict(zip(self._row_placeholders, row))
+    feed_dict[self._index_placeholder] = index
+    return feed_dict
+
+
+def enqueue_data(data,
+                 capacity,
+                 shuffle=False,
+                 min_after_dequeue=None,
+                 seed=None):
+  """Creates a queue filled from a numpy array or pandas `DataFrame`.
+
+    Returns a queue filled with the rows of the given array or `DataFrame`. In
+    the case of a pandas `DataFrame`, the first enqueued `Tensor` corresponds to
+    the index of the `DataFrame`. For numpy arrays, the first enqueued `Tensor`
+    contains the row number.
+
+  Args:
+    data: a numpy `ndarray or` pandas `DataFrame` that will be read into the
+      queue.
+    capacity: the capacity of the queue.
+    shuffle: whether or not to shuffle the rows of the array.
+    min_after_dequeue: minimum number of elements that can remain in the queue
+    after a dequeue operation. Only used when `shuffle` is true. If not set,
+    defaults to `capacity` / 4.
+    seed: used to seed RandomShuffleQueue. Only used when `shuffle` is True.
+
+  Returns:
+    A queue filled with the rows of the given array or `DataFrame`.
+
+  Raises:
+    TypeError: `data` is not a Pandas `DataFrame` or a numpy `ndarray`.
+  """
+  # TODO(jamieas): create multithreaded version of enqueue_data.
+  if isinstance(data, np.ndarray):
+    dtypes = [tf.int64, tf.as_dtype(data.dtype)]
+    shapes = [(), data.shape[1:]]
+    get_feed_fn = _ArrayFeedFn
+  elif HAS_PANDAS and isinstance(data, pd.DataFrame):
+    dtypes = [tf.as_dtype(dt) for dt in [data.index.dtype] + list(data.dtypes)]
+    shapes = [() for _ in dtypes]
+    get_feed_fn = _PandasFeedFn
+  else:
+    raise TypeError(
+        "data must be either a numpy array or pandas DataFrame if pandas is "
+        "installed; got {}".format(
+            type(data).__name__))
+
+  placeholders = [tf.placeholder(*type_and_shape)
+                  for type_and_shape in zip(dtypes, shapes)]
+  if shuffle:
+    min_after_dequeue = (capacity / 4 if min_after_dequeue is None else
+                         min_after_dequeue)
+    queue = tf.RandomShuffleQueue(capacity,
+                                  min_after_dequeue,
+                                  dtypes=dtypes,
+                                  shapes=shapes,
+                                  seed=seed)
+  else:
+    queue = tf.FIFOQueue(capacity, dtypes=dtypes, shapes=shapes)
+  enqueue_op = queue.enqueue(placeholders)
+  feed_fn = get_feed_fn(placeholders, data)
+  queue_runner = fqr.FeedingQueueRunner(queue=queue,
+                                        enqueue_ops=[enqueue_op],
+                                        feed_fn=feed_fn)
+  tf.train.add_queue_runner(queue_runner)
+  return queue

tensorflow/contrib/learn/python/learn/dataframe/queues/feeding_queue_runner.py

@@ -0,0 +1,105 @@
+# Copyright 2016 Google Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""A `QueueRunner` that takes a feed function as an argument."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.framework import errors
+from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.training import queue_runner as qr
+
+
+class FeedingQueueRunner(qr.QueueRunner):
+  """A queue runner that allows the feeding of values such as numpy arrays."""
+
+  def __init__(self, queue=None, enqueue_ops=None, close_op=None,
+               cancel_op=None, feed_fn=None):
+    """Initialize the queue runner.
+
+    For further documentation, see `queue_runner.py`. Note that
+    `FeedingQueueRunner` does not support construction from protobuffer nor
+    serialization to protobuffer.
+
+    Args:
+      queue: A `Queue`.
+      enqueue_ops: List of enqueue ops to run in threads later.
+      close_op: Op to close the queue. Pending enqueue ops are preserved.
+      cancel_op: Op to close the queue and cancel pending enqueue ops.
+      feed_fn: a function that returns a dictionary mapping fed `Tensor`s to
+      values.
+    """
+    super(FeedingQueueRunner, self).__init__(queue, enqueue_ops, close_op,
+                                             cancel_op)
+    self._feed_fn = feed_fn
+
+  # pylint: disable=broad-except
+  def _run(self, sess, enqueue_op, coord=None):
+    """Execute the enqueue op in a loop, close the queue in case of error.
+
+    Args:
+      sess: A `Session`.
+      enqueue_op: The `Operation` to run.
+      coord: Optional `Coordinator` object for reporting errors and checking
+        for stop conditions.
+
+    """
+    # TODO(jamieas): Reduce code duplication with `QueueRunner`.
+    decremented = False
+    try:
+      while True:
+        if coord and coord.should_stop():
+          break
+        try:
+          feed_dict = None if self._feed_fn is None else self._feed_fn()
+          sess.run(enqueue_op, feed_dict=feed_dict)
+        except errors.OutOfRangeError:
+          # This exception indicates that a queue was closed.
+          with self._lock:
+            self._runs -= 1
+            decremented = True
+            if self._runs == 0:
+              try:
+                sess.run(self._close_op)
+              except Exception as e:
+                # Intentionally ignore errors from close_op.
+                logging.vlog(1, "Ignored exception: %s", str(e))
+            return
+    except Exception as e:
+      # This catches all other exceptions.
+      if coord:
+        coord.request_stop(e)
+      else:
+        logging.error("Exception in QueueRunner: %s", str(e))
+        with self._lock:
+          self._exceptions_raised.append(e)
+        raise
+    finally:
+      # Make sure we account for all terminations: normal or errors.
+      if not decremented:
+        with self._lock:
+          self._runs -= 1
+
+  def _init_from_proto(self, queue_runner_def):
+    raise NotImplementedError(
+        "{} does not support initialization from proto.".format(type(
+            self).__name__))
+
+  def to_proto(self):
+    raise NotImplementedError(
+        "{} does not support serialization to proto.".format(type(
+            self).__name__))

tensorflow/contrib/learn/python/learn/estimators/__init__.py

@@ -18,11 +18,17 @@ from __future__ import print_function
 
 from tensorflow.contrib.learn.python.learn.estimators.autoencoder import TensorFlowDNNAutoencoder
 from tensorflow.contrib.learn.python.learn.estimators.base import TensorFlowEstimator, TensorFlowBaseTransformer
+from tensorflow.contrib.learn.python.learn.estimators.dnn import DNNClassifier
+from tensorflow.contrib.learn.python.learn.estimators.dnn import DNNRegressor
 from tensorflow.contrib.learn.python.learn.estimators.dnn import TensorFlowDNNClassifier
 from tensorflow.contrib.learn.python.learn.estimators.dnn import TensorFlowDNNRegressor
+from tensorflow.contrib.learn.python.learn.estimators.dnn_linear_combined import DNNLinearCombinedClassifier
+from tensorflow.contrib.learn.python.learn.estimators.dnn_linear_combined import DNNLinearCombinedRegressor
 from tensorflow.contrib.learn.python.learn.estimators.estimator import BaseEstimator
 from tensorflow.contrib.learn.python.learn.estimators.estimator import Estimator
 from tensorflow.contrib.learn.python.learn.estimators.estimator import ModeKeys
+from tensorflow.contrib.learn.python.learn.estimators.linear import LinearClassifier
+from tensorflow.contrib.learn.python.learn.estimators.linear import LinearRegressor
 from tensorflow.contrib.learn.python.learn.estimators.linear import TensorFlowClassifier
 from tensorflow.contrib.learn.python.learn.estimators.linear import TensorFlowLinearClassifier
 from tensorflow.contrib.learn.python.learn.estimators.linear import TensorFlowLinearRegressor

tensorflow/contrib/learn/python/learn/estimators/base.py

@@ -45,12 +45,13 @@ from tensorflow.python.ops import variable_scope as vs
 from tensorflow.python.training import training as train
 
 from tensorflow.contrib.layers import optimizers
-from tensorflow.contrib.learn.python.learn import trainer
 from tensorflow.contrib.learn.python.learn.io.data_feeder import setup_train_data_feeder
 from tensorflow.contrib.learn.python.learn.io.data_feeder import setup_predict_data_feeder
 from tensorflow.contrib.learn.python.learn.ops.dropout_ops import DROPOUTS
-from tensorflow.contrib.learn.python.learn import monitors
+from tensorflow.contrib.learn.python.learn import monitors as monitors_lib
+from tensorflow.contrib.learn.python.learn.utils import checkpoints
 
+from tensorflow.contrib.learn.python.learn.estimators import estimator
 from tensorflow.contrib.learn.python.learn.estimators import _sklearn
 from tensorflow.contrib.learn.python.learn.estimators._sklearn import NotFittedError
 from tensorflow.contrib.learn.python.learn.estimators.run_config import RunConfig
@@ -63,7 +64,35 @@ def _write_with_backup(filename, content):
     f.write(content)
 
 
-class TensorFlowEstimator(_sklearn.BaseEstimator):
+def _copy_dir(dir_in, dir_out):
+  gfile.MakeDirs(dir_out)
+  for name in gfile.ListDirectory(dir_in):
+    name_in = os.path.join(dir_in, name)
+    name_out = os.path.join(dir_out, name)
+    if gfile.IsDirectory(name_in):
+      gfile.MakeDirs(name_out)
+      _copy_dir(name_in, name_out)
+    else:
+      gfile.Copy(name_in, name_out, overwrite=True)
+
+
+def _new_tf_model_fn(model_fn, class_weight):
+  """Backward compatibility way of adding class weight and IS_TRAINING.
+
+  TODO(ipolosukhin): Remove this function after new layers are available.
+  Specifically:
+   * dropout and batch norm should work via update ops.
+   * class weights should be retrieved from weights column or hparams.
+  """
+  def _model_fn(features, targets, mode):
+    ops.get_default_graph().add_to_collection('IS_TRAINING', mode == 'train')
+    if class_weight is not None:
+      constant_op.constant(class_weight, name='class_weight')
+    return model_fn(features, targets)
+  return _model_fn
+
+
+class TensorFlowEstimator(estimator.Estimator):
   """Base class for all TensorFlow estimators.
 
   Parameters:
@@ -109,101 +138,21 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
                continue_training=False,
                config=None,
                verbose=1):
-    self.model_fn = model_fn
+    super(TensorFlowEstimator, self).__init__(
+        model_fn=_new_tf_model_fn(model_fn, class_weight),
+        classification=n_classes > 1,
+        learning_rate=learning_rate,
+        optimizer=optimizer,
+        clip_gradients=clip_gradients,
+        config=config)
     self.n_classes = n_classes
     self.batch_size = batch_size
     self.steps = steps
     self.verbose = verbose
-    self.optimizer = optimizer
-    self.learning_rate = learning_rate
-    self.clip_gradients = clip_gradients
     self.continue_training = continue_training
-    self._initialized = False
-    self.class_weight = class_weight
-    self._config = config
+    self._data_feeder = None
 
-  def _setup_training(self):
-    """Sets up graph, model and trainer."""
-    # Create config if not given.
-    if self._config is None:
-      self._config = RunConfig(verbose=self.verbose)
-    # Create new graph.
-    self._graph = ops.Graph()
-    self._graph.add_to_collection('IS_TRAINING', True)
-    with self._graph.as_default():
-      random_seed.set_random_seed(self._config.tf_random_seed)
-      self._global_step = variables.Variable(0,
-                                             name='global_step',
-                                             trainable=False)
-
-      # Setting up inputs and outputs.
-      self._inp, self._out = self._data_feeder.input_builder()
-
-      # If class weights are provided, add them to the graph.
-      # Different loss functions can use this tensor by name.
-      if self.class_weight:
-        self._class_weight_node = constant_op.constant(self.class_weight,
-                                                       name='class_weight')
-
-      # Add histograms for X and y if they are floats.
-      if self._data_feeder.input_dtype in (np.float32, np.float64):
-        logging_ops.histogram_summary('X', self._inp)
-      if self._data_feeder.output_dtype in (np.float32, np.float64)\
-        and self._out is not None:
-        logging_ops.histogram_summary('y', self._out)
-
-      # Create model's graph.
-      self._model_predictions, self._model_loss = self.model_fn(self._inp,
-                                                                self._out)
-
-      # Set up a single operator to merge all the summaries
-      self._summaries = logging_ops.merge_all_summaries()
-
-      # Create trainer and augment graph with gradients and optimizer.
-      # Additionally creates initialization ops.
-      learning_rate = self.learning_rate
-      optimizer = self.optimizer
-      if callable(learning_rate):
-        learning_rate = learning_rate(self._global_step)
-      if callable(optimizer):
-        optimizer = optimizer(learning_rate)
-      self._train = optimizers.optimize_loss(self._model_loss,
-                                             self._global_step,
-                                             learning_rate=learning_rate,
-                                             optimizer=optimizer,
-                                             clip_gradients=self.clip_gradients)
-
-      # Update ops during training, e.g. batch_norm_ops
-      self._train = control_flow_ops.group(self._train, *
-                                           ops.get_collection('update_ops'))
-
-      # Get all initializers for all trainable variables.
-      self._initializers = variables.initialize_all_variables()
-
-      # Create model's saver capturing all the nodes created up until now.
-      self._saver = train.Saver(max_to_keep=self._config.keep_checkpoint_max,
-                                keep_checkpoint_every_n_hours=
-                                self._config.keep_checkpoint_every_n_hours)
-
-      # Enable monitor to create validation data dict with appropriate
-      # tf placeholders
-      self._monitor.create_val_feed_dict(self._inp, self._out)
-
-      # Create session to run model with.
-      self._session = session.Session(self._config.master,
-                                      config=self._config.tf_config)
-
-      # Run parameter initializers.
-      self._session.run(self._initializers)
-
-  def _setup_summary_writer(self, logdir):
-    """Sets up summary writer to prepare for later optional visualization."""
-    self._summary_writer = train.SummaryWriter(
-        os.path.join(logdir,
-                     datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')),
-        graph=self._session.graph)
-
-  def fit(self, X, y, monitor=None, logdir=None):
+  def fit(self, x, y, steps=None, monitors=None, logdir=None):
     """Builds a neural network model given provided `model_fn` and training
     data X and y.
 
@@ -214,63 +163,44 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
     To restart learning, create new estimator.
 
     Args:
-      X: matrix or tensor of shape [n_samples, n_features...]. Can be
+      x: matrix or tensor of shape [n_samples, n_features...]. Can be
       iterator that returns arrays of features. The training input
       samples for fitting the model.
       y: vector or matrix [n_samples] or [n_samples, n_outputs]. Can be
       iterator that returns array of targets. The training target values
       (class labels in classification, real numbers in regression).
-      monitor: Monitor object to print training progress and invoke early
-        stopping
+      steps: int, number of steps to train.
+             If None or 0, train for `self.steps`.
+      monitors: List of `BaseMonitor` objects to print training progress and
+        invoke early stopping.
       logdir: the directory to save the log file that can be used for
       optional visualization.
 
     Returns:
       Returns self.
     """
-    # Sets up data feeder.
-    self._data_feeder = setup_train_data_feeder(X, y, self.n_classes,
-                                                self.batch_size)
-
-    if monitor is None:
-      self._monitor = monitors.default_monitor(verbose=self.verbose)
-    else:
-      self._monitor = monitor
-
-    if not self.continue_training or not self._initialized:
-      # Sets up model and trainer.
-      self._setup_training()
-      self._initialized = True
-    else:
-      self._data_feeder.set_placeholders(self._inp, self._out)
-
-    # Sets up summary writer for later optional visualization.
-    # Due to not able to setup _summary_writer in __init__ as it's not a
-    # parameter of the model, here we need to check if such variable exists
-    # and if it's None or not (in case it was setup in a previous run).
-    # It is initialized only in the case where it wasn't before and log dir
-    # is provided.
-    if logdir:
-      if (not hasattr(self, '_summary_writer') or
-          (hasattr(self, '_summary_writer') and self._summary_writer is None)):
-        self._setup_summary_writer(logdir)
-    else:
-      self._summary_writer = None
-
-    # Train model for given number of steps.
-    trainer.train(self._session,
-                  self._train,
-                  self._model_loss,
-                  self._global_step,
-                  self._data_feeder.get_feed_dict_fn(),
-                  steps=self.steps,
-                  monitor=self._monitor,
-                  summary_writer=self._summary_writer,
-                  summaries=self._summaries,
-                  feed_params_fn=self._data_feeder.get_feed_params)
+    if logdir is not None:
+      self._model_dir = logdir
+    self._data_feeder = setup_train_data_feeder(
+        x, y, n_classes=self.n_classes, batch_size=self.batch_size)
+    self._train_model(input_fn=self._data_feeder.input_builder,
+                      feed_fn=self._data_feeder.get_feed_dict_fn(),
+                      steps=steps or self.steps,
+                      monitors=monitors)
     return self
 
-  def partial_fit(self, X, y):
+  def evaluate(self, x=None, y=None, input_fn=None, steps=None):
+    """See base class."""
+    feed_fn = None
+    if x is not None:
+      eval_data_feeder = setup_train_data_feeder(
+          x, y, n_classes=self.n_classes, batch_size=self.batch_size, epochs=1)
+      input_fn, feed_fn = (eval_data_feeder.input_builder,
+                           eval_data_feeder.get_feed_dict_fn())
+    return self._evaluate_model(
+        input_fn=input_fn, feed_fn=feed_fn, steps=steps or self.steps)
+
+  def partial_fit(self, x, y):
     """Incremental fit on a batch of samples.
 
     This method is expected to be called several times consecutively
@@ -282,7 +212,7 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
     to converge, and you want to split up training into subparts.
 
     Args:
-      X: matrix or tensor of shape [n_samples, n_features...]. Can be
+      x: matrix or tensor of shape [n_samples, n_features...]. Can be
       iterator that returns arrays of features. The training input
       samples for fitting the model.
       y: vector or matrix [n_samples] or [n_samples, n_outputs]. Can be
@@ -292,33 +222,30 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
     Returns:
       Returns self.
     """
-    return self.fit(X, y)
-
-  def _predict(self, X, axis=-1, batch_size=None):
-    if not self._initialized:
-      raise _sklearn.NotFittedError()
+    return self.fit(x, y)
 
+  def _predict(self, x, axis=-1, batch_size=None):
+    if self._graph is None:
+      raise NotFittedError()
     # Use the batch size for fitting if the user did not specify one.
     if batch_size is None:
       batch_size = self.batch_size
 
-    self._graph.add_to_collection('IS_TRAINING', False)
-    predict_data_feeder = setup_predict_data_feeder(X, batch_size=batch_size)
-    preds = []
-    dropouts = self._graph.get_collection(DROPOUTS)
-    feed_dict = {prob: 1.0 for prob in dropouts}
-    for data in predict_data_feeder:
-      feed_dict[self._inp] = data
-      predictions_for_batch = self._session.run(self._model_predictions,
-                                                feed_dict)
-      if self.n_classes > 1 and axis != -1:
-        preds.append(predictions_for_batch.argmax(axis=axis))
-      else:
-        preds.append(predictions_for_batch)
+    predict_data_feeder = setup_train_data_feeder(
+        x, None, n_classes=None,
+        batch_size=batch_size,
+        shuffle=False, epochs=1)
 
-    return np.concatenate(preds, axis=0)
+    preds = self._infer_model(
+        input_fn=predict_data_feeder.input_builder,
+        feed_fn=predict_data_feeder.get_feed_dict_fn())
+    if self.n_classes > 1 and axis != -1:
+      preds = preds['predictions'].argmax(axis=axis)
+    else:
+      preds = preds['predictions']
+    return preds
 
-  def predict(self, X, axis=1, batch_size=None):
+  def predict(self, x, axis=1, batch_size=None):
     """Predict class or regression for X.
 
     For a classification model, the predicted class for each sample in X is
@@ -326,7 +253,7 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
     returned.
 
     Args:
-      X: array-like matrix, [n_samples, n_features...] or iterator.
+      x: array-like matrix, [n_samples, n_features...] or iterator.
       axis: Which axis to argmax for classification.
         By default axis 1 (next after batch) is used.
         Use 2 for sequence predictions.
@@ -338,13 +265,13 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
       y: array of shape [n_samples]. The predicted classes or predicted
       value.
     """
-    return self._predict(X, axis=axis, batch_size=batch_size)
+    return self._predict(x, axis=axis, batch_size=batch_size)
 
-  def predict_proba(self, X, batch_size=None):
+  def predict_proba(self, x, batch_size=None):
     """Predict class probability of the input samples X.
 
     Args:
-      X: array-like matrix, [n_samples, n_features...] or iterator.
+      x: array-like matrix, [n_samples, n_features...] or iterator.
       batch_size: If test set is too big, use batch size to split
         it into mini batches. By default the batch_size member variable is used.
 
@@ -352,7 +279,7 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
       y: array of shape [n_samples, n_classes]. The predicted
       probabilities for each class.
     """
-    return self._predict(X, batch_size=batch_size)
+    return self._predict(x, batch_size=batch_size)
 
   def get_tensor(self, name):
     """Returns tensor by name.
@@ -374,7 +301,9 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
     Returns:
       Numpy array - value of the tensor.
     """
-    return self._session.run(self.get_tensor(name))
+    if name.endswith(':0'):
+      name = name[:-2]
+    return checkpoints.load_variable(self.model_dir, name)
 
   def get_variable_names(self):
     """Returns list of all variable names in this model.
@@ -382,8 +311,7 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
     Returns:
       List of names.
     """
-    with self._graph.as_default():
-      return [v.name for v in variables.all_variables()]
+    return [name for name, _ in checkpoints.list_variables(self.model_dir)]
 
   def save(self, path):
     """Saves checkpoints and graph to given path.
@@ -391,17 +319,12 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
     Args:
       path: Folder to save model to.
     """
-    if not self._initialized:
-      raise _sklearn.NotFittedError()
+    if self._graph is None:
+      raise NotFittedError
 
-    # Currently Saver requires absolute path to work correctly.
-    path = os.path.abspath(path)
+    # Copy model dir into new path.
+    _copy_dir(self.model_dir, path)
 
-    if not os.path.exists(path):
-      os.makedirs(path)
-    if not os.path.isdir(path):
-      raise ValueError('Path %s should be a directory to save'
-                       'checkpoints and graph.' % path)
     # Save model definition.
     all_params = self.get_params()
     params = {}
@@ -414,25 +337,6 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
         default=lambda o: o.__dict__ if hasattr(o, '__dict__') else None)
     _write_with_backup(os.path.join(path, 'model.def'), model_def)
 
-    # Save checkpoints.
-    endpoints = '%s\n%s\n%s\n%s' % (self._inp.name, self._out.name,
-                                    self._model_predictions.name,
-                                    self._model_loss.name)
-    _write_with_backup(os.path.join(path, 'endpoints'), endpoints)
-
-    # Save graph definition.
-    _write_with_backup(
-        os.path.join(path, 'graph.pbtxt'), str(self._graph.as_graph_def()))
-
-    # Save saver definition.
-    _write_with_backup(
-        os.path.join(path, 'saver.pbtxt'), str(self._saver.as_saver_def()))
-
-    # Save checkpoints.
-    self._saver.save(self._session,
-                     os.path.join(path, 'model'),
-                     global_step=self._global_step)
-
   def _restore(self, path):
     """Restores this estimator from given path.
 
@@ -442,57 +346,7 @@ class TensorFlowEstimator(_sklearn.BaseEstimator):
     Args:
       path: Path to checkpoints and other information.
     """
-    # Currently Saver requires absolute path to work correctly.
-    path = os.path.abspath(path)
-
-    self._graph = ops.Graph()
-    with self._graph.as_default():
-      endpoints_filename = os.path.join(path, 'endpoints')
-      if not os.path.exists(endpoints_filename):
-        raise ValueError("Restore folder doesn't contain endpoints.")
-      with gfile.Open(endpoints_filename) as foutputs:
-        endpoints = foutputs.read().split('\n')
-      graph_filename = os.path.join(path, 'graph.pbtxt')
-      if not os.path.exists(graph_filename):
-        raise ValueError("Restore folder doesn't contain graph definition.")
-      with gfile.Open(graph_filename) as fgraph:
-        graph_def = graph_pb2.GraphDef()
-        text_format.Merge(fgraph.read(), graph_def)
-        (self._inp, self._out, self._model_predictions,
-         self._model_loss) = importer.import_graph_def(
-             graph_def, name='', return_elements=endpoints)
-      saver_filename = os.path.join(path, 'saver.pbtxt')
-      if not os.path.exists(saver_filename):
-        raise ValueError("Restore folder doesn't contain saver definition.")
-      with gfile.Open(saver_filename) as fsaver:
-        saver_def = train.SaverDef()
-        text_format.Merge(fsaver.read(), saver_def)
-        self._saver = train.Saver(saver_def=saver_def)
-
-      # Restore trainer
-      self._global_step = self._graph.get_tensor_by_name('global_step:0')
-      self._train = self._graph.get_operation_by_name('OptimizeLoss/train')
-
-      # Restore summaries.
-      self._summaries = self._graph.get_operation_by_name(
-          'MergeSummary/MergeSummary')
-
-      # Restore session.
-      if not isinstance(self._config, RunConfig):
-        self._config = RunConfig(verbose=self.verbose)
-      self._session = session.Session(self._config.master,
-                                      config=self._config.tf_config)
-      checkpoint_path = train.latest_checkpoint(path)
-      if checkpoint_path is None:
-        raise ValueError(
-            'Missing checkpoint files in the %s. Please '
-            'make sure you are you have checkpoint file that describes '
-            'latest checkpoints and appropriate checkpoints are there. '
-            'If you have moved the folder, you at this point need to '
-            'update manually update the paths in the checkpoint file.' % path)
-      self._saver.restore(self._session, checkpoint_path)
-    # Set to be initialized.
-    self._initialized = True
+    raise NotImplementedError
 
   # pylint: disable=unused-argument
   @classmethod
@@ -549,7 +403,7 @@ class TensorFlowBaseTransformer(TensorFlowEstimator, _sklearn.TransformerMixin):
 
     def fit(self, X, y=None, monitor=None, logdir=None):
         """Fit a transformer."""
-        return(super(TensorFlowBaseTransformer, self).fit(X, y, monitor=None, logdir=None))
+        return(super(TensorFlowBaseTransformer, self).fit(X, y, monitors=None, logdir=None))
 
     def fit_transform(self, X, y=None, monitor=None, logdir=None):
         """Fit transformer and transform X using trained transformer."""

tensorflow/contrib/learn/python/learn/estimators/dnn.py

@@ -16,11 +16,176 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-from tensorflow.contrib.learn.python.learn.estimators import _sklearn
-from tensorflow.contrib.learn.python.learn.estimators.base import TensorFlowEstimator
+from tensorflow.contrib import layers
 from tensorflow.contrib.learn.python.learn import models
+from tensorflow.contrib.learn.python.learn.estimators import _sklearn
+from tensorflow.contrib.learn.python.learn.estimators import dnn_linear_combined
+from tensorflow.contrib.learn.python.learn.estimators.base import TensorFlowEstimator
+from tensorflow.python.ops import nn
 
 
+# TODO(ipolosukhin): Merge thirdparty DNN with this.
+class DNNClassifier(dnn_linear_combined.DNNLinearCombinedClassifier):
+  """A classifier for TensorFlow DNN models.
+
+    Example:
+      ```
+      installed_app_id = sparse_column_with_hash_bucket("installed_id", 1e6)
+      impression_app_id = sparse_column_with_hash_bucket("impression_id", 1e6)
+
+      installed_emb = embedding_column(installed_app_id, dimension=16,
+                                       combiner="sum")
+      impression_emb = embedding_column(impression_app_id, dimension=16,
+                                        combiner="sum")
+
+      estimator = DNNClassifier(
+          feature_columns=[installed_emb, impression_emb],
+          hidden_units=[1024, 512, 256])
+
+      # Input builders
+      def input_fn_train: # returns X, Y
+        pass
+      estimator.train(input_fn_train)
+
+      def input_fn_eval: # returns X, Y
+        pass
+      estimator.evaluate(input_fn_eval)
+      estimator.predict(x)
+      ```
+
+    Input of `fit`, `train`, and `evaluate` should have following features,
+      otherwise there will be a `KeyError`:
+        if `weight_column_name` is not `None`, a feature with
+          `key=weight_column_name` whose value is a `Tensor`.
+        for each `column` in `feature_columns`:
+        - if `column` is a `SparseColumn`, a feature with `key=column.name`
+          whose `value` is a `SparseTensor`.
+        - if `column` is a `RealValuedColumn, a feature with `key=column.name`
+          whose `value` is a `Tensor`.
+        - if `feauture_columns` is None, then `input` must contains only real
+          valued `Tensor`.
+
+  Parameters:
+    hidden_units: List of hidden units per layer. All layers are fully
+      connected. Ex. [64, 32] means first layer has 64 nodes and second one has
+      32.
+    feature_columns: An iterable containing all the feature columns used by the
+      model. All items in the set should be instances of classes derived from
+      `FeatureColumn`.
+    model_dir: Directory to save model parameters, graph and etc.
+    n_classes: number of target classes. Default is binary classification.
+      It must be greater than 1.
+    weight_column_name: A string defining feature column name representing
+      weights. It is used to down weight or boost examples during training. It
+      will be multiplied by the loss of the example.
+    optimizer: An instance of `tf.Optimizer` used to train the model. If `None`,
+      will use an Adagrad optimizer.
+    activation_fn: Activation function applied to each layer. If `None`, will
+      use `tf.nn.relu`.
+  """
+
+  def __init__(self,
+               hidden_units,
+               feature_columns=None,
+               model_dir=None,
+               n_classes=2,
+               weight_column_name=None,
+               optimizer=None,
+               activation_fn=nn.relu):
+    super(DNNClassifier, self).__init__(n_classes=n_classes,
+                                        weight_column_name=weight_column_name,
+                                        dnn_feature_columns=feature_columns,
+                                        dnn_optimizer=optimizer,
+                                        dnn_hidden_units=hidden_units,
+                                        dnn_activation_fn=activation_fn)
+
+  def _get_train_ops(self, features, targets):
+    """See base class."""
+    if self._dnn_feature_columns is None:
+      self._dnn_feature_columns = layers.infer_real_valued_columns(features)
+    return super(DNNClassifier, self)._get_train_ops(features, targets)
+
+
+class DNNRegressor(dnn_linear_combined.DNNLinearCombinedRegressor):
+  """A regressor for TensorFlow DNN models.
+
+    Example:
+      ```
+      installed_app_id = sparse_column_with_hash_bucket("installed_id", 1e6)
+      impression_app_id = sparse_column_with_hash_bucket("impression_id", 1e6)
+
+      installed_emb = embedding_column(installed_app_id, dimension=16,
+                                       combiner="sum")
+      impression_emb = embedding_column(impression_app_id, dimension=16,
+                                        combiner="sum")
+
+      estimator = DNNRegressor(
+          feature_columns=[installed_emb, impression_emb],
+          hidden_units=[1024, 512, 256])
+
+      # Input builders
+      def input_fn_train: # returns X, Y
+        pass
+      estimator.train(input_fn_train)
+
+      def input_fn_eval: # returns X, Y
+        pass
+      estimator.evaluate(input_fn_eval)
+      estimator.predict(x)
+      ```
+
+    Input of `fit`, `train`, and `evaluate` should have following features,
+      otherwise there will be a `KeyError`:
+        if `weight_column_name` is not `None`, a feature with
+          `key=weight_column_name` whose value is a `Tensor`.
+        for each `column` in `feature_columns`:
+        - if `column` is a `SparseColumn`, a feature with `key=column.name`
+          whose `value` is a `SparseTensor`.
+        - if `column` is a `RealValuedColumn, a feature with `key=column.name`
+          whose `value` is a `Tensor`.
+        - if `feauture_columns` is None, then `input` must contains only real
+          valued `Tensor`.
+
+
+
+  Parameters:
+    hidden_units: List of hidden units per layer. All layers are fully
+      connected. Ex. [64, 32] means first layer has 64 nodes and second one has
+      32.
+    feature_columns: An iterable containing all the feature columns used by the
+      model. All items in the set should be instances of classes derived from
+      `FeatureColumn`.
+    model_dir: Directory to save model parameters, graph and etc.
+    weight_column_name: A string defining feature column name representing
+      weights. It is used to down weight or boost examples during training. It
+      will be multiplied by the loss of the example.
+    optimizer: An instance of `tf.Optimizer` used to train the model. If `None`,
+      will use an Adagrad optimizer.
+    activation_fn: Activation function applied to each layer. If `None`, will
+      use `tf.nn.relu`.
+  """
+
+  def __init__(self,
+               hidden_units,
+               feature_columns=None,
+               model_dir=None,
+               weight_column_name=None,
+               optimizer=None,
+               activation_fn=nn.relu):
+    super(DNNRegressor, self).__init__(weight_column_name=weight_column_name,
+                                       dnn_feature_columns=feature_columns,
+                                       dnn_optimizer=optimizer,
+                                       dnn_hidden_units=hidden_units,
+                                       dnn_activation_fn=activation_fn)
+
+  def _get_train_ops(self, features, targets):
+    """See base class."""
+    if self._dnn_feature_columns is None:
+      self._dnn_feature_columns = layers.infer_real_valued_columns(features)
+    return super(DNNRegressor, self)._get_train_ops(features, targets)
+
+
+# TODO(ipolosukhin): Deprecate this class in favor of DNNClassifier.
 class TensorFlowDNNClassifier(TensorFlowEstimator, _sklearn.ClassifierMixin):
   """TensorFlow DNN Classifier model.
 
@@ -84,16 +249,16 @@ class TensorFlowDNNClassifier(TensorFlowEstimator, _sklearn.ClassifierMixin):
   @property
   def weights_(self):
     """Returns weights of the DNN weight layers."""
-    return [self._session.run(w)
+    return [self.get_tensor_value(w.name)
             for w in self._graph.get_collection('dnn_weights')
-           ] + [self.get_tensor_value('logistic_regression/weights:0')]
+           ] + [self.get_tensor_value('logistic_regression/weights')]
 
   @property
   def bias_(self):
     """Returns bias of the DNN's bias layers."""
-    return [self._session.run(b)
+    return [self.get_tensor_value(b.name)
             for b in self._graph.get_collection('dnn_biases')
-           ] + [self.get_tensor_value('logistic_regression/bias:0')]
+           ] + [self.get_tensor_value('logistic_regression/bias')]
 
 
 class TensorFlowDNNRegressor(TensorFlowEstimator, _sklearn.RegressorMixin):
@@ -157,13 +322,13 @@ class TensorFlowDNNRegressor(TensorFlowEstimator, _sklearn.RegressorMixin):
   @property
   def weights_(self):
     """Returns weights of the DNN weight layers."""
-    return [self._session.run(w)
+    return [self.get_tensor_value(w.name)
             for w in self._graph.get_collection('dnn_weights')
-           ] + [self.get_tensor_value('linear_regression/weights:0')]
+           ] + [self.get_tensor_value('linear_regression/weights')]
 
   @property
   def bias_(self):
     """Returns bias of the DNN's bias layers."""
-    return [self._session.run(b)
+    return [self.get_tensor_value(b.name)
             for b in self._graph.get_collection('dnn_biases')
-           ] + [self.get_tensor_value('linear_regression/bias:0')]
+           ] + [self.get_tensor_value('linear_regression/bias')]

tensorflow/contrib/learn/python/learn/estimators/dnn_linear_combined.py

@@ -0,0 +1,490 @@
+#  Copyright 2015 Google Inc. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+"""TensorFlow estimators for Linear and DNN joined training models."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import math
+
+import six
+
+from tensorflow.contrib import layers
+from tensorflow.contrib import metrics as metrics_lib
+from tensorflow.contrib.framework.python.ops import variables as variables
+from tensorflow.contrib.learn.python.learn.estimators import estimator
+from tensorflow.python.framework import ops
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import gradients
+from tensorflow.python.ops import logging_ops
+from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import nn
+from tensorflow.python.ops import parsing_ops
+from tensorflow.python.ops import state_ops
+
+
+# TODO(ispir): Increase test coverage
+class _DNNLinearCombinedBaseEstimator(estimator.BaseEstimator):
+  """An estimator for TensorFlow Linear and DNN joined training models.
+
+    Input of `fit`, `train`, and `evaluate` should have following features,
+      otherwise there will be a `KeyError`:
+        if `weight_column_name` is not `None`, a feature with
+          `key=weight_column_name` whose value is a `Tensor`.
+        for each `column` in `dnn_feature_columns` + `linear_feature_columns`:
+        - if `column` is a `SparseColumn`, a feature with `key=column.name`
+          whose `value` is a `SparseTensor`.
+        - if `column` is a `RealValuedColumn, a feature with `key=column.name`
+          whose `value` is a `Tensor`.
+
+  Parameters:
+    model_dir: Directory to save model parameters, graph and etc.
+    n_classes: number of target classes. Default is binary classification.
+    weight_column_name: A string defining feature column name representing
+      weights. It is used to down weight or boost examples during training. It
+      will be multiplied by the loss of the example.
+    linear_feature_columns: An iterable containing all the feature columns used
+      by linear part of the model. All items in the set should be instances of
+      classes derived from `FeatureColumn`.
+    linear_optimizer: An instance of `tf.Optimizer` used to apply gradients to
+      the linear part of the model. If `None`, will use a FTRL optimizer.
+    dnn_feature_columns: An iterable containing all the feature columns used by
+      deep part of the model. All items in the set should be instances of
+      classes derived from `FeatureColumn`.
+    dnn_hidden_units: List of hidden units per layer. All layers are fully
+      connected.
+    dnn_optimizer: An instance of `tf.Optimizer` used to apply gradients to the
+      deep part of the model. If `None`, will use an Adagrad optimizer.
+    dnn_activation_fn: Activation function applied to each layer. If `None`,
+      will use `tf.nn.relu`.
+
+    Raises:
+      ValueError: If both linear_feature_columns and dnn_features_columns are
+        empty at the same time.
+  """
+
+  def __init__(self,
+               model_dir=None,
+               n_classes=2,
+               weight_column_name=None,
+               linear_feature_columns=None,
+               linear_optimizer=None,
+               dnn_feature_columns=None,
+               dnn_optimizer=None,
+               dnn_hidden_units=None,
+               dnn_activation_fn=nn.relu):
+    super(_DNNLinearCombinedBaseEstimator, self).__init__(model_dir=model_dir)
+    self._n_classes = n_classes
+    self._weight_column_name = weight_column_name
+    self._linear_feature_columns = linear_feature_columns
+    self._linear_optimizer = linear_optimizer
+    self._dnn_feature_columns = dnn_feature_columns
+    self._dnn_optimizer = dnn_optimizer
+    self._dnn_hidden_units = dnn_hidden_units
+    self._dnn_activation_fn = dnn_activation_fn
+    if self._dnn_activation_fn is None:
+      self._dnn_activation_fn = nn.relu
+    self._dnn_weight_collection = "DNNLinearCombined_dnn"
+    self._linear_weight_collection = "DNNLinearCombined_linear"
+
+  def _get_train_ops(self, features, targets):
+    """See base class."""
+    global_step = variables.get_global_step()
+    assert global_step
+    loss = self._loss(
+        self._logits(features), targets, self._get_weight_tensor(features))
+    logging_ops.scalar_summary("loss", loss)
+
+    linear_vars = self._get_linear_vars()
+    dnn_vars = self._get_dnn_vars()
+    grads = gradients.gradients(loss, dnn_vars + linear_vars)
+    dnn_grads = grads[0:len(dnn_vars)]
+    linear_grads = grads[len(dnn_vars):]
+
+    train_ops = self._get_linear_training_ops(
+        linear_grads, linear_vars) + self._get_dnn_training_ops(dnn_grads,
+                                                                dnn_vars)
+
+    train_step = control_flow_ops.group(*train_ops, name="combined_training_op")
+    with ops.control_dependencies([train_step]):
+      with ops.get_default_graph().colocate_with(global_step):
+        return state_ops.assign_add(global_step, 1).op, loss
+
+  def _get_eval_ops(self, features, targets, metrics):
+    """See base class."""
+    predictions = self._get_predict_ops(features)
+    result = {}
+    for name, metric in six.iteritems(metrics):
+      result[name] = metric(predictions, targets,
+                            self._get_weight_tensor(features))
+    return result
+
+  def _get_default_metric_functions(self):
+    """See base class."""
+    def _compute_loss(logits, targets, weights=None):
+      return self._loss(logits, targets, weight_tensor=weights)
+
+    def _compute_accuracy(logits, targets, weights=None):
+      if self._n_classes > 2:
+        _, predictions = nn.top_k(logits, 1)
+      else:
+        predictions = array_ops.reshape(logits, [-1])
+        predictions = math_ops.greater(predictions,
+                                       array_ops.zeros_like(predictions))
+        targets = array_ops.reshape(targets, [-1])
+      return metrics_lib.accuracy(
+          math_ops.to_int32(predictions), math_ops.to_int32(targets), weights)
+
+    result = {"loss": _compute_loss}
+    if self._n_classes > 1:
+      result["accuracy"] = _compute_accuracy
+    return result
+
+  def _get_predict_ops(self, features):
+    return self._logits(features)
+
+  def _get_feature_ops_from_example(self, examples_batch):
+    column_types = layers.create_dict_for_parse_example(
+        self._get_linear_feature_columns() + self._get_dnn_feature_columns())
+    features = parsing_ops.parse_example(examples_batch, column_types)
+    return features
+
+  def _num_label_columns(self):
+    return 1 if self._n_classes <= 2 else self._n_classes
+
+  def _get_linear_feature_columns(self):
+    return sorted(
+        set(self._linear_feature_columns),
+        key=lambda x: x.key) if self._linear_feature_columns else None
+
+  def _get_dnn_feature_columns(self):
+    return sorted(set(
+        self._dnn_feature_columns)) if self._dnn_feature_columns else None
+
+  def _dnn_logits(self, features):
+    net = layers.input_from_feature_columns(
+        features,
+        self._get_dnn_feature_columns(),
+        weight_collections=[self._dnn_weight_collection])
+    for layer_id, num_hidden_units in enumerate(self._dnn_hidden_units):
+      net = layers.legacy_fully_connected(
+          net,
+          num_hidden_units,
+          activation_fn=self._dnn_activation_fn,
+          weight_collections=[self._dnn_weight_collection],
+          bias_collections=[self._dnn_weight_collection],
+          name="hiddenlayer_%d" % layer_id)
+      self._add_hidden_layer_summary(net, "hiddenlayer_%d" % layer_id)
+    logit = layers.legacy_fully_connected(
+        net,
+        self._num_label_columns(),
+        weight_collections=[self._dnn_weight_collection],
+        bias_collections=[self._dnn_weight_collection],
+        name="dnn_logit")
+    self._add_hidden_layer_summary(logit, "dnn_logit")
+    return logit
+
+  def _add_hidden_layer_summary(self, value, tag):
+    # TODO(zakaria): Move this code to tf.learn and add test.
+    logging_ops.scalar_summary("%s:fraction_of_zero_values" % tag,
+                               nn.zero_fraction(value))
+    logging_ops.histogram_summary("%s:activation" % tag, value)
+
+  def _linear_logits(self, features):
+    logits, _, _ = layers.weighted_sum_from_feature_columns(
+        columns_to_tensors=features,
+        feature_columns=self._get_linear_feature_columns(),
+        num_outputs=self._num_label_columns(),
+        weight_collections=[self._linear_weight_collection],
+        name="linear")
+    return logits
+
+  def _get_feature_dict(self, features):
+    if isinstance(features, dict):
+      return features
+    return {"": features}
+
+  def _logits(self, features):
+    if not (self._get_linear_feature_columns() or
+            self._get_dnn_feature_columns()):
+      raise ValueError("Either linear_feature_columns or dnn_feature_columns "
+                       "should be defined.")
+
+    features = self._get_feature_dict(features)
+    if self._get_linear_feature_columns() and self._get_dnn_feature_columns():
+      return self._linear_logits(features) + self._dnn_logits(features)
+    elif self._get_dnn_feature_columns():
+      return self._dnn_logits(features)
+    else:
+      return self._linear_logits(features)
+
+  def _get_weight_tensor(self, features):
+    if not self._weight_column_name:
+      return None
+    else:
+      return math_ops.to_float(features[self._weight_column_name])
+
+  def _loss(self, logits, target, weight_tensor):
+    if self._n_classes < 2:
+      loss_vec = math_ops.square(logits - math_ops.to_float(target))
+    elif self._n_classes == 2:
+      loss_vec = nn.sigmoid_cross_entropy_with_logits(logits,
+                                                      math_ops.to_float(target))
+    else:
+      loss_vec = nn.sparse_softmax_cross_entropy_with_logits(
+          logits, array_ops.reshape(target, [-1]))
+
+    if weight_tensor is None:
+      return math_ops.reduce_mean(loss_vec, name="loss")
+    else:
+      loss_vec = array_ops.reshape(loss_vec, shape=(-1,))
+      loss_vec = math_ops.mul(
+          loss_vec, array_ops.reshape(weight_tensor, shape=(-1,)))
+      return math_ops.div(
+          math_ops.reduce_sum(loss_vec),
+          math_ops.to_float(math_ops.reduce_sum(weight_tensor)),
+          name="loss")
+
+  def _get_linear_vars(self):
+    if self._get_linear_feature_columns():
+      return ops.get_collection(self._linear_weight_collection)
+    return []
+
+  def _get_linear_training_ops(self, linear_grads, linear_vars):
+    if self._get_linear_feature_columns():
+      self._linear_optimizer = self._get_optimizer(
+          self._linear_optimizer,
+          default_optimizer="Ftrl",
+          default_learning_rate=1. / math.sqrt(len(
+              self._get_linear_feature_columns())))
+      return [
+          self._linear_optimizer.apply_gradients(zip(linear_grads, linear_vars))
+      ]
+    return []
+
+  def _get_dnn_vars(self):
+    if self._get_dnn_feature_columns():
+      return ops.get_collection(self._dnn_weight_collection)
+    return []
+
+  def _get_dnn_training_ops(self, dnn_grads, dnn_vars):
+    if self._get_dnn_feature_columns():
+      self._dnn_optimizer = self._get_optimizer(self._dnn_optimizer,
+                                                default_optimizer="Adagrad",
+                                                default_learning_rate=0.05)
+      return [self._dnn_optimizer.apply_gradients(zip(dnn_grads, dnn_vars))]
+    return []
+
+  def _get_optimizer(self, optimizer, default_optimizer, default_learning_rate):
+    if optimizer is None:
+      optimizer = default_optimizer
+    if isinstance(optimizer, six.string_types):
+      optimizer = layers.OPTIMIZER_CLS_NAMES[optimizer](
+          learning_rate=default_learning_rate)
+    return optimizer
+
+
+class DNNLinearCombinedClassifier(_DNNLinearCombinedBaseEstimator):
+  """A classifier for TensorFlow Linear and DNN joined training models.
+
+  Example:
+    ```
+    installed_app_id = sparse_column_with_hash_bucket("installed_id", 1e6)
+    impression_app_id = sparse_column_with_hash_bucket("impression_id", 1e6)
+
+    installed_x_impression = crossed_column(
+        [installed_app_id, impression_app_id])
+
+    installed_emb = embedding_column(installed_app_id, dimension=16,
+                                     combiner="sum")
+    impression_emb = embedding_column(impression_app_id, dimension=16,
+                                      combiner="sum")
+
+    estimator = DNNLinearCombinedClassifier(
+        # common settings
+        n_classes, weight_column_name,
+        # wide settings
+        linear_feature_columns=[installed_x_impression],
+        linear_optimizer=tf.train.FtrlOptimizer(...),
+        # deep settings
+        dnn_feature_columns=[installed_emb, impression_emb],
+        dnn_hidden_units=[1000, 500, 100],
+        dnn_optimizer=tf.train.AdagradOptimizer(...))
+
+    # Input builders
+    def input_fn_train: # returns X, Y
+      ...
+    def input_fn_eval: # returns X, Y
+      ...
+    estimator.train(input_fn_train)
+    estimator.evaluate(input_fn_eval)
+    estimator.predict(x)
+    ```
+
+    Input of `fit`, `train`, and `evaluate` should have following features,
+      otherwise there will be a `KeyError`:
+        if `weight_column_name` is not `None`, a feature with
+          `key=weight_column_name` whose value is a `Tensor`.
+        for each `column` in `dnn_feature_columns` + `linear_feature_columns`:
+        - if `column` is a `SparseColumn`, a feature with `key=column.name`
+          whose `value` is a `SparseTensor`.
+        - if `column` is a `RealValuedColumn, a feature with `key=column.name`
+          whose `value` is a `Tensor`.
+
+  Parameters:
+    model_dir: Directory to save model parameters, graph and etc.
+    n_classes: number of target classes. Default is binary classification.
+    weight_column_name: A string defining feature column name representing
+      weights. It is used to down weight or boost examples during training. It
+      will be multiplied by the loss of the example.
+    linear_feature_columns: An iterable containing all the feature columns used
+      by linear part of the model. All items in the set must be instances of
+      classes derived from `FeatureColumn`.
+    linear_optimizer: An instance of `tf.Optimizer` used to apply gradients to
+      the linear part of the model. If `None`, will use a FTRL optimizer.
+    dnn_feature_columns: An iterable containing all the feature columns used by
+      deep part of the model. All items in the set must be instances of
+      classes derived from `FeatureColumn`.
+    dnn_hidden_units: List of hidden units per layer. All layers are fully
+      connected.
+    dnn_optimizer: An instance of `tf.Optimizer` used to apply gradients to the
+      deep part of the model. If `None`, will use an Adagrad optimizer.
+    dnn_activation_fn: Activation function applied to each layer. If `None`,
+      will use `tf.nn.relu`.
+
+    Raises:
+      ValueError: If both linear_feature_columns and dnn_features_columns are
+        empty at the same time.
+      ValueError: If both n_classes < 2.
+  """
+
+  def __init__(self,
+               model_dir=None,
+               n_classes=2,
+               weight_column_name=None,
+               linear_feature_columns=None,
+               linear_optimizer=None,
+               dnn_feature_columns=None,
+               dnn_optimizer=None,
+               dnn_hidden_units=None,
+               dnn_activation_fn=nn.relu):
+    if n_classes < 2:
+      raise ValueError("n_classes should be greater than 1. Given: {}".format(
+          n_classes))
+
+    super(DNNLinearCombinedClassifier, self).__init__(
+        model_dir=model_dir,
+        n_classes=n_classes,
+        weight_column_name=weight_column_name,
+        linear_feature_columns=linear_feature_columns,
+        linear_optimizer=linear_optimizer,
+        dnn_feature_columns=dnn_feature_columns,
+        dnn_optimizer=dnn_optimizer,
+        dnn_hidden_units=dnn_hidden_units,
+        dnn_activation_fn=dnn_activation_fn)
+
+
+class DNNLinearCombinedRegressor(_DNNLinearCombinedBaseEstimator):
+  """A regressor for TensorFlow Linear and DNN joined training models.
+
+  Example:
+    ```
+    installed_app_id = sparse_column_with_hash_bucket("installed_id", 1e6)
+    impression_app_id = sparse_column_with_hash_bucket("impression_id", 1e6)
+
+    installed_x_impression = crossed_column(
+        [installed_app_id, impression_app_id])
+
+    installed_emb = embedding_column(installed_app_id, dimension=16,
+                                     combiner="sum")
+    impression_emb = embedding_column(impression_app_id, dimension=16,
+                                      combiner="sum")
+
+    estimator = DNNLinearCombinedClassifier(
+        # common settings
+        n_classes, weight_column_name,
+        # wide settings
+        linear_feature_columns=[installed_x_impression],
+        linear_optimizer=tf.train.FtrlOptimizer(...),
+        # deep settings
+        dnn_feature_columns=[installed_emb, impression_emb],
+        dnn_hidden_units=[1000, 500, 100],
+        dnn_optimizer=tf.train.AdagradOptimizer(...))
+
+    # Input builders
+    def input_fn_train: # returns X, Y
+      ...
+    def input_fn_eval: # returns X, Y
+      ...
+    estimator.train(input_fn_train)
+    estimator.evaluate(input_fn_eval)
+    estimator.predict(x)
+    ```
+
+    Input of `fit`, `train`, and `evaluate` should have following features,
+      otherwise there will be a `KeyError`:
+        if `weight_column_name` is not `None`, a feature with
+          `key=weight_column_name` whose value is a `Tensor`.
+        for each `column` in `dnn_feature_columns` + `linear_feature_columns`:
+        - if `column` is a `SparseColumn`, a feature with `key=column.name`
+          whose `value` is a `SparseTensor`.
+        - if `column` is a `RealValuedColumn, a feature with `key=column.name`
+          whose `value` is a `Tensor`.
+
+  Parameters:
+    model_dir: Directory to save model parameters, graph and etc.
+    weight_column_name: A string defining feature column name representing
+      weights. It is used to down weight or boost examples during training. It
+      will be multiplied by the loss of the example.
+    linear_feature_columns: An iterable containing all the feature columns used
+      by linear part of the model. All items in the set must be instances of
+      classes derived from `FeatureColumn`.
+    linear_optimizer: An instance of `tf.Optimizer` used to apply gradients to
+      the linear part of the model. If `None`, will use a FTRL optimizer.
+    dnn_feature_columns: An iterable containing all the feature columns used by
+      deep part of the model. All items in the set must be instances of
+      classes derived from `FeatureColumn`.
+    dnn_hidden_units: List of hidden units per layer. All layers are fully
+      connected.
+    dnn_optimizer: An instance of `tf.Optimizer` used to apply gradients to the
+      deep part of the model. If `None`, will use an Adagrad optimizer.
+    dnn_activation_fn: Activation function applied to each layer. If None, will
+      use `tf.nn.relu`.
+
+    Raises:
+      ValueError: If both linear_feature_columns and dnn_features_columns are
+        empty at the same time.
+  """
+
+  def __init__(self,
+               model_dir=None,
+               weight_column_name=None,
+               linear_feature_columns=None,
+               linear_optimizer=None,
+               dnn_feature_columns=None,
+               dnn_optimizer=None,
+               dnn_hidden_units=None,
+               dnn_activation_fn=nn.relu):
+    super(DNNLinearCombinedRegressor, self).__init__(
+        model_dir=model_dir,
+        n_classes=0,
+        weight_column_name=weight_column_name,
+        linear_feature_columns=linear_feature_columns,
+        linear_optimizer=linear_optimizer,
+        dnn_feature_columns=dnn_feature_columns,
+        dnn_optimizer=dnn_optimizer,
+        dnn_hidden_units=dnn_hidden_units,
+        dnn_activation_fn=dnn_activation_fn)

tensorflow/contrib/learn/python/learn/estimators/dnn_linear_combined_test.py

@@ -0,0 +1,254 @@
+#  Copyright 2015 Google Inc. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+"""Tests for DNNLinearCombinedEstimators."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import tensorflow as tf
+
+
+def _get_quantile_based_buckets(feature_values, num_buckets):
+  quantiles = np.percentile(
+      np.array(feature_values), ([100 * (i + 1.) / (num_buckets + 1.)
+                                  for i in range(num_buckets)]))
+  return list(quantiles)
+
+
+def _prepare_iris_data_for_logistic_regression():
+  # Converts iris data to a logistic regression problem.
+  iris = tf.contrib.learn.datasets.load_iris()
+  ids = np.where((iris.target == 0) | (iris.target == 1))
+  iris = tf.contrib.learn.datasets.base.Dataset(data=iris.data[ids],
+                                                target=iris.target[ids])
+  return iris
+
+
+def _iris_input_fn():
+  iris = _prepare_iris_data_for_logistic_regression()
+  return {
+      'feature': tf.constant(iris.data, dtype=tf.float32)
+  }, tf.constant(iris.target, shape=[100, 1], dtype=tf.int32)
+
+
+class DNNLinearCombinedClassifierTest(tf.test.TestCase):
+
+  def testLogisticRegression_MatrixData(self):
+    """Tests binary classification using matrix data as input."""
+    iris = _prepare_iris_data_for_logistic_regression()
+    cont_features = [
+        tf.contrib.layers.real_valued_column('feature', dimension=4)]
+    bucketized_feature = [tf.contrib.layers.bucketized_column(
+        cont_features[0], _get_quantile_based_buckets(iris.data, 10))]
+
+    classifier = tf.contrib.learn.DNNLinearCombinedClassifier(
+        linear_feature_columns=bucketized_feature,
+        dnn_feature_columns=cont_features,
+        dnn_hidden_units=[3, 3])
+
+    classifier.train(_iris_input_fn, steps=100)
+    scores = classifier.evaluate(input_fn=_iris_input_fn, steps=100)
+    self.assertGreater(scores['accuracy'], 0.9)
+
+  def testLogisticRegression_TensorData(self):
+    """Tests binary classification using Tensor data as input."""
+    def _input_fn():
+      iris = _prepare_iris_data_for_logistic_regression()
+      features = {}
+      for i in range(4):
+        # The following shows how to provide the Tensor data for
+        # RealValuedColumns.
+        features.update({
+            str(i): tf.reshape(tf.constant(iris.data[:, i], dtype=tf.float32),
+                               [-1, 1])})
+      # The following shows how to provide the SparseTensor data for
+      # a SparseColumn.
+      features['dummy_sparse_column'] = tf.SparseTensor(
+          values=['en', 'fr', 'zh'],
+          indices=[[0, 0], [0, 1], [60, 0]],
+          shape=[len(iris.target), 2])
+      target = tf.reshape(tf.constant(iris.target, dtype=tf.float32), [-1, 1])
+      return features, target
+
+    iris = _prepare_iris_data_for_logistic_regression()
+    cont_features = [tf.contrib.layers.real_valued_column(str(i))
+                     for i in range(4)]
+    linear_features = [
+        tf.contrib.layers.bucketized_column(
+            cont_features[i], _get_quantile_based_buckets(iris.data[:, str(i)],
+                                                          10)) for i in range(4)
+    ]
+    linear_features.append(tf.contrib.layers.sparse_column_with_hash_bucket(
+        'dummy_sparse_column', hash_bucket_size=100))
+
+    classifier = tf.contrib.learn.DNNLinearCombinedClassifier(
+        linear_feature_columns=linear_features,
+        dnn_feature_columns=cont_features,
+        dnn_hidden_units=[3, 3])
+
+    classifier.train(input_fn=_input_fn, steps=100)
+    scores = classifier.evaluate(input_fn=_input_fn, steps=100)
+    self.assertGreater(scores['accuracy'], 0.9)
+
+  def testMultiClass(self):
+    """Tests multi-class classification using matrix data as input.
+
+    Please see testLogisticRegression_TensorData() for how to use Tensor
+    data as input instead.
+    """
+    iris = tf.contrib.learn.datasets.load_iris()
+    cont_features = [
+        tf.contrib.layers.real_valued_column('feature', dimension=4)]
+    bucketized_features = [
+        tf.contrib.layers.bucketized_column(
+            cont_features[0], _get_quantile_based_buckets(iris.data, 10))]
+
+    classifier = tf.contrib.learn.DNNLinearCombinedClassifier(
+        n_classes=3,
+        linear_feature_columns=bucketized_features,
+        dnn_feature_columns=cont_features,
+        dnn_hidden_units=[3, 3])
+
+    classifier.train(_iris_input_fn, steps=100)
+    scores = classifier.evaluate(input_fn=_iris_input_fn, steps=100)
+    self.assertGreater(scores['accuracy'], 0.9)
+
+  def testWeightColumn(self):
+    """Tests weight column."""
+
+    def _input_fn_train():
+      # Create 4 rows, one of them (y = x), three of them (y=Not(x))
+      # First row has more weight than others. Model should fit (y=x) better
+      # than (y=Not(x)) due to the relative higher weight of the first row.
+      target = tf.constant([[1], [0], [0], [0]])
+      features = {
+          'x': tf.ones(shape=[4, 1], dtype=tf.float32),
+          'w': tf.constant([[100.], [3.], [2.], [2.]])
+      }
+      return features, target
+
+    def _input_fn_eval():
+      # Create 4 rows (y = x)
+      target = tf.constant([[1], [1], [1], [1]])
+      features = {
+          'x': tf.ones(shape=[4, 1], dtype=tf.float32),
+          'w': tf.constant([[1.], [1.], [1.], [1.]])
+      }
+      return features, target
+
+    classifier = tf.contrib.learn.DNNLinearCombinedClassifier(
+        weight_column_name='w',
+        linear_feature_columns=[tf.contrib.layers.real_valued_column('x')],
+        dnn_feature_columns=[tf.contrib.layers.real_valued_column('x')],
+        dnn_hidden_units=[3, 3])
+
+    classifier.train(input_fn=_input_fn_train, steps=100)
+    scores = classifier.evaluate(input_fn=_input_fn_eval,
+                                 steps=100)
+    # If there is no weight column, model should learn y=Not(x). All examples in
+    # eval data set are y=x. So if weight column is ignored, then accuracy
+    # should be zero.
+    self.assertGreater(scores['accuracy'], 0.9)
+
+  def testEvaluationShouldUseWeightColumn(self):
+    """Tests weight column in evaluation."""
+
+    def _input_fn_train():
+      # Create 4 rows, one of them (y = x), three of them (y=Not(x))
+      # First row has more weight than others. Model should fit (y=x) better
+      # than (y=Not(x)) due to the relative higher weight of the first row.
+      target = tf.constant([[1], [0], [0], [0]])
+      features = {
+          'x': tf.ones(shape=[4, 1], dtype=tf.float32),
+          'w': tf.constant([[100.], [3.], [2.], [2.]])
+      }
+      return features, target
+
+    classifier = tf.contrib.learn.DNNLinearCombinedClassifier(
+        weight_column_name='w',
+        linear_feature_columns=[tf.contrib.layers.real_valued_column('x')],
+        dnn_feature_columns=[tf.contrib.layers.real_valued_column('x')],
+        dnn_hidden_units=[3, 3])
+
+    classifier.train(input_fn=_input_fn_train, steps=100)
+    scores = classifier.evaluate(input_fn=_input_fn_train, steps=100)
+    # If weight column is ignored, then accuracy should be 0.25. If it's not
+    # ignored, then it should be greater than 0.6.
+    self.assertGreater(scores['accuracy'], 0.6)
+
+  def testCustomOptimizerByObject(self):
+    """Tests binary classification using matrix data as input."""
+    iris = _prepare_iris_data_for_logistic_regression()
+    cont_features = [
+        tf.contrib.layers.real_valued_column('feature', dimension=4)]
+    bucketized_features = [
+        tf.contrib.layers.bucketized_column(
+            cont_features[0], _get_quantile_based_buckets(iris.data, 10))]
+
+    classifier = tf.contrib.learn.DNNLinearCombinedClassifier(
+        linear_feature_columns=bucketized_features,
+        linear_optimizer=tf.train.FtrlOptimizer(learning_rate=0.1),
+        dnn_feature_columns=cont_features,
+        dnn_hidden_units=[3, 3],
+        dnn_optimizer=tf.train.AdagradOptimizer(learning_rate=0.1))
+
+    classifier.train(_iris_input_fn, steps=100)
+    scores = classifier.evaluate(input_fn=_iris_input_fn, steps=100)
+    self.assertGreater(scores['accuracy'], 0.9)
+
+  def testCustomOptimizerByString(self):
+    """Tests binary classification using matrix data as input."""
+    iris = _prepare_iris_data_for_logistic_regression()
+    cont_features = [
+        tf.contrib.layers.real_valued_column('feature', dimension=4)]
+    bucketized_features = [
+        tf.contrib.layers.bucketized_column(
+            cont_features[0], _get_quantile_based_buckets(iris.data, 10))]
+
+    classifier = tf.contrib.learn.DNNLinearCombinedClassifier(
+        linear_feature_columns=bucketized_features,
+        linear_optimizer='Ftrl',
+        dnn_feature_columns=cont_features,
+        dnn_hidden_units=[3, 3],
+        dnn_optimizer='Adagrad')
+
+    classifier.train(_iris_input_fn, steps=100)
+    scores = classifier.evaluate(input_fn=_iris_input_fn, steps=100)
+    self.assertGreater(scores['accuracy'], 0.9)
+
+
+class DNNLinearCombinedRegressorTest(tf.test.TestCase):
+
+  def testRegression(self):
+    """Tests a regression problem."""
+
+    def _input_fn_train():
+      # Create 4 rows of (y = x)
+      target = tf.constant([[100.], [3.], [2.], [2.]])
+      features = {'x': tf.constant([[100.], [3.], [2.], [2.]])}
+      return features, target
+
+    classifier = tf.contrib.learn.DNNLinearCombinedRegressor(
+        linear_feature_columns=[tf.contrib.layers.real_valued_column('x')],
+        dnn_feature_columns=[tf.contrib.layers.real_valued_column('x')],
+        dnn_hidden_units=[3, 3])
+
+    classifier.train(input_fn=_input_fn_train, steps=100)
+    classifier.evaluate(input_fn=_input_fn_train, steps=1)
+
+
+if __name__ == '__main__':
+  tf.test.main()

tensorflow/contrib/learn/python/learn/estimators/dnn_test.py

@@ -0,0 +1,77 @@
+#  Copyright 2015 Google Inc. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+"""Tests for DNNEstimators."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow as tf
+
+
+def _iris_input_fn():
+  iris = tf.contrib.learn.datasets.load_iris()
+  return {
+      'feature': tf.constant(iris.data, dtype=tf.float32)
+  }, tf.constant(iris.target, shape=[150, 1], dtype=tf.int32)
+
+
+class DNNClassifierTest(tf.test.TestCase):
+
+  def testMultiClass(self):
+    """Tests multi-class classification using matrix data as input."""
+    cont_features = [
+        tf.contrib.layers.real_valued_column('feature', dimension=4)]
+
+    classifier = tf.contrib.learn.DNNClassifier(n_classes=3,
+                                                feature_columns=cont_features,
+                                                hidden_units=[3, 3])
+
+    classifier.train(_iris_input_fn, steps=1000)
+    classifier.evaluate(input_fn=_iris_input_fn, steps=100)
+    # TODO(ispir): Enable accuracy check after resolving the randomness issue.
+    # self.assertGreater(scores['accuracy/mean'], 0.6)
+
+
+class DNNRegressorTest(tf.test.TestCase):
+
+  def testRegression(self):
+    """Tests multi-class classification using matrix data as input."""
+    cont_features = [
+        tf.contrib.layers.real_valued_column('feature', dimension=4)]
+
+    regressor = tf.contrib.learn.DNNRegressor(feature_columns=cont_features,
+                                              hidden_units=[3, 3])
+
+    regressor.train(_iris_input_fn, steps=1000)
+    regressor.evaluate(input_fn=_iris_input_fn, steps=100)
+
+
+def boston_input_fn():
+  boston = tf.contrib.learn.datasets.load_boston()
+  features = tf.cast(tf.reshape(tf.constant(boston.data), [-1, 13]), tf.float32)
+  target = tf.cast(tf.reshape(tf.constant(boston.target), [-1, 1]), tf.float32)
+  return features, target
+
+
+class InferedColumnTest(tf.test.TestCase):
+
+  def testTrain(self):
+    est = tf.contrib.learn.DNNRegressor(hidden_units=[3, 3])
+    est.train(input_fn=boston_input_fn, steps=1)
+    _ = est.evaluate(input_fn=boston_input_fn, steps=1)
+
+
+if __name__ == '__main__':
+  tf.test.main()

tensorflow/contrib/learn/python/learn/estimators/estimator.py

@@ -21,12 +21,16 @@ import abc
 import os
 import tempfile
 import time
+import types
 
+import numpy as np
 import six
 
 from tensorflow.contrib import framework as contrib_framework
 from tensorflow.contrib import layers
 from tensorflow.contrib import losses
+from tensorflow.contrib.learn.python.learn import graph_actions
+from tensorflow.contrib.learn.python.learn import monitors as monitors_lib
 from tensorflow.contrib.learn.python.learn.estimators import _sklearn as sklearn
 from tensorflow.contrib.learn.python.learn.estimators import run_config
 from tensorflow.contrib.learn.python.learn.estimators import tensor_signature
@@ -37,6 +41,8 @@ from tensorflow.contrib.learn.python.learn.io import data_feeder
 
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import random_seed
+from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import logging_ops
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.training import device_setter
 from tensorflow.python.training import saver
@@ -89,7 +95,7 @@ def _get_predict_input_fn(x, batch_size):
 
   df = data_feeder.setup_train_data_feeder(x, None,
                                            n_classes=None,
-                                           batch_size=batch_size)
+                                           batch_size=batch_size, epochs=1)
   return df.input_builder, df.get_feed_dict_fn()
 
 
@@ -112,7 +118,7 @@ class BaseEstimator(sklearn.BaseEstimator):
   # TODO(wicke): Remove this once launcher takes over config functionality
   _Config = run_config.RunConfig  # pylint: disable=invalid-name
 
-  def __init__(self, model_dir=None):
+  def __init__(self, model_dir=None, config=None):
     # Model directory.
     self._model_dir = model_dir
     if self._model_dir is None:
@@ -121,7 +127,10 @@ class BaseEstimator(sklearn.BaseEstimator):
                    self._model_dir)
 
     # Create a run configuration
-    self._config = BaseEstimator._Config()
+    if config is None:
+      self._config = BaseEstimator._Config()
+    else:
+      self._config = config
 
     # Set device function depending if there are replicas or not.
     if self._config.num_ps_replicas > 0:
@@ -136,6 +145,12 @@ class BaseEstimator(sklearn.BaseEstimator):
     self._features_info = None
     self._targets_info = None
 
+    self._graph = None
+
+  @property
+  def model_dir(self):
+    return self._model_dir
+
   @abc.abstractproperty
   def _get_train_ops(self, features, targets):
     """Method that builds model graph and returns trainer ops.
@@ -204,7 +219,7 @@ class BaseEstimator(sklearn.BaseEstimator):
     """
     return {}
 
-  def fit(self, x, y, steps, batch_size=32, monitor=None):
+  def fit(self, x, y, steps, batch_size=32, monitors=None):
     """Trains a model given training data X and y.
 
     Args:
@@ -216,8 +231,8 @@ class BaseEstimator(sklearn.BaseEstimator):
          (class labels in classification, real numbers in regression).
       steps: number of steps to train model for.
       batch_size: minibatch size to use on the input, defaults to 32.
-      monitor: monitor object to print training progress and invoke
-               early stopping.
+      monitors: List of `BaseMonitor` subclass instances. Used for callbacks
+                inside the training loop.
 
     Returns:
       Returns self.
@@ -226,24 +241,24 @@ class BaseEstimator(sklearn.BaseEstimator):
     return self._train_model(input_fn=input_fn,
                              feed_fn=feed_fn,
                              steps=steps,
-                             monitor=monitor)
+                             monitors=monitors)
 
-  def train(self, input_fn, steps, monitor=None):
+  def train(self, input_fn, steps, monitors=None):
     """Trains a model given input builder function.
 
     Args:
       input_fn: Input builder function, returns tuple of dicts or
                 dict and Tensor.
       steps: number of steps to train model for.
-      monitor: monitor object to print training progress and invoke
-               early stopping.
+      monitors: List of `BaseMonitor` subclass instances. Used for callbacks
+                inside the training loop.
 
     Returns:
       Returns self.
     """
-    return self._train_model(input_fn=input_fn, steps=steps, monitor=monitor)
+    return self._train_model(input_fn=input_fn, steps=steps, monitors=monitors)
 
-  def partial_fit(self, x, y, steps=1, batch_size=32, monitor=None):
+  def partial_fit(self, x, y, steps=1, batch_size=32, monitors=None):
     """Incremental fit on a batch of samples.
 
     This method is expected to be called several times consecutively
@@ -263,8 +278,8 @@ class BaseEstimator(sklearn.BaseEstimator):
         (class label in classification, real numbers in regression).
       steps: number of steps to train model for.
       batch_size: minibatch size to use on the input, defaults to 32.
-      monitor: Monitor object to print training progress and invoke
-               early stopping.
+      monitors: List of `BaseMonitor` subclass instances. Used for callbacks
+                inside the training loop.
 
     Returns:
       Returns self.
@@ -273,7 +288,7 @@ class BaseEstimator(sklearn.BaseEstimator):
     return self._train_model(input_fn=input_fn,
                              feed_fn=feed_fn,
                              steps=steps,
-                             monitor=monitor)
+                             monitors=monitors)
 
   def evaluate(self, x=None, y=None, input_fn=None, feed_fn=None,
                batch_size=32, steps=100, metrics=None):
@@ -323,7 +338,7 @@ class BaseEstimator(sklearn.BaseEstimator):
 
     Args:
       x: features.
-      batch_size: OVerride default batch size.
+      batch_size: Override default batch size.
 
     Returns:
       Numpy array of predicted probabilities.
@@ -338,20 +353,24 @@ class BaseEstimator(sklearn.BaseEstimator):
                          (str(features), str(self._features_info)))
     else:
       self._features_info = tensor_signature.create_signatures(features)
-    if self._targets_info is not None:
-      if not tensor_signature.tensors_compatible(targets, self._targets_info):
-        raise ValueError('Targets are incompatible with given information. '
-                         'Given targets: %s, required signatures: %s.' %
-                         (str(targets), str(self._targets_info)))
-    else:
-      self._targets_info = tensor_signature.create_signatures(targets)
+    if targets is not None:
+      if self._targets_info is not None:
+        if not tensor_signature.tensors_compatible(targets, self._targets_info):
+          raise ValueError('Targets are incompatible with given information. '
+                           'Given targets: %s, required signatures: %s.' %
+                           (str(targets), str(self._targets_info)))
+      else:
+        self._targets_info = tensor_signature.create_signatures(targets)
 
   def _train_model(self,
                    input_fn,
                    steps,
                    feed_fn=None,
+                   init_op=None,
+                   init_feed_fn=None,
+                   init_fn=None,
                    device_fn=None,
-                   monitor=None,
+                   monitors=None,
                    log_every_steps=100,
                    fail_on_nan_loss=True):
     if self._config.execution_mode not in ('all', 'train'):
@@ -369,24 +388,42 @@ class BaseEstimator(sklearn.BaseEstimator):
     # Device allocation
     device_fn = device_fn or self._device_fn
 
-    with ops.Graph().as_default() as g, g.device(device_fn):
+    self._graph = ops.Graph()
+    with self._graph.as_default() as g, g.device(device_fn):
       random_seed.set_random_seed(self._config.tf_random_seed)
       global_step = contrib_framework.create_global_step(g)
       features, targets = input_fn()
       self._check_inputs(features, targets)
       train_op, loss_op = self._get_train_ops(features, targets)
+
+      # Add default monitors.
+      if monitors is None:
+        monitors = []
+      monitors += monitors_lib.get_default_monitors(
+          loss_op=loss_op,
+          summary_op=logging_ops.get_summary_op(),
+          save_summary_steps=100)
+
+      # Setup monitors.
+      for monitor in monitors:
+        monitor.set_estimator(self)
+
       return train(
           graph=g,
           output_dir=self._model_dir,
           train_op=train_op,
           loss_op=loss_op,
           global_step_tensor=global_step,
+          init_op=init_op,
+          init_feed_dict=init_feed_fn() if init_feed_fn is not None else None,
+          init_fn=init_fn,
           log_every_steps=log_every_steps,
           supervisor_is_chief=(self._config.task == 0),
           supervisor_master=self._config.master,
           feed_fn=feed_fn,
           max_steps=steps,
-          fail_on_nan_loss=fail_on_nan_loss)
+          fail_on_nan_loss=fail_on_nan_loss,
+          monitors=monitors)
 
   def _evaluate_model(self, input_fn, steps, feed_fn=None, metrics=None):
     if self._config.execution_mode not in ('all', 'evaluate', 'eval_evalset'):
@@ -413,22 +450,41 @@ class BaseEstimator(sklearn.BaseEstimator):
           max_steps=steps)
       return eval_results
 
-  def _infer_model(self, x, batch_size=None, axis=None, proba=False):
+  def _infer_model(self,
+                   x=None, input_fn=None, feed_fn=None,
+                   batch_size=None, axis=None, proba=False):
     # Converts inputs into tf.DataFrame / tf.Series.
     batch_size = -1 if batch_size is None else batch_size
-    input_fn, feed_fn = _get_predict_input_fn(x, batch_size)
+    if x is not None:
+      input_fn, feed_fn = _get_predict_input_fn(x, batch_size)
 
     checkpoint_path = saver.latest_checkpoint(self._model_dir)
     with ops.Graph().as_default() as g:
       random_seed.set_random_seed(self._config.tf_random_seed)
       contrib_framework.create_global_step(g)
       features, _ = input_fn()
-      feed_dict = feed_fn() if feed_fn is not None else None
       predictions = self._get_predict_ops(features)
       if not isinstance(predictions, dict):
         predictions = {'predictions': predictions}
       # TODO(ipolosukhin): Support batching
-      return infer(checkpoint_path, predictions, feed_dict=feed_dict)
+      if feed_fn is None:
+        return infer(checkpoint_path, predictions)
+      preds = {}
+      while True:
+        try:
+          feed_dict = feed_fn()
+        except StopIteration:
+          break
+        if feed_dict is None:
+          break
+        outputs = infer(checkpoint_path, predictions, feed_dict=feed_dict)
+        for key in outputs:
+          if key not in preds:
+            preds[key] = []
+          preds[key].append(outputs[key])
+      for key in preds:
+        preds[key] = np.concatenate(preds[key], axis=0)
+      return preds
 
 
 class Estimator(BaseEstimator):
@@ -449,16 +505,18 @@ class Estimator(BaseEstimator):
                  (like tf.train.GradientDescentOptimizer).
     clip_gradients: clip_norm value for call to `clip_by_global_norm`. None
                     denotes no gradient clipping.
+    config: Configuration object.
   """
 
   def __init__(self,
                model_fn=None,
                model_dir=None,
                classification=True,
-               learning_rate=0.01,
-               optimizer='SGD',
-               clip_gradients=None):
-    super(Estimator, self).__init__(model_dir=model_dir)
+               learning_rate=0.1,
+               optimizer='Adagrad',
+               clip_gradients=None,
+               config=None):
+    super(Estimator, self).__init__(model_dir=model_dir, config=config)
 
     self._model_fn = model_fn
     self._classification = classification
@@ -486,12 +544,25 @@ class Estimator(BaseEstimator):
       Tuple of train `Operation` and loss `Tensor`.
     """
     _, loss = self._model_fn(features, targets, ModeKeys.TRAIN)
+    # TODO(ipolosukhin): Move this to TensorFlowEstimator when
+    # moving out training.
+    if isinstance(self.learning_rate, types.FunctionType):
+      learning_rate = self.learning_rate(contrib_framework.get_global_step())
+    else:
+      learning_rate = self.learning_rate
+    if isinstance(self.optimizer, types.FunctionType):
+      optimizer = self.optimizer(learning_rate)
+    else:
+      optimizer = self.optimizer
     train_op = layers.optimize_loss(
         loss,
         contrib_framework.get_global_step(),
-        learning_rate=self.learning_rate,
-        optimizer=self.optimizer,
+        learning_rate=learning_rate,
+        optimizer=optimizer,
         clip_gradients=self.clip_gradients)
+    # Add update ops.
+    train_op = control_flow_ops.group(
+        train_op, *ops.get_collection('update_ops'))
     return train_op, loss
 
   def _get_eval_ops(self, features, targets, metrics):

tensorflow/contrib/learn/python/learn/estimators/estimator_test.py

@@ -37,6 +37,24 @@ def linear_model_fn(features, target, unused_mode):
   return tf.contrib.learn.models.linear_regression_zero_init(features, target)
 
 
+class CheckCallsMonitor(tf.contrib.learn.monitors.BaseMonitor):
+
+  def __init__(self):
+    self.calls = None
+    self.expect_calls = None
+
+  def begin(self, max_steps):
+    self.calls = 0
+    self.expect_calls = max_steps
+
+  def step_end(self, step, outputs):
+    self.calls += 1
+    return False
+
+  def end(self):
+    assert self.calls == self.expect_calls
+
+
 class EstimatorTest(tf.test.TestCase):
 
   def testTrain(self):
@@ -64,6 +82,12 @@ class EstimatorTest(tf.test.TestCase):
     with self.assertRaises(ValueError):
       est.train(input_fn=other_input_fn, steps=1)
 
+  def testMonitors(self):
+    est = tf.contrib.learn.Estimator(model_fn=linear_model_fn,
+                                     classification=False)
+    est.train(input_fn=boston_input_fn, steps=21,
+              monitors=[CheckCallsMonitor()])
+
 
 if __name__ == '__main__':
   tf.test.main()

tensorflow/contrib/learn/python/learn/estimators/linear.py

@@ -16,11 +16,151 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-from tensorflow.contrib.learn.python.learn.estimators import _sklearn
-from tensorflow.contrib.learn.python.learn.estimators.base import TensorFlowEstimator
+from tensorflow.contrib import layers
 from tensorflow.contrib.learn.python.learn import models
+from tensorflow.contrib.learn.python.learn.estimators import _sklearn
+from tensorflow.contrib.learn.python.learn.estimators import dnn_linear_combined
+from tensorflow.contrib.learn.python.learn.estimators.base import TensorFlowEstimator
 
 
+class LinearClassifier(dnn_linear_combined.DNNLinearCombinedClassifier):
+  """Linear classifier model.
+
+    Example:
+    ```
+    installed_app_id = sparse_column_with_hash_bucket("installed_id", 1e6)
+    impression_app_id = sparse_column_with_hash_bucket("impression_id", 1e6)
+
+    installed_x_impression = crossed_column(
+        [installed_app_id, impression_app_id])
+
+    estimator = LinearClassifier(
+        feature_columns=[impression_app_id, installed_x_impression])
+
+    # Input builders
+    def input_fn_train: # returns X, Y
+      ...
+    def input_fn_eval: # returns X, Y
+      ...
+    estimator.train(input_fn_train)
+    estimator.evaluate(input_fn_eval)
+    estimator.predict(x)
+    ```
+
+    Input of `fit`, `train`, and `evaluate` should have following features,
+      otherwise there will be a `KeyError`:
+        if `weight_column_name` is not `None`, a feature with
+          `key=weight_column_name` whose value is a `Tensor`.
+        for each `column` in `feature_columns`:
+        - if `column` is a `SparseColumn`, a feature with `key=column.name`
+          whose `value` is a `SparseTensor`.
+        - if `column` is a `RealValuedColumn, a feature with `key=column.name`
+          whose `value` is a `Tensor`.
+        - if `feauture_columns` is None, then `input` must contains only real
+          valued `Tensor`.
+
+
+  Parameters:
+    feature_columns: An iterable containing all the feature columns used by the
+      model. All items in the set should be instances of classes derived from
+      `FeatureColumn`.
+    model_dir: Directory to save model parameters, graph and etc.
+    n_classes: number of target classes. Default is binary classification.
+    weight_column_name: A string defining feature column name representing
+      weights. It is used to down weight or boost examples during training. It
+      will be multiplied by the loss of the example.
+    optimizer: An instance of `tf.Optimizer` used to train the model. If `None`,
+      will use an Ftrl optimizer.
+  """
+
+  def __init__(self,
+               feature_columns=None,
+               model_dir=None,
+               n_classes=2,
+               weight_column_name=None,
+               optimizer=None):
+    super(LinearClassifier, self).__init__(
+        model_dir=model_dir,
+        n_classes=n_classes,
+        weight_column_name=weight_column_name,
+        linear_feature_columns=feature_columns,
+        linear_optimizer=optimizer)
+
+  def _get_train_ops(self, features, targets):
+    """See base class."""
+    if self._linear_feature_columns is None:
+      self._linear_feature_columns = layers.infer_real_valued_columns(features)
+    return super(LinearClassifier, self)._get_train_ops(features, targets)
+
+
+class LinearRegressor(dnn_linear_combined.DNNLinearCombinedRegressor):
+  """Linear regressor model.
+
+    Example:
+    ```
+    installed_app_id = sparse_column_with_hash_bucket("installed_id", 1e6)
+    impression_app_id = sparse_column_with_hash_bucket("impression_id", 1e6)
+
+    installed_x_impression = crossed_column(
+        [installed_app_id, impression_app_id])
+
+    estimator = LinearRegressor(
+        feature_columns=[impression_app_id, installed_x_impression])
+
+    # Input builders
+    def input_fn_train: # returns X, Y
+      ...
+    def input_fn_eval: # returns X, Y
+      ...
+    estimator.train(input_fn_train)
+    estimator.evaluate(input_fn_eval)
+    estimator.predict(x)
+    ```
+
+    Input of `fit`, `train`, and `evaluate` should have following features,
+      otherwise there will be a KeyError:
+        if `weight_column_name` is not None:
+          key=weight_column_name, value=a `Tensor`
+        for column in `feature_columns`:
+        - if isinstance(column, `SparseColumn`):
+            key=column.name, value=a `SparseTensor`
+        - if isinstance(column, `RealValuedColumn`):
+            key=column.name, value=a `Tensor`
+        - if `feauture_columns` is None:
+            input must contains only real valued `Tensor`.
+
+  Parameters:
+    feature_columns: An iterable containing all the feature columns used by the
+      model. All items in the set should be instances of classes derived from
+      `FeatureColumn`.
+    model_dir: Directory to save model parameters, graph and etc.
+    weight_column_name: A string defining feature column name representing
+      weights. It is used to down weight or boost examples during training. It
+      will be multiplied by the loss of the example.
+    optimizer: An instance of `tf.Optimizer` used to train the model. If `None`,
+      will use an Ftrl optimizer.
+  """
+
+  def __init__(self,
+               feature_columns=None,
+               model_dir=None,
+               n_classes=2,
+               weight_column_name=None,
+               optimizer=None):
+    super(LinearRegressor, self).__init__(
+        model_dir=model_dir,
+        weight_column_name=weight_column_name,
+        linear_feature_columns=feature_columns,
+        linear_optimizer=optimizer)
+
+  def _get_train_ops(self, features, targets):
+    """See base class."""
+    if self._linear_feature_columns is None:
+      self._linear_feature_columns = layers.infer_real_valued_columns(features)
+    return super(LinearRegressor, self)._get_train_ops(features, targets)
+
+
+# TODO(ipolosukhin): Deprecate this class in favor of LinearClassifier.
 class TensorFlowLinearRegressor(TensorFlowEstimator, _sklearn.RegressorMixin):
   """TensorFlow Linear Regression model."""
 
@@ -50,12 +190,12 @@ class TensorFlowLinearRegressor(TensorFlowEstimator, _sklearn.RegressorMixin):
   @property
   def weights_(self):
     """Returns weights of the linear regression."""
-    return self.get_tensor_value('linear_regression/weights:0')
+    return self.get_tensor_value('linear_regression/weights')
 
   @property
   def bias_(self):
     """Returns bias of the linear regression."""
-    return self.get_tensor_value('linear_regression/bias:0')
+    return self.get_tensor_value('linear_regression/bias')
 
 
 class TensorFlowLinearClassifier(TensorFlowEstimator, _sklearn.ClassifierMixin):
@@ -89,12 +229,12 @@ class TensorFlowLinearClassifier(TensorFlowEstimator, _sklearn.ClassifierMixin):
   @property
   def weights_(self):
     """Returns weights of the linear classifier."""
-    return self.get_tensor_value('logistic_regression/weights:0')
+    return self.get_tensor_value('logistic_regression/weights')
 
   @property
   def bias_(self):
     """Returns weights of the linear classifier."""
-    return self.get_tensor_value('logistic_regression/bias:0')
+    return self.get_tensor_value('logistic_regression/bias')
 
 
 TensorFlowRegressor = TensorFlowLinearRegressor

tensorflow/contrib/learn/python/learn/estimators/linear_test.py

@@ -0,0 +1,155 @@
+#  Copyright 2015 Google Inc. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+"""Tests for estimators.linear."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow as tf
+
+
+class LinearClassifierTest(tf.test.TestCase):
+
+  def testTrain(self):
+    """Tests that loss goes down with training."""
+
+    def input_fn():
+      return {
+          'age': tf.constant([1]),
+          'language': tf.SparseTensor(values=['english'],
+                                      indices=[[0, 0]],
+                                      shape=[1, 1])
+      }, tf.constant([[1.]])
+
+    language = tf.contrib.layers.sparse_column_with_hash_bucket('language', 100)
+    age = tf.contrib.layers.real_valued_column('age')
+
+    classifier = tf.contrib.learn.LinearClassifier(
+        feature_columns=[age, language])
+    loss1 = classifier.train(input_fn, steps=100)
+    loss2 = classifier.train(input_fn, steps=200)
+
+    self.assertLess(loss2, loss1)
+    self.assertLess(loss2, 0.01)
+
+  def testTrainOptimizerWithL1Reg(self):
+    """Tests l1 regularized model has higher loss."""
+
+    def input_fn():
+      return {
+          'language': tf.SparseTensor(values=['hindi'],
+                                      indices=[[0, 0]],
+                                      shape=[1, 1])
+      }, tf.constant([[1.]])
+
+    language = tf.contrib.layers.sparse_column_with_hash_bucket('language', 100)
+    classifier_no_reg = tf.contrib.learn.LinearClassifier(
+        feature_columns=[language])
+    classifier_with_reg = tf.contrib.learn.LinearClassifier(
+        feature_columns=[language],
+        optimizer=tf.train.FtrlOptimizer(learning_rate=1.0,
+                                         l1_regularization_strength=100.))
+    loss_no_reg = classifier_no_reg.train(input_fn, steps=100)
+    loss_with_reg = classifier_with_reg.train(input_fn, steps=100)
+    self.assertLess(loss_no_reg, loss_with_reg)
+
+  def testTrainWithMissingFeature(self):
+    """Tests that training works with missing features."""
+
+    def input_fn():
+      return {
+          'language': tf.SparseTensor(values=['Swahili', 'turkish'],
+                                      indices=[[0, 0], [2, 0]],
+                                      shape=[3, 1])
+      }, tf.constant([[1.], [1.], [1.]])
+
+    language = tf.contrib.layers.sparse_column_with_hash_bucket('language', 100)
+    classifier = tf.contrib.learn.LinearClassifier(feature_columns=[language])
+    loss = classifier.train(input_fn, steps=100)
+    self.assertLess(loss, 0.01)
+
+  def testEval(self):
+    """Tests that eval produces correct metrics.
+    """
+
+    def input_fn():
+      return {
+          'age': tf.constant([[1], [2]]),
+          'language': tf.SparseTensor(values=['greek', 'chinise'],
+                                      indices=[[0, 0], [1, 0]],
+                                      shape=[2, 1]),
+      }, tf.constant([[1.], [0.]])
+
+    language = tf.contrib.layers.sparse_column_with_hash_bucket('language', 100)
+    age = tf.contrib.layers.real_valued_column('age')
+    classifier = tf.contrib.learn.LinearClassifier(
+        feature_columns=[age, language])
+
+    # Evaluate on untrained model
+    classifier.evaluate(input_fn=input_fn, steps=2)
+    # TODO(ispir): Enable accuracy check after resolving the randomness issue.
+    # self.assertAlmostEqual(.5, evaluated_values['accuracy/mean'])
+
+    # Evaluate on trained mdoel
+    classifier.train(input_fn, steps=100)
+    classifier.evaluate(input_fn=input_fn, steps=2)
+
+    # TODO(ispir): Enable accuracy check after resolving the randomness issue.
+    # self.assertLess(evaluated_values['loss/mean'], 0.3)
+    # self.assertGreater(evaluated_values['accuracy/mean'], .95)
+
+
+class LinearRegressorTest(tf.test.TestCase):
+
+  def testRegression(self):
+    """Tests that loss goes down with training."""
+
+    def input_fn():
+      return {
+          'age': tf.constant([1]),
+          'language': tf.SparseTensor(values=['english'],
+                                      indices=[[0, 0]],
+                                      shape=[1, 1])
+      }, tf.constant([[10.]])
+
+    language = tf.contrib.layers.sparse_column_with_hash_bucket('language', 100)
+    age = tf.contrib.layers.real_valued_column('age')
+
+    classifier = tf.contrib.learn.LinearRegressor(
+        feature_columns=[age, language])
+    loss1 = classifier.train(input_fn, steps=100)
+    loss2 = classifier.train(input_fn, steps=200)
+
+    self.assertLess(loss2, loss1)
+    self.assertLess(loss2, 0.01)
+
+
+def boston_input_fn():
+  boston = tf.contrib.learn.datasets.load_boston()
+  features = tf.cast(tf.reshape(tf.constant(boston.data), [-1, 13]), tf.float32)
+  target = tf.cast(tf.reshape(tf.constant(boston.target), [-1, 1]), tf.float32)
+  return features, target
+
+
+class InferedColumnTest(tf.test.TestCase):
+
+  def testTrain(self):
+    est = tf.contrib.learn.LinearRegressor()
+    est.train(input_fn=boston_input_fn, steps=1)
+    _ = est.evaluate(input_fn=boston_input_fn, steps=1)
+
+
+if __name__ == '__main__':
+  tf.test.main()

tensorflow/contrib/learn/python/learn/estimators/rnn.py

@@ -111,12 +111,12 @@ class TensorFlowRNNClassifier(TensorFlowEstimator, _sklearn.ClassifierMixin):
   @property
   def bias_(self):
     """Returns bias of the rnn layer."""
-    return self.get_tensor_value('logistic_regression/bias:0')
+    return self.get_tensor_value('logistic_regression/bias')
 
   @property
   def weights_(self):
     """Returns weights of the rnn layer."""
-    return self.get_tensor_value('logistic_regression/weights:0')
+    return self.get_tensor_value('logistic_regression/weights')
 
 
 class TensorFlowRNNRegressor(TensorFlowEstimator, _sklearn.RegressorMixin):
@@ -201,9 +201,9 @@ class TensorFlowRNNRegressor(TensorFlowEstimator, _sklearn.RegressorMixin):
   @property
   def bias_(self):
     """Returns bias of the rnn layer."""
-    return self.get_tensor_value('linear_regression/bias:0')
+    return self.get_tensor_value('linear_regression/bias')
 
   @property
   def weights_(self):
     """Returns weights of the rnn layer."""
-    return self.get_tensor_value('linear_regression/weights:0')
+    return self.get_tensor_value('linear_regression/weights')

tensorflow/contrib/learn/python/learn/graph_actions.py

@@ -28,12 +28,14 @@ from six import reraise
 from tensorflow.contrib.framework.python.ops import ops as contrib_ops
 from tensorflow.contrib.framework.python.ops import variables as contrib_variables
 from tensorflow.contrib.layers.python.layers import summaries
+from tensorflow.contrib.learn.python.learn import monitors as monitors_lib
 from tensorflow.core.util.event_pb2 import SessionLog
 from tensorflow.python.client import session as tf_session
 from tensorflow.python.framework import errors
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_util
 from tensorflow.python.ops import data_flow_ops
+from tensorflow.python.ops import logging_ops
 from tensorflow.python.ops import variables
 from tensorflow.python.platform import gfile
 from tensorflow.python.platform import tf_logging as logging
@@ -98,23 +100,24 @@ def _prepare_session(graph,
                      start_services,
                      global_step_tensor,
                      init_op=None,
+                     init_feed_dict=None,
                      init_fn=None,
                      supervisor_is_chief=True,
                      supervisor_master='',
-                     supervisor_save_model_secs=600,
-                     supervisor_save_summaries_secs=10):
+                     supervisor_save_model_secs=600):
   """Starts a session using the supervisor."""
   if global_step_tensor is None:
     global_step_tensor = Supervisor.USE_DEFAULT
   supervisor = Supervisor(
       graph,
       init_op=init_op or Supervisor.USE_DEFAULT,
+      init_feed_dict=init_feed_dict,
       is_chief=supervisor_is_chief,
       logdir=output_dir,
       saver=_make_saver(graph),
       global_step=global_step_tensor,
+      summary_op=None,
       save_model_secs=supervisor_save_model_secs,
-      save_summaries_secs=supervisor_save_summaries_secs,
       init_fn=init_fn)
   session = supervisor.PrepareSession(master=supervisor_master,
                                       start_standard_services=start_services)
@@ -122,6 +125,23 @@ def _prepare_session(graph,
   return supervisor, session
 
 
+def _run_with_monitors(session, step, tensors, feed_dict, monitors):
+  """Runs session for given tensors with monitor callbacks."""
+  for monitor in monitors:
+    tensors = monitor.step_begin(step, tensors)
+
+  outputs = session.run(tensors, feed_dict=feed_dict)
+  outputs = dict(zip(
+      [t.name if isinstance(t, ops.Tensor) else t for t in tensors],
+      outputs))
+
+  should_stop = False
+  for monitor in monitors:
+    induce_stop = monitor.step_end(step, outputs)
+    should_stop = should_stop or induce_stop
+  return outputs, should_stop
+
+
 # TODO(ptucker): Add unit test.
 # TODO(wicke): switch to forced named kwargs
 def train(graph,
@@ -130,15 +150,17 @@ def train(graph,
           loss_op,
           global_step_tensor=None,
           init_op=None,
+          init_feed_dict=None,
           init_fn=None,
           log_every_steps=10,
           supervisor_is_chief=True,
           supervisor_master='',
           supervisor_save_model_secs=600,
-          supervisor_save_summaries_secs=10,
+          supervisor_save_summaries_steps=100,
           feed_fn=None,
           max_steps=None,
-          fail_on_nan_loss=True):
+          fail_on_nan_loss=True,
+          monitors=None):
   """Train a model.
 
   Given `graph`, a directory to write outputs to (`output_dir`), and some ops,
@@ -164,6 +186,8 @@ def train(graph,
       one is extracted from the graph using the same logic as in `Supervisor`.
     init_op: An op that initializes the graph. If `None`, use `Supervisor`'s
       default.
+    init_feed_dict: A dictionary that maps `Tensor` objects to feed values.
+      This feed dictionary will be used when `init_op` is evaluated.
     init_fn: Optional callable passed to Supervisor to initialize the model.
     log_every_steps: Output logs regularly. The logs contain timing data and the
       current loss.
@@ -172,13 +196,15 @@ def train(graph,
     supervisor_master: The master string to use when preparing the session.
     supervisor_save_model_secs: Save a checkpoint every
       `supervisor_save_model_secs` seconds when training.
-    supervisor_save_summaries_secs: Save summaries every
-      `supervisor_save_summaries_secs` seconds when training.
+    supervisor_save_summaries_steps: Save summaries every
+      `supervisor_save_summaries_steps` seconds when training.
     feed_fn: A function that is called every iteration to produce a `feed_dict`
       passed to `session.run` calls. Optional.
     max_steps: Train until `global_step_tensor` evaluates to this value.
     fail_on_nan_loss: If true, raise `NanLossDuringTrainingError` if `loss_op`
       evaluates to `NaN`. If false, continue training as if nothing happened.
+    monitors: List of `BaseMonitor` subclass instances. Used for callbacks
+      inside the training loop.
 
   Returns:
     The final loss value.
@@ -194,17 +220,30 @@ def train(graph,
       graph, global_step_tensor)
   if global_step_tensor is None:
     raise ValueError('No "global_step" was provided or found in the graph.')
+
+  # TODO(ipolosukhin): Replace all functionality of Supervisor with Monitors.
+  if not monitors:
+    monitors = monitors_lib.get_default_monitors(
+        loss_op=loss_op,
+        summary_op=logging_ops.get_summary_op(),
+        save_summary_steps=supervisor_save_summaries_steps,
+        output_dir=output_dir)
+
+  # Start monitors, can create graph parts.
+  for monitor in monitors:
+    monitor.begin(max_steps=max_steps)
+
   supervisor, session = _prepare_session(
       graph=graph,
       output_dir=output_dir,
       start_services=True,
       global_step_tensor=global_step_tensor,
       init_op=init_op,
+      init_feed_dict=init_feed_dict,
       init_fn=init_fn,
       supervisor_is_chief=supervisor_is_chief,
       supervisor_master=supervisor_master,
-      supervisor_save_model_secs=supervisor_save_model_secs,
-      supervisor_save_summaries_secs=supervisor_save_summaries_secs)
+      supervisor_save_model_secs=supervisor_save_model_secs)
 
   with session:
     get_current_step = lambda: session.run(global_step_tensor)
@@ -215,13 +254,18 @@ def train(graph,
     loss_value = None
     logging.info('Training steps [%d,%s)', last_step, 'inf'
                  if max_steps is None else str(max_steps))
+
     excinfo = None
     try:
       while not supervisor.ShouldStop() and (
           (max_steps is None) or (last_step < max_steps)):
         start_time = time.time()
         feed_dict = feed_fn() if feed_fn is not None else None
-        _, loss_value = session.run([train_op, loss_op], feed_dict=feed_dict)
+
+        outputs, should_stop = _run_with_monitors(
+            session, last_step + 1, [train_op, loss_op], feed_dict, monitors)
+
+        loss_value = outputs[loss_op.name]
         if np.isnan(loss_value):
           failure_message = 'Model diverged with loss = NaN.'
           if fail_on_nan_loss:
@@ -230,6 +274,9 @@ def train(graph,
           else:
             logging.warning(failure_message)
 
+        if should_stop:
+          break
+
         this_step = get_current_step()
 
         if this_step <= last_step:
@@ -268,12 +315,11 @@ def train(graph,
           logging.info('Saving checkpoint for step %d to checkpoint: %s.' % (
               last_step, ckpt_path))
           supervisor.saver.save(session, ckpt_path, global_step=last_step)
-          if supervisor.summary_op is not None:
-            summary_strs = session.run(supervisor.summary_op)
-            supervisor.summary_writer.add_summary(summary_strs, last_step)
-            supervisor.summary_writer.add_session_log(
-                SessionLog(status=SessionLog.STOP), last_step)
-            supervisor.summary_writer.close()
+
+          # Finish monitors.
+          for monitor in monitors:
+            monitor.end()
+
       # catch OutOfRangeError which is thrown when queue is out of data (and for
       # other reasons as well).
       except errors.OutOfRangeError as e:
@@ -354,6 +400,9 @@ def evaluate(graph,
     if isinstance(value, ops.Tensor):
       summaries.summarize_tensor(value, tag=key)
 
+  # Create or get summary op.
+  summary_op = logging_ops.get_summary_op()
+
   # TODO(wicke): Don't use supervisor here, or switch to output_dir=eval_dir.
   supervisor, session = _prepare_session(
       graph=graph,
@@ -363,8 +412,7 @@ def evaluate(graph,
       init_op=init_op,
       supervisor_is_chief=True,
       supervisor_master=supervisor_master,
-      supervisor_save_model_secs=None,
-      supervisor_save_summaries_secs=None)
+      supervisor_save_model_secs=None)
   global_step_tensor = supervisor.global_step
 
   with session:
@@ -394,14 +442,17 @@ def evaluate(graph,
                          ', '.join('%s = %s' % (k, v)
                                    for k, v in eval_results.items()))
       finally:
-        # Make our own summary writer and write a summary to the eval dir
-        if supervisor.summary_op is not None:
+        # Make our own summary writer and write a summary to the eval dir.
+        # Only is feed_fn is not provided.
+        # TODO(ipolosukhin): Convert evaluation to use streaming_metrics,
+        # then we can save for non feed_fn as well.
+        if summary_op is not None and feed_fn is None:
           summary_writer = None
           try:
             summary_writer = SummaryWriter(output_dir,
                                            graph_def=session.graph_def)
 
-            summary_str = session.run(supervisor.summary_op)
+            summary_str = session.run(summary_op)
             if summary_str:
               summary_writer.add_summary(summary_str, current_global_step)
           finally:
@@ -414,7 +465,10 @@ def evaluate(graph,
     # catch OutOfRangeError which is thrown when queue is out of data (and for
     # other reasons as well).
     except errors.OutOfRangeError as e:
-      logging.warn('Input queue exhausted: %s.', e)
+      logging.warn('Input queue is exhausted: %s.', e)
+    # catch StopIteration which is thrown is DataReader is out of data.
+    except StopIteration as e:
+      logging.info('Input iterator is exhausted: %s.', e)
 
   return eval_results, current_global_step
 

tensorflow/contrib/learn/python/learn/io/data_feeder.py

@@ -13,6 +13,9 @@
 #  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 #  See the License for the specific language governing permissions and
 #  limitations under the License.
+
+# TODO(ipolosukhin): Replace this module with feed-dict queue runners & queues.
+
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
@@ -33,6 +36,8 @@ from .dask_io import HAS_DASK, extract_dask_data, extract_dask_labels
 
 def _get_in_out_shape(x_shape, y_shape, n_classes, batch_size):
   """Returns shape for input and output of the data feeder."""
+  if batch_size < 0:
+    batch_size = x_shape[0]
   x_shape = list(x_shape[1:]) if len(x_shape) > 1 else [1]
   input_shape = [batch_size] + x_shape
   if y_shape is None:
@@ -65,7 +70,8 @@ def _is_iterable(X):
   return hasattr(X, 'next') or hasattr(X, '__next__')
 
 
-def setup_train_data_feeder(X, y, n_classes, batch_size):
+def setup_train_data_feeder(
+    X, y, n_classes, batch_size, shuffle=True, epochs=None):
   """Create data feeder, to sample inputs from dataset.
 
   If X and y are iterators, use StreamingDataFeeder.
@@ -95,7 +101,8 @@ def setup_train_data_feeder(X, y, n_classes, batch_size):
       raise ValueError('Both X and y should be iterators for '
                        'streaming learning to work.')
     data_feeder_cls = StreamingDataFeeder
-  return data_feeder_cls(X, y, n_classes, batch_size)
+  return data_feeder_cls(
+      X, y, n_classes, batch_size, shuffle=shuffle, epochs=epochs)
 
 
 def _batch_data(X, batch_size):
@@ -189,7 +196,7 @@ class DataFeeder(object):
     output_dtype: dtype of output.
   """
 
-  def __init__(self, X, y, n_classes, batch_size, random_state=None):
+  def __init__(self, X, y, n_classes, batch_size, shuffle=True, random_state=None, epochs=None):
     x_dtype = np.int64 if X.dtype == np.int64 else np.float32
     y_dtype = (
         np.int64 if n_classes is not None and n_classes > 1 else np.float32)
@@ -201,6 +208,7 @@ class DataFeeder(object):
       self.y = y
     self.n_classes = n_classes
     self.batch_size = batch_size
+    self.max_epochs = epochs
     self.input_shape, self.output_shape = _get_in_out_shape(self.X.shape, None
                                                             if self.y is None
                                                             else self.y.shape,
@@ -213,9 +221,13 @@ class DataFeeder(object):
       self.output_dtype = np.float32
     else:
       self.output_dtype = y.dtype
+    self.shuffle = shuffle
     self.random_state = np.random.RandomState(
         42) if random_state is None else random_state
-    self.indices = self.random_state.permutation(self.X.shape[0])
+    if self.shuffle:
+      self.indices = self.random_state.permutation(self.X.shape[0])
+    else:
+      self.indices = np.array(range(self.X.shape[0]))
     self.offset = 0
     self.epoch = 0
     self._epoch_placeholder = None
@@ -283,16 +295,19 @@ class DataFeeder(object):
       from X and y.
     """
     def _feed_dict_fn():
+      if self.max_epochs is not None and self.epoch + 1 > self.max_epochs:
+        raise StopIteration
       assert self._input_placeholder != None
       feed_dict = {}
       if self._epoch_placeholder is not None:
         feed_dict[self._epoch_placeholder.name] = [self.epoch]
 
-      # take random indices
+      # take next batch of indices
       if self.batch_size < 0:
         batch_indices = self.indices
       else:
-        batch_indices = self.indices[self.offset:self.offset + self.batch_size]
+        end = min(self.X.shape[0], self.offset + self.batch_size)
+        batch_indices = self.indices[self.offset:end]
 
       # assign input features from random indices
       inp = np.array(self.X[batch_indices]).reshape((batch_indices.shape[0], 1)) \
@@ -300,10 +315,13 @@ class DataFeeder(object):
       feed_dict[self._input_placeholder.name] = inp
 
       # move offset and reset it if necessary
-      self.offset += self.batch_size
-      if self.offset >= self.X.shape[0]:
-        self.indices = self.random_state.permutation(self.X.shape[0])
-        self.offset = 0
+      if self.batch_size > 0:
+        self.offset += self.batch_size
+        if self.offset >= self.X.shape[0]:
+          self.indices = self.random_state.permutation(self.X.shape[0])
+          self.offset = 0
+          self.epoch += 1
+      else:
         self.epoch += 1
 
       # return early if there are no labels
@@ -359,15 +377,21 @@ class StreamingDataFeeder(DataFeeder):
     output_dtype: dtype of output.
   """
 
-  def __init__(self, X, y, n_classes, batch_size):
+  def __init__(self, X, y, n_classes, batch_size, shuffle=False, epochs=None):
     X_first_el = six.next(X)
-    y_first_el = six.next(y)
     self.X = itertools.chain([X_first_el], X)
-    self.y = itertools.chain([y_first_el], y)
+    if y is not None:
+      y_first_el = six.next(y)
+      self.y = itertools.chain([y_first_el], y)
+    else:
+      y_first_el = None
+      self.y = None
     self.n_classes = n_classes
     self.batch_size = batch_size
     self.input_shape, self.output_shape = _get_in_out_shape(
-        [1] + list(X_first_el.shape), [1] + list(y_first_el.shape), n_classes,
+        [1] + list(X_first_el.shape),
+        [1] + list(y_first_el.shape) if y is not None else None,
+        n_classes,
         batch_size)
     self.input_dtype = X_first_el.dtype
     # Convert float64 to float32, as all the parameters in the model are
@@ -396,21 +420,37 @@ class StreamingDataFeeder(DataFeeder):
       A function that when called samples a random subset of batch size
       from X and y.
     """
+    self.stopped = False
 
     def _feed_dict_fn():
+      if self.stopped:
+        raise StopIteration
       inp = np.zeros(self.input_shape, dtype=self.input_dtype)
-      out = np.zeros(self.output_shape, dtype=self.output_dtype)
+      if self.y is not None:
+        out = np.zeros(self.output_shape, dtype=self.output_dtype)
       for i in xrange(self.batch_size):
-        inp[i, :] = six.next(self.X)
-        y = six.next(self.y)
-        if self.n_classes > 1:
-          if len(self.output_shape) == 2:
-            out.itemset((i, y), 1.0)
+        # Add handling when queue ends.
+        try:
+          inp[i, :] = six.next(self.X)
+        except StopIteration:
+          self.stopped = True
+          inp = inp[:i, :]
+          if self.y is not None:
+            out = out[:i]
+          break
+
+        if self.y is not None:
+          y = six.next(self.y)
+          if self.n_classes > 1:
+            if len(self.output_shape) == 2:
+              out.itemset((i, y), 1.0)
+            else:
+              for idx, value in enumerate(y):
+                out.itemset(tuple([i, idx, value]), 1.0)
           else:
-            for idx, value in enumerate(y):
-              out.itemset(tuple([i, idx, value]), 1.0)
-        else:
-          out[i] = y
+            out[i] = y
+      if self.y is None:
+        return {self._input_placeholder.name: inp}
       return {self._input_placeholder.name: inp,
               self._output_placeholder.name: out}
 
@@ -443,7 +483,7 @@ class DaskDataFeeder(object):
     input_dtype: dtype of input.
     output_dtype: dtype of output.
   """
-  def __init__(self, X, y, n_classes, batch_size, random_state=None):
+  def __init__(self, X, y, n_classes, batch_size, shuffle=True, random_state=None):
     import dask.dataframe as dd
     # TODO(terrytangyuan): check X and y dtypes in dask_io like pandas
     self.X = X

tensorflow/contrib/learn/python/learn/monitors.py

@@ -17,221 +17,199 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import os
-import sys
-import numpy as np
-
-from tensorflow.core.framework import summary_pb2
-from tensorflow.python.training import training as train
-from tensorflow.contrib.learn.python.learn.io.data_feeder import setup_train_data_feeder
-
-# pylint: disable=too-many-instance-attributes
-# pylint: disable=too-few-public-methods
-# pylint: disable=too-many-arguments
-# pylint: disable=attribute-defined-outside-init
-
-
-def default_monitor(verbose=1):
-  """Returns very simple monitor object to summarize training progress.
-
-  Args:
-    verbose: Level of verbosity of output.
-
-  Returns:
-    Default monitor object.
-  """
-  return BaseMonitor(verbose=verbose)
+from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.training import summary_io
 
 
 class BaseMonitor(object):
-  """Base class for all learning monitors.
-
-  Stores and reports training loss throughout learning
-
-  Parameters:
-    print_steps: Number of steps in between printing cost.
-    early_stopping_rounds:  Activates early stopping if this is not None.
-                            Loss needs to decrease at least every
-                            <early_stopping_rounds>
-                            round(s) to continue training. (default: None)
-    verbose: Level of verbosity of output.
+  """Base class for Monitors.
 
+  Defines basic interfaces of Monitors.
   """
 
-  def __init__(self, print_steps=100, early_stopping_rounds=None, verbose=1):
-    self.print_steps = print_steps
-    self.early_stopping_rounds = early_stopping_rounds
-
-    self.converged = False
-    self.min_loss = np.inf
-    self.min_loss_i = 0
-    self.last_loss_seen = np.inf
-    self.steps = 0
-    self.print_train_loss_buffer = []
-    self.all_train_loss_buffer = []
-    self.verbose = verbose
-    self.epoch = None
-
-  def update(self, global_step, step_number, training_loss, sess,
-             feed_params_fn, loss_expression_tensor):
-    """Adds training_loss to monitor.
-
-    Triggers printed output if appropriate
-
-    Args:
-      global_step: global step number
-      step_number: current step in training
-      training_loss: float value of training loss
-      sess: session for computation (used to calculate validation loss)
-      feed_params_fn: function generating dict with information like
-        epoch. Sometimes None.
-      loss_expression_tensor: Tensor applied to validation data to
-        calculate val loss
-
-    """
-    self.steps = step_number
-    self.global_step = global_step
-    self.print_train_loss_buffer.append(training_loss)
-    self.all_train_loss_buffer.append(training_loss)
-    self.sess = sess
-    self.loss_expression_tensor = loss_expression_tensor
-    self._set_last_loss_seen()
-    if self.last_loss_seen < self.min_loss:
-      self.min_loss = self.last_loss_seen
-      self.min_loss_i = self.steps
-    self._set_epoch(feed_params_fn)
-    self.report()
-
-  def _set_last_loss_seen(self):
-    """Sets last_loss_seen attribute to most recent training error."""
-    self.last_loss_seen = self.all_train_loss_buffer[-1]
-    self._estimator = None
-
-  def report(self):
-    """Checks whether to report, and prints loss information if appropriate."""
-    if self.verbose and (self.steps % self.print_steps == 0):
-      self._set_training_summary()
-      print(self._summary_str)
-
   def set_estimator(self, estimator):
     self._estimator = estimator
 
-  def monitor_inducing_stop(self):
-    """Returns True if the monitor requests the model stop.
-
-    Returns:
-      True if the monitor requests the model stop (e.g. for early stopping).
-    """
-    if self.early_stopping_rounds is None:
-      return False
-    stop_now = (self.steps - self.min_loss_i >= self.early_stopping_rounds)
-    if stop_now:
-      sys.stderr.write("Stopping. Best step:\n step {} with loss {}\n"
-                       .format(self.min_loss_i, self.min_loss))
-    return stop_now
-
-  def create_val_feed_dict(self, inp, out):
-    """Validation requires access to TensorFlow placeholders.
-
-    Not used in this Monitor
+  def begin(self, max_steps=None):
+    """Callback at the beginning of training/evaluation.
 
     Args:
-      inp: not used
-      out: not used
+      max_steps: Maximum steps this training will run until.
     """
     pass
 
-  def _set_epoch(self, feed_params_fn):
-    """Sets self.epoch from a function providing this info in a dict."""
-    if feed_params_fn:
-      feed_params = feed_params_fn()
-      self.epoch = feed_params["epoch"] if "epoch" in feed_params else None
-
-  def _set_training_summary(self):
-    """Returns the string to be written describing training progress"""
-    avg_train_loss = np.mean(self.print_train_loss_buffer)
-    self.print_train_loss_buffer = []
-    if self.epoch:
-      self._summary_str = (
-          "Step #{step}, epoch #{epoch}, avg. train loss: {loss:.5f}"
-          .format(step=self.steps,
-                  loss=avg_train_loss,
-                  epoch=self.epoch))
-    else:
-      self._summary_str = ("Step #{step}, avg. train loss: {loss:.5f}"
-                           .format(step=self.global_step,
-                                   loss=avg_train_loss))
-    self._modify_summary_string()
-
-  def _modify_summary_string(self):
-    """Makes monitor specific changes to printed summary.
-
-    Nothing interesting in BaseMonitor
-    """
+  def end(self):
+    """Callback at the end of training/evaluation."""
     pass
 
+  def epoch_begin(self, epoch):
+    pass
 
-class ValidationMonitor(BaseMonitor):
-  """Monitor that reports validation score and uses it for early stopping.
+  def epoch_end(self, epoch):
+    pass
+
+  def step_begin(self, step, tensors):  # pylint: disable=unused-argument
+    """Callback before training step begins.
+
+    Use this callback to:
+     - override which tensors to run.
+
+    Args:
+      step: int, global step of the model.
+      tensors: list of `Tensors` that going to be passed to session.run.
+
+    Returns:
+      Dict of `Tensors` that going to be ran.
+    """
+    return tensors
+
+  def step_end(self, step, output):  # pylint: disable=unused-argument
+    """Callback after training step finished.
+
+    Use this callback to:
+     - log results.
+     - save checkpoints.
+     - compute validation score.
+     - perform early stopping.
+
+    Args:
+      step: `int`, global step of the model.
+      output: `dict` of `np.array` results executed.
+
+    Returns:
+      `bool`, `True` if model should stop, `False` or `None` if continue.
+    """
+    return False
+
+
+class EveryN(BaseMonitor):
+  """Base class for monitors that execute callbacks every n steps / seconds.
 
   Parameters:
-    val_X: Validation features
-    val_y: Validation labels
-    n_classes: Number of labels in output. 0 for regression
-    print_steps: Number of steps in between printing cost.
-    early_stopping_rounds:  Activates early stopping if this is not None.
-                            Loss needs to decrease at least every
-                            <early_stopping_rounds>
-                            round(s) to continue training. (default: None)
+    every_n_steps: int, calls `every_n_step_{begin,end}` every this many steps.
+    first_n_steps: int, calls `every_n_step_{begin,end}` for first n steps.
+
+  TODO(ipolosukhin): Add also every n seconds.
   """
 
-  def __init__(self,
-               val_X,
-               val_y,
-               n_classes=0,
-               print_steps=100,
-               early_stopping_rounds=None):
-    super(ValidationMonitor, self).__init__(
-        print_steps=print_steps,
-        early_stopping_rounds=early_stopping_rounds)
-    self.val_feeder = setup_train_data_feeder(val_X, val_y, n_classes, -1)
-    self.print_val_loss_buffer = []
-    self.all_val_loss_buffer = []
-    self._summary_writer = None
+  def __init__(
+      self, every_n_steps=100, first_n_steps=1):
+    self._every_n_steps = every_n_steps
+    self._first_n_steps = first_n_steps
+    self._max_steps = None
 
-  def create_val_feed_dict(self, inp, out):
-    """Set tensorflow placeholders and create validation data feed."""
-    self.val_feeder.set_placeholders(inp, out)
-    self.val_dict = self.val_feeder.get_feed_dict_fn()()
+  def begin(self, max_steps=None):
+    self._max_steps = max_steps
+
+  def every_n_step_begin(self, step, tensors):  # pylint: disable=unused-argument
+    return tensors
+
+  def every_n_step_end(self, step, outputs):  # pylint: disable=unused-argument
+    return False
+
+  def step_begin(self, step, tensors):
+    if (step <= self._first_n_steps or step % self._every_n_steps == 0 or
+        step == self._max_steps):
+      tensors = self.every_n_step_begin(step, tensors)
+    return tensors
+
+  def step_end(self, step, output):
+    to_stop = False
+    if (step <= self._first_n_steps or step % self._every_n_steps == 0 or
+        step == self._max_steps):
+      to_stop = self.every_n_step_end(step, output)
+    return to_stop
+
+
+class PrintTensor(EveryN):
+
+  def __init__(self, tensor_names, every_n=100, first_n=1):
+    super(PrintTensor, self).__init__(every_n, first_n)
+    self._tensor_names = tensor_names
+
+  def every_n_step_begin(self, unused_step, tensors):
+    return tensors + self._tensor_names
+
+  def every_n_step_end(self, step, outputs):
+    stats = []
+    for name in self._tensor_names:
+      if name in outputs:
+        stats.append("%s = %s" % (name, str(outputs[name])))
+    logging.info("Step %d: %s" % (step, ", ".join(stats)))
+
+
+class SummarySaver(EveryN):
+  """Saves summary every N seconds."""
+
+  def __init__(self, summary_op, save_steps=100, output_dir=None):
+    # TODO(ipolosukhin): Implement every N seconds.
+    super(SummarySaver, self).__init__(every_n_steps=save_steps)
+    self._summary_op = summary_op
+    self._summary_writer = None
+    if output_dir:
+      self._summary_writer = summary_io.SummaryWriter(output_dir)
 
   def set_estimator(self, estimator):
-    super(ValidationMonitor, self).set_estimator(estimator)
-    if estimator._output_dir is None:
-      return
-    self._summary_writer = train.SummaryWriter(os.path.join(estimator._output_dir, 'eval'))
+    super(SummarySaver, self).set_estimator(estimator)
+    self._summary_writer = summary_io.SummaryWriter(self._estimator.model_dir)
 
-  def _set_last_loss_seen(self):
-    """Sets self.last_loss_seen to most recent validation loss.
+  def every_n_step_begin(self, unused_step, tensors):
+    return tensors + [self._summary_op]
 
-    Also stores this value to appropriate buffers
-    """
-    [val_loss] = self.sess.run(
-        [self.loss_expression_tensor],
-        feed_dict=self.val_dict)
-    self.last_loss_seen = val_loss
-    self.all_val_loss_buffer.append(val_loss)
-    self.print_val_loss_buffer.append(val_loss)
-    if self._summary_writer is not None:
-      summary = summary_pb2.Summary()
-      value = summary.value.add()
-      value.tag = "loss"
-      value.simple_value = float(val_loss)
-      self._summary_writer.add_summary(summary, self.global_step)
+  def every_n_step_end(self, step, outputs):
+    summary_strs = outputs[self._summary_op.name]
+    if self._summary_writer:
+      self._summary_writer.add_summary(summary_strs, step)
+    return False
 
-  def _modify_summary_string(self):
-    """Flushes validation print buffer into summary string."""
-    avg_val_loss = np.mean(self.print_val_loss_buffer)
-    self.print_val_loss_buffer = []
-    val_loss_string = "avg. val loss: {val_loss:.5f}".format(
-        val_loss=avg_val_loss)
-    self._summary_str = (", ".join([self._summary_str, val_loss_string]))
+  def end(self):
+    self._summary_writer.flush()
+
+
+class ValidationMonitor(EveryN):
+  """Runs evaluation every n steps.
+
+  Can do early stopping on validation loss if `early_stopping_rounds` provided.
+
+  """
+
+  def __init__(self, x=None, y=None, input_fn=None,
+               every_n_steps=100, early_stopping_rounds=None):
+    super(ValidationMonitor, self).__init__(every_n_steps=every_n_steps)
+    if x is None and input_fn is None:
+      raise ValueError("Either x or input_fn should be provided.")
+    self.x = x
+    self.y = y
+    self.input_fn = input_fn
+    self.min_loss_step = 0
+    self.min_loss = None
+    self.early_stopping_rounds = early_stopping_rounds
+
+  def every_n_step_end(self, step, unused_outputs):
+    outputs = self._estimator.evaluate(
+        x=self.x, y=self.y, input_fn=self.input_fn)
+    stats = []
+    for name in outputs:
+      stats.append("%s = %s" % (name, str(outputs[name])))
+    logging.info("Validation (step %d): %s" % (step, ", ".join(stats)))
+    if self.early_stopping_rounds is not None:
+      if self.min_loss is None or outputs["loss"] < self.min_loss:
+        self.min_loss = outputs["loss"]
+        self.min_loss_step = step
+      stop_now = (step - self.min_loss_step >= self.early_stopping_rounds)
+      if stop_now:
+        logging.info("Stopping. Best step: {} with loss {}."
+                     .format(self.min_loss_step, self.min_loss))
+        return True
+    return False
+
+
+def get_default_monitors(loss_op=None, summary_op=None, save_summary_steps=100,
+                         output_dir=None):
+  monitors = []
+  if loss_op is not None:
+    monitors.append(PrintTensor([loss_op.name]))
+  if summary_op is not None:
+    monitors.append(SummarySaver(summary_op, save_steps=save_summary_steps,
+                                 output_dir=output_dir))
+  return monitors

tensorflow/contrib/learn/python/learn/ops/dnn_ops.py

@@ -17,9 +17,12 @@ from __future__ import division
 from __future__ import print_function
 
 from tensorflow.contrib import layers
+from tensorflow.contrib.learn.python.learn.ops import dropout_ops
+from tensorflow.python.framework import ops
+from tensorflow.python.ops import array_ops as array_ops_
+from tensorflow.python.ops import control_flow_ops
 from tensorflow.python.ops import nn
 from tensorflow.python.ops import variable_scope as vs
-from tensorflow.contrib.learn.python.learn.ops import dropout_ops
 
 
 def dnn(tensor_in, hidden_units, activation=nn.relu, dropout=None):
@@ -49,5 +52,9 @@ def dnn(tensor_in, hidden_units, activation=nn.relu, dropout=None):
         if activation is not None:
           tensor_in = activation(tensor_in)
         if dropout is not None:
-          tensor_in = dropout_ops.dropout(tensor_in, prob=(1.0 - dropout))
+          is_training = array_ops_.squeeze(ops.get_collection('IS_TRAINING'))
+          tensor_in = control_flow_ops.cond(
+              is_training,
+              lambda: dropout_ops.dropout(tensor_in, prob=(1.0 - dropout)),
+              lambda: tensor_in)
     return tensor_in

tensorflow/contrib/learn/python/learn/tests/dataframe/test_feeding_queue_runner.py

@@ -0,0 +1,120 @@
+# Copyright 2016 Google Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Tests `FeedingQueueRunner` using arrays and `DataFrames`."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+
+import numpy as np
+import tensorflow as tf
+import tensorflow.contrib.learn.python.learn.dataframe.queues.feeding_functions as ff
+
+# pylint: disable=g-import-not-at-top
+try:
+  import pandas as pd
+  HAS_PANDAS = True
+except ImportError:
+  HAS_PANDAS = False
+
+
+def get_rows(array, row_indices):
+  rows = [array[i] for i in row_indices]
+  return np.vstack(rows)
+
+
+class FeedingQueueRunnerTestCase(tf.test.TestCase):
+  """Tests for `FeedingQueueRunner`."""
+
+  def testArrayFeeding(self):
+    with tf.Graph().as_default():
+      array = np.arange(32).reshape([16, 2])
+      q = ff.enqueue_data(array, capacity=100)
+      batch_size = 3
+      dq_op = q.dequeue_many(batch_size)
+      with tf.Session() as sess:
+        coord = tf.train.Coordinator()
+        threads = tf.train.start_queue_runners(sess=sess, coord=coord)
+        for i in range(100):
+          indices = [j % array.shape[0]
+                     for j in range(batch_size * i, batch_size * (i + 1))]
+          expected_dq = get_rows(array, indices)
+          dq = sess.run(dq_op)
+          np.testing.assert_array_equal(indices, dq[0])
+          np.testing.assert_array_equal(expected_dq, dq[1])
+        coord.request_stop()
+        coord.join(threads)
+
+  def testPandasFeeding(self):
+    if not HAS_PANDAS:
+      return
+    with tf.Graph().as_default():
+      array1 = np.arange(32)
+      array2 = np.arange(32, 64)
+      df = pd.DataFrame({"a": array1, "b": array2}, index=np.arange(64, 96))
+      q = ff.enqueue_data(df, capacity=100)
+      batch_size = 5
+      dq_op = q.dequeue_many(5)
+      with tf.Session() as sess:
+        coord = tf.train.Coordinator()
+        threads = tf.train.start_queue_runners(sess=sess, coord=coord)
+        for i in range(100):
+          indices = [j % array1.shape[0]
+                     for j in range(batch_size * i, batch_size * (i + 1))]
+          expected_df_indices = df.index[indices]
+          expected_rows = df.iloc[indices]
+          dq = sess.run(dq_op)
+          np.testing.assert_array_equal(expected_df_indices, dq[0])
+          for col_num, col in enumerate(df.columns):
+            np.testing.assert_array_equal(expected_rows[col].values,
+                                          dq[col_num + 1])
+        coord.request_stop()
+        coord.join(threads)
+
+  def testShuffle(self):
+    array_size = 7
+    batch_size = 3
+    iterations = 1000
+    mean = batch_size * iterations * 1.0 / array_size
+    tolerance = 3
+    with tf.Graph().as_default():
+      array = np.arange(array_size)
+      q = ff.enqueue_data(array, capacity=100, shuffle=True, seed=1234)
+      dq_op = q.dequeue_many(batch_size)
+      with tf.Session() as sess:
+        coord = tf.train.Coordinator()
+        threads = tf.train.start_queue_runners(sess=sess, coord=coord)
+        counts = {x: 0 for x in array}
+        for _ in range(iterations):
+          dq = sess.run(dq_op)
+          for dqed in dq[1]:
+            self.assertIn(dqed, array)
+            counts[dqed] += 1
+        for k, v in counts.items():
+          self.assertGreater(
+              mean + tolerance, v,
+              "Value {} occurred {} times, expected {:.2f} +/- {}".format(
+                  k, v, mean, tolerance))
+          self.assertLess(
+              mean - tolerance, v,
+              "Value {} occurred {} times, expected {:.2f} +/- {}".format(
+                  k, v, mean, tolerance))
+        coord.request_stop()
+        coord.join(threads)
+
+if __name__ == "__main__":
+  tf.test.main()

tensorflow/contrib/learn/python/learn/tests/test_base.py

@@ -14,7 +14,6 @@
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
-from __future__ import unicode_literals
 
 import random
 
@@ -61,13 +60,14 @@ class BaseTest(tf.test.TestCase):
     classifier.fit(iris.data, [float(x) for x in iris.target])
     self.assertEqual(
         classifier.get_variable_names(),
-        ["global_step:0", "logistic_regression/weights:0",
-         "logistic_regression/bias:0",
+        ["OptimizeLoss/learning_rate",
+         "OptimizeLoss/logistic_regression/bias/Adagrad",
          "OptimizeLoss/logistic_regression/softmax_classifier/"
-         "softmax_cross_entropy_loss/value/avg:0",
-         "OptimizeLoss/learning_rate:0",
-         "OptimizeLoss/logistic_regression/weights/Adagrad:0",
-         "OptimizeLoss/logistic_regression/bias/Adagrad:0"])
+           "softmax_cross_entropy_loss/value/avg",
+         "OptimizeLoss/logistic_regression/weights/Adagrad",
+         "global_step",
+         "logistic_regression/bias",
+         "logistic_regression/weights"])
 
   def testIrisSummaries(self):
     iris = datasets.load_iris()
@@ -84,9 +84,11 @@ class BaseTest(tf.test.TestCase):
                                                   steps=250)
     classifier.fit(iris.data, iris.target)
     score1 = accuracy_score(iris.target, classifier.predict(iris.data))
-    classifier.fit(iris.data, iris.target)
+    classifier.fit(iris.data, iris.target, steps=500)
     score2 = accuracy_score(iris.target, classifier.predict(iris.data))
-    self.assertGreater(score2, score1, "Failed with score = {0}".format(score2))
+    self.assertGreater(
+        score2, score1,
+        "Failed with score2 {0} <= score1 {1}".format(score2, score1))
 
   def testIrisStreaming(self):
     iris = datasets.load_iris()

tensorflow/contrib/learn/python/learn/tests/test_early_stopping.py

@@ -38,7 +38,8 @@ class EarlyStoppingTest(tf.test.TestCase):
 
     X_train, X_val, y_train, y_val = train_test_split(
         X_train, y_train, test_size=0.2)
-    val_monitor = learn.monitors.ValidationMonitor(X_val, y_val, n_classes=3)
+    val_monitor = learn.monitors.ValidationMonitor(X_val, y_val,
+                                                   early_stopping_rounds=100)
 
     # classifier without early stopping - overfitting
     classifier1 = learn.TensorFlowDNNClassifier(hidden_units=[10, 20, 10],
@@ -52,7 +53,7 @@ class EarlyStoppingTest(tf.test.TestCase):
                                                 n_classes=3,
                                                 steps=1000)
 
-    classifier2.fit(X_train, y_train, val_monitor)
+    classifier2.fit(X_train, y_train, monitors=[val_monitor])
     score2 = accuracy_score(y_test, classifier2.predict(X_test))
 
     # self.assertGreater(score2, score1, "No improvement using early stopping.")

tensorflow/contrib/learn/python/learn/tests/test_nonlinear.py

@@ -17,6 +17,7 @@ from __future__ import print_function
 
 import random
 
+import numpy as np
 import tensorflow as tf
 from tensorflow.contrib.learn.python import learn
 from tensorflow.contrib.learn.python.learn import datasets
@@ -48,7 +49,7 @@ class NonLinearTest(tf.test.TestCase):
     regressor = learn.TensorFlowDNNRegressor(hidden_units=[10, 20, 10],
                                              n_classes=0,
                                              batch_size=boston.data.shape[0],
-                                             steps=200,
+                                             steps=300,
                                              learning_rate=0.001)
     regressor.fit(boston.data, boston.target)
     score = mean_squared_error(boston.target, regressor.predict(boston.data))
@@ -79,6 +80,7 @@ class NonLinearTest(tf.test.TestCase):
                                                dropout=0.1)
     classifier.fit(iris.data, iris.target)
     score = accuracy_score(iris.target, classifier.predict(iris.data))
+    # If the quality is lower - dropout is not working.
     self.assertGreater(score, 0.9, "Failed with score = {0}".format(score))
 
   def testDNNDropout0_9(self):
@@ -94,7 +96,6 @@ class NonLinearTest(tf.test.TestCase):
 
   def testRNN(self):
     random.seed(42)
-    import numpy as np
     data = np.array(
         list([[2, 1, 2, 2, 3], [2, 2, 3, 4, 5], [3, 3, 1, 2, 1], [2, 4, 5, 4, 1]
              ]),
@@ -103,7 +104,8 @@ class NonLinearTest(tf.test.TestCase):
     labels = np.array(list([1, 0, 1, 0]), dtype=np.float32)
     # targets for regression
     targets = np.array(list([10, 16, 10, 16]), dtype=np.float32)
-    test_data = np.array(list([[1, 3, 3, 2, 1], [2, 3, 4, 5, 6]]))
+    test_data = np.array(list([[1, 3, 3, 2, 1], [2, 3, 4, 5, 6]]),
+                         dtype=np.float32)
 
     def input_fn(X):
       return tf.split(1, 5, X)
@@ -144,7 +146,6 @@ class NonLinearTest(tf.test.TestCase):
 
   def testBidirectionalRNN(self):
     random.seed(42)
-    import numpy as np
     data = np.array(
         list([[2, 1, 2, 2, 3], [2, 2, 3, 4, 5], [3, 3, 1, 2, 1], [2, 4, 5, 4, 1]
              ]),
@@ -161,19 +162,20 @@ class NonLinearTest(tf.test.TestCase):
                                                input_op_fn=input_fn,
                                                bidirectional=True)
     classifier.fit(data, labels)
-    predictions = classifier.predict(np.array(list([[1, 3, 3, 2, 1], [2, 3, 4,
-                                                                      5, 6]])))
+    test_data = np.array(list([[1, 3, 3, 2, 1], [2, 3, 4,
+                                                 5, 6]]), dtype=np.float32)
+    predictions = classifier.predict(test_data)
     self.assertAllClose(predictions, np.array([1, 0]))
 
-  def testDNNAutoencoder(self):
-    import numpy as np
-    iris = datasets.load_iris()
-    autoencoder = learn.TensorFlowDNNAutoencoder(hidden_units=[10, 20])
-    transformed = autoencoder.fit_transform(iris.data[1:2])
-    expected = np.array([[ -3.57627869e-07, 1.17000043e+00, 1.01902664e+00, 1.19209290e-07,
-                            0.00000000e+00, 1.19209290e-07, -5.96046448e-08, -2.38418579e-07,
-                            9.74681854e-01, 1.19209290e-07]])
-    self.assertAllClose(transformed, expected)
+  # def testDNNAutoencoder(self):
+  #   import numpy as np
+  #   iris = datasets.load_iris()
+  #   autoencoder = learn.TensorFlowDNNAutoencoder(hidden_units=[10, 20])
+  #   transformed = autoencoder.fit_transform(iris.data[1:2])
+  #   expected = np.array([[ -3.57627869e-07, 1.17000043e+00, 1.01902664e+00, 1.19209290e-07,
+  #                           0.00000000e+00, 1.19209290e-07, -5.96046448e-08, -2.38418579e-07,
+  #                           9.74681854e-01, 1.19209290e-07]])
+  #   self.assertAllClose(transformed, expected)
 
 
 if __name__ == "__main__":

tensorflow/contrib/learn/python/learn/tests/test_saver.py

@@ -33,10 +33,10 @@ class SaverTest(tf.test.TestCase):
     classifier = learn.TensorFlowLinearClassifier(n_classes=3)
     classifier.fit(iris.data, iris.target)
     classifier.save(path)
-    new_classifier = learn.TensorFlowEstimator.restore(path)
-    self.assertEqual(type(new_classifier), type(classifier))
-    score = accuracy_score(iris.target, new_classifier.predict(iris.data))
-    self.assertGreater(score, 0.5, 'Failed with score = {0}'.format(score))
+#     new_classifier = learn.TensorFlowEstimator.restore(path)
+#     self.assertEqual(type(new_classifier), type(classifier))
+#     score = accuracy_score(iris.target, new_classifier.predict(iris.data))
+#     self.assertGreater(score, 0.5, 'Failed with score = {0}'.format(score))
 
   def testCustomModel(self):
     path = tf.test.get_temp_dir() + '/tmp.saver2'
@@ -49,10 +49,10 @@ class SaverTest(tf.test.TestCase):
     classifier = learn.TensorFlowEstimator(model_fn=custom_model, n_classes=3)
     classifier.fit(iris.data, iris.target)
     classifier.save(path)
-    new_classifier = learn.TensorFlowEstimator.restore(path)
-    self.assertEqual(type(new_classifier), type(classifier))
-    score = accuracy_score(iris.target, new_classifier.predict(iris.data))
-    self.assertGreater(score, 0.5, 'Failed with score = {0}'.format(score))
+#     new_classifier = learn.TensorFlowEstimator.restore(path)
+#     self.assertEqual(type(new_classifier), type(classifier))
+#     score = accuracy_score(iris.target, new_classifier.predict(iris.data))
+#     self.assertGreater(score, 0.5, 'Failed with score = {0}'.format(score))
 
   def testDNN(self):
     path = tf.test.get_temp_dir() + '/tmp_saver3'
@@ -62,10 +62,10 @@ class SaverTest(tf.test.TestCase):
                                                n_classes=3)
     classifier.fit(iris.data, iris.target)
     classifier.save(path)
-    new_classifier = learn.TensorFlowEstimator.restore(path)
-    self.assertEqual(type(new_classifier), type(classifier))
-    score = accuracy_score(iris.target, new_classifier.predict(iris.data))
-    self.assertGreater(score, 0.5, 'Failed with score = {0}'.format(score))
+#     new_classifier = learn.TensorFlowEstimator.restore(path)
+#     self.assertEqual(type(new_classifier), type(classifier))
+#     score = accuracy_score(iris.target, new_classifier.predict(iris.data))
+#     self.assertGreater(score, 0.5, 'Failed with score = {0}'.format(score))
 
   def testNoFolder(self):
     with self.assertRaises(ValueError):
@@ -80,7 +80,7 @@ class SaverTest(tf.test.TestCase):
     classifier.fit(iris.data, iris.target)
     classifier.save(path)
     os.remove(os.path.join(path, 'checkpoint'))
-    with self.assertRaises(ValueError):
+    with self.assertRaises(NotImplementedError):
       learn.TensorFlowEstimator.restore(path)
 
 if __name__ == '__main__':

tensorflow/contrib/learn/python/learn/trainer.py

@@ -1,4 +1,7 @@
-"""Generic trainer for TensorFlow models."""
+"""Generic trainer for TensorFlow models.
+
+This module is deprecated, please use graph_actions.
+"""
 #  Copyright 2015-present The Scikit Flow Authors. All Rights Reserved.
 #
 #  Licensed under the Apache License, Version 2.0 (the "License");
@@ -19,6 +22,8 @@ from __future__ import print_function
 
 from six.moves import xrange  # pylint: disable=redefined-builtin
 
+from tensorflow.python.platform import tf_logging as logging
+
 
 def train(session,
           train_op,
@@ -45,6 +50,8 @@ def train(session,
     summaries: Joined object of all summaries that should be ran.
     feed_params_fn: Feed params function.
   """
+  logging.warning("learn.trainer.train is deprecated. "
+                  "Please use learn.graph_actions.train instead.")
   for step in xrange(steps):
     feed_dict = feed_dict_fn()
     if summaries is not None:
@@ -55,13 +62,5 @@ def train(session,
       global_step_value, loss_value, _ = session.run(
           [global_step, loss, train_op],
           feed_dict=feed_dict)
-    monitor.update(step,
-                   global_step_value,
-                   loss_value,
-                   session,
-                   feed_params_fn,
-                   loss_expression_tensor=loss)
     if summaries is not None and summary_writer and summ is not None:
       summary_writer.add_summary(summ, global_step_value)
-    if monitor.monitor_inducing_stop():
-      break

tensorflow/contrib/learn/python/learn/utils/checkpoints.py

@@ -20,16 +20,23 @@ from __future__ import print_function
 
 import six
 
-from tensorflow.python.framework import dtypes
-from tensorflow.python.ops import script_ops
+from tensorflow.python.ops import gen_io_ops
 from tensorflow.python.ops import state_ops
 from tensorflow.python.ops import variable_scope as vs
+from tensorflow.python.ops import variables
 from tensorflow.python.platform import gfile
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.training import saver
 from tensorflow.python.training import training as train
 
 
+def _get_checkpoint_filename(filepattern):
+  """Returns checkpoint filename given directory or specific filepattern."""
+  if gfile.IsDirectory(filepattern):
+    return saver.latest_checkpoint(filepattern)
+  return filepattern
+
+
 def load_checkpoint(filepattern):
   """Returns CheckpointReader for latest checkpoint.
 
@@ -42,13 +49,11 @@ def load_checkpoint(filepattern):
   Raises:
     ValueError: if checkpoint_dir doesn't have 'checkpoint' file or checkpoints.
   """
-  if gfile.IsDirectory(filepattern):
-    filename = saver.latest_checkpoint(filepattern)
-    if filename is None:
-      raise ValueError("Couldn't find 'checkpoint' file or checkpoints in "
-                       "given directory %s" % filepattern)
-    return train.NewCheckpointReader(filename)
-  return train.NewCheckpointReader(filepattern)
+  filename = _get_checkpoint_filename(filepattern)
+  if filename is None:
+    raise ValueError("Couldn't find 'checkpoint' file or checkpoints in "
+                     "given directory %s" % filepattern)
+  return train.NewCheckpointReader(filename)
 
 
 def load_variable(checkpoint_dir, name):
@@ -83,43 +88,49 @@ def list_variables(checkpoint_dir):
   return result
 
 
-def _checkpoint_initializer(variable, checkpoint_reader, tensor_name):
-  """Assigns variable to value that will be loaded from checkpoint's tensor.
+# pylint: disable=protected-access
+# Currently variable_scope doesn't provide very good APIs to access
+# all variables under scope and retrieve and check existing scopes.
+# TODO(ipolosukhin): Refactor variable_scope module to provide nicer APIs.
+
+
+def _set_checkpoint_initializer(variable, file_pattern, tensor_name, slice_spec,
+                                name="checkpoint_initializer"):
+  """Sets variable initializer to assign op form value in checkpoint's tensor.
 
   Args:
     variable: `Variable` object.
-    checkpoint_reader: `CheckpointReader` object.
+    file_pattern: string, where to load checkpoints from.
     tensor_name: Name of the `Tensor` to load from checkpoint reader.
-
-  Returns:
-    `Tensor` that returns value of `tensor_name` in checkpoint.
-
-  Raises:
-    ValueError: if shape or dtype of `variable` doesn't match with Tensor in
-                checkpoint.
+    slice_spec: Slice specification for loading partitioned variables.
+    name: Name of the operation.
   """
-  # Currently to avoid putting the whole tensor into the graph, this adds a
-  # py_func function to the graph, that will return actual value.
-  # TODO(ipolosukhin): Rewrite this as C++ op, that loads checkpoint at time.
-  tensor = checkpoint_reader.get_tensor(tensor_name)
-  def _tensor():
-    return tensor
-  if not variable.get_shape().is_compatible_with(tensor.shape):
-    raise ValueError(
-        "Shape of variable %s (%s) doesn't match with shape of "
-        "tensor %s (%s) from checkpoint reader." % (
-            variable.name, str(variable.get_shape()),
-            tensor_name, str(tensor.shape)
-        ))
-  if not dtypes.as_dtype(tensor.dtype).is_compatible_with(variable.dtype):
-    raise ValueError(
-        "DType of variable %s (%s) doesn't match with dtype of "
-        "tensor %s (%s) from checkpoint reader." % (
-            variable.name, str(variable.dtype),
-            tensor_name, str(dtypes.as_dtype(tensor.dtype))
-        ))
-  return state_ops.assign(
-      variable, script_ops.py_func(_tensor, [], [tensor.dtype])[0])
+  base_type = variable.dtype.base_dtype
+  restore_op = gen_io_ops._restore_slice(
+      file_pattern,
+      tensor_name,
+      slice_spec,
+      base_type,
+      preferred_shard=-1,
+      name=name)
+  variable._initializer_op = state_ops.assign(variable, restore_op)
+
+
+def _set_variable_or_list_initializer(variable_or_list, file_pattern,
+                                      tensor_name):
+  if isinstance(variable_or_list, (list, tuple)):
+    # A set of slices.
+    slice_name = None
+    for v in variable_or_list:
+      if slice_name is None:
+        slice_name = v._save_slice_info.full_name
+      elif slice_name != v._save_slice_info.full_name:
+        raise ValueError("Slices must all be from the same tensor: %s != %s" %
+                         (slice_name, v._save_slice_info.full_name))
+      _set_checkpoint_initializer(v, file_pattern, tensor_name,
+                                  v._save_slice_info.spec)
+  else:
+    _set_checkpoint_initializer(variable_or_list, file_pattern, tensor_name, "")
 
 
 def init_from_checkpoint(checkpoint_dir, assignment_map):
@@ -133,8 +144,15 @@ def init_from_checkpoint(checkpoint_dir, assignment_map):
       current `scope_name` from `checkpoint_scope_name` with matching variable
       names.
     `'scope_name/variable_name': 'checkpoint_scope_name/some_other_variable'` -
-    will initalize `scope_name/variable_name` variable
-    from `checkpoint_scope_name/some_other_variable`.
+      will initalize `scope_name/variable_name` variable
+      from `checkpoint_scope_name/some_other_variable`.
+    `variable: 'scope_varaible_name'` - will initialize given variable with
+      variable from the checkpoint.
+    `'scope_name/': '/'` - will load all variables in current `scope_name` from
+      checkpoint's root (e.g. no scope).
+
+  Supports loading into partitioned variables, which are represented as
+  '<variable>/part_<part #>'.
 
   Example:
   ```python
@@ -142,13 +160,18 @@ def init_from_checkpoint(checkpoint_dir, assignment_map):
     with tf.variable_scope('test'):
       m = tf.get_variable('my_var')
     with tf.variable_scope('test2'):
-      m = tf.get_variable('my_var')
+      var2 = tf.get_variable('my_var')
     ...
     # Specify which variables to intialize from checkpoint.
     init_from_checkpoint(checkpoint_dir, {
       'test/my_var': 'some_var',
       'test2/', 'some_scope/'})
     ...
+    # Or use `Variable` objects to identify what to initialize.
+    init_from_checkpoint(checkpoint_dir, {
+      var2: 'some_scope/var2',
+    })
+    ...
     # Initialize variables as usual.
     session.run(tf.get_all_variables())
   ```
@@ -163,26 +186,50 @@ def init_from_checkpoint(checkpoint_dir, assignment_map):
     tf.errors.OpError: If missing checkpoints or tensors in checkpoints.
     ValueError: If missing variables in current graph.
   """
+  filepattern = _get_checkpoint_filename(checkpoint_dir)
   reader = load_checkpoint(checkpoint_dir)
   variable_map = reader.get_variable_to_shape_map()
   for current_name, tensor_name in six.iteritems(assignment_map):
     scopes = ""
-    if "/" in current_name:
-      scopes = current_name[:current_name.rindex("/")]
-      current_name = current_name[current_name.rindex("/") + 1:]
-    if current_name:
+    var = None
+    # Check if this is Variable object.
+    if isinstance(current_name, variables.Variable):
+      var = current_name
+    else:
+      var_scope = vs._get_default_variable_store()
+      # Check if this is variable in var_store.
+      var = var_scope._vars.get(current_name, None)
+      # Also check if variable is partitioned as list.
+      if var is None:
+        if current_name + "/part_0" in var_scope._vars:
+          var = []
+          i = 0
+          while current_name + "/part_%d" % i in var_scope._vars:
+            var.append(var_scope._vars[current_name + "/part_%d" % i])
+            i += 1
+    if var is not None:
       # If 1 to 1 mapping was provided, find variable in the scope.
       if tensor_name not in variable_map:
         raise ValueError("Tensor %s is not found in %s checkpoint" % (
             tensor_name, checkpoint_dir
         ))
-      with vs.variable_scope(scopes, reuse=True):
-        var = vs.get_variable(current_name)
-        var._initializer_op = _checkpoint_initializer(var, reader, tensor_name)  # pylint: disable=protected-access
-        logging.info("Initialize variable %s from checkpoint %s with %s" % (
-            var.name, checkpoint_dir, tensor_name
-        ))
+      if isinstance(var, variables.Variable):
+        # Additional at-call-time checks.
+        if not var.get_shape().is_compatible_with(variable_map[tensor_name]):
+          raise ValueError(
+              "Shape of variable %s (%s) doesn't match with shape of "
+              "tensor %s (%s) from checkpoint reader." % (
+                  var.name, str(var.get_shape()),
+                  tensor_name, str(variable_map[tensor_name])
+              ))
+      _set_variable_or_list_initializer(var, filepattern, tensor_name)
+      logging.info("Initialize variable %s from checkpoint %s with %s" % (
+          current_name, checkpoint_dir, tensor_name
+      ))
     else:
+      if "/" in current_name:
+        scopes = current_name[:current_name.rindex("/")]
+        current_name = current_name[current_name.rindex("/") + 1:]
       if not tensor_name.endswith("/"):
         raise ValueError(
             "Assignment map with scope only name (%s) "
@@ -191,21 +238,23 @@ def init_from_checkpoint(checkpoint_dir, assignment_map):
                 scopes, tensor_name
             ))
       # If scope to scope mapping was provided, find all variables in the scope.
-      # TODO(ipolosukhin): Refactor variable_scope module to provide nicer APIs.
-      var_scope = vs._get_default_variable_store()  # pylint: disable=protected-access
-      for var_name in var_scope._vars:  # pylint: disable=protected-access
+      for var_name in var_scope._vars:
         if var_name.startswith(scopes):
           # Lookup name with specified prefix and suffix from current variable.
-          full_tensor_name = tensor_name + var_name[len(scopes) + 1:]
+          # If tensor_name given is '/' (root), don't use it for full name.
+          if tensor_name != "/":
+            full_tensor_name = tensor_name + var_name[len(scopes) + 1:]
+          else:
+            full_tensor_name = var_name[len(scopes) + 1:]
           if full_tensor_name not in variable_map:
             raise ValueError(
                 "Tensor %s (%s in %s) is not found in %s checkpoint" % (
                     full_tensor_name, var_name[len(scopes) + 1:], tensor_name,
                     checkpoint_dir
                 ))
-          var = var_scope._vars[var_name]  # pylint: disable=protected-access
-          var._initializer_op = _checkpoint_initializer(  # pylint: disable=protected-access
-              var, reader, full_tensor_name)
+          var = var_scope._vars[var_name]
+          _set_variable_or_list_initializer(var, filepattern, full_tensor_name)
           logging.info("Initialize variable %s from checkpoint %s with %s" % (
               var_name, checkpoint_dir, tensor_name
           ))
+# pylint: enable=protected-access

tensorflow/contrib/learn/python/learn/utils/checkpoints_test.py

@@ -41,6 +41,24 @@ def _create_checkpoints(sess, checkpoint_dir):
   return v1_value, v2_value, v3_value, v4_value
 
 
+def _create_partition_checkpoints(sess, checkpoint_dir):
+  checkpoint_prefix = os.path.join(checkpoint_dir, "model")
+  checkpoint_state_name = "checkpoint"
+  # TODO(ipolosukhin): Enable this when get_variable partitioning works.
+#   v1 = tf.get_variable("var1", [100, 100],
+#                        partitioner=tf.variable_axis_size_partitioner(axis=0,
+#                                                         max_shard_bytes=512))
+  v1 = tf.create_partitioned_variables(
+      shape=[100, 100], slicing=[5, 1], name="var1",
+      initializer=tf.truncated_normal_initializer(0.5))
+  sess.run(tf.initialize_all_variables())
+  v1_value = sess.run(v1)
+  saver = tf.train.Saver()
+  saver.save(sess, checkpoint_prefix, global_step=0,
+             latest_filename=checkpoint_state_name)
+  return v1_value
+
+
 class CheckpointsTest(tf.test.TestCase):
 
   def testNoCheckpoints(self):
@@ -97,7 +115,7 @@ class CheckpointsTest(tf.test.TestCase):
         })
         checkpoints.init_from_checkpoint(checkpoint_dir, {
             "some_scope/some_other_scope/my2": "var2",
-            "my3": "var3",
+            my3: "var3",
         })
 
         session.run(tf.initialize_all_variables())
@@ -107,7 +125,60 @@ class CheckpointsTest(tf.test.TestCase):
         self.assertAllEqual(my4.eval(session), v4)
 
         # Check that tensors are not explicitly in the graph.
-        self.assertLess(len(str(session.graph.as_graph_def())), 22000)
+        self.assertLess(len(str(session.graph.as_graph_def())), 26000)
+
+  def testInitFromRootCheckpoint(self):
+    checkpoint_dir = self.get_temp_dir()
+    with self.test_session() as session:
+      v1, v2, v3, v4 = _create_checkpoints(session, checkpoint_dir)
+
+    # New graph and session.
+    with tf.Graph().as_default() as g:
+      with self.test_session(graph=g) as session:
+        with tf.variable_scope("some_scope"):
+          my1 = tf.get_variable("var1", [1, 10])
+          my2 = tf.get_variable("var2", [10, 10])
+          my3 = tf.get_variable("var3", [100, 100])
+          with tf.variable_scope("useful_scope"):
+            my4 = tf.get_variable("var4", [9, 9])
+
+        checkpoints.init_from_checkpoint(checkpoint_dir, {
+            "some_scope/": "/",
+        })
+
+        session.run(tf.initialize_all_variables())
+        self.assertAllEqual(my1.eval(session), v1)
+        self.assertAllEqual(my2.eval(session), v2)
+        self.assertAllEqual(my3.eval(session), v3)
+        self.assertAllEqual(my4.eval(session), v4)
+
+  def testInitFromPartitionVar(self):
+    checkpoint_dir = self.get_temp_dir()
+    with self.test_session() as session:
+      v1 = _create_partition_checkpoints(session, checkpoint_dir)
+
+    # New graph and session.
+    with tf.Graph().as_default() as g:
+      with self.test_session(graph=g) as session:
+        with tf.variable_scope("some_scope"):
+          # TODO(ipolosukhin): Enable this when get_variable partitioning works.
+          # Currently get_variable with partitioner doesn't return Variable,
+          # but returns a concat op.
+#           my1 = tf.get_variable(
+#               "my1", [100, 100],
+#               partitioner=tf.variable_axis_size_partitioner(axis=0,
+#                                                          max_shard_bytes=100))
+          my1 = tf.create_partitioned_variables(
+              shape=[100, 100], slicing=[5, 1], name="my1",
+              initializer=tf.truncated_normal_initializer(0.5))
+
+        checkpoints.init_from_checkpoint(checkpoint_dir, {
+            "some_scope/my1": "var1",
+        })
+
+        session.run(tf.initialize_all_variables())
+        my1_values = session.run(my1)
+        self.assertAllEqual(my1_values, v1)
 
   def testInitFromCheckpointMissing(self):
     checkpoint_dir = self.get_temp_dir()
@@ -142,11 +213,6 @@ class CheckpointsTest(tf.test.TestCase):
           checkpoints.init_from_checkpoint(checkpoint_dir, {
               "some_scope/my1": "var1"})
 
-        # DType mismatch.
-        with self.assertRaises(ValueError):
-          checkpoints.init_from_checkpoint(checkpoint_dir, {
-              "some_scope/my2": "var1"})
-
         # Variable 'my1' and 'my2' are missing in given checkpoint scope.
         with self.assertRaises(ValueError):
           checkpoints.init_from_checkpoint(checkpoint_dir, {

tensorflow/contrib/metrics/__init__.py

@@ -146,8 +146,10 @@ from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_mean_rela
 from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_mean_squared_error
 from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_percentage_less
 from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_precision
+from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_precision_at_thresholds
 from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_recall
 from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_recall_at_k
+from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_recall_at_thresholds
 from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_root_mean_squared_error
 from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_sparse_precision_at_k
 from tensorflow.contrib.metrics.python.ops.metric_ops import streaming_sparse_recall_at_k

tensorflow/contrib/metrics/python/ops/metric_ops.py

@@ -28,7 +28,6 @@ from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import check_ops
-from tensorflow.python.ops import logging_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import nn
 from tensorflow.python.ops import sparse_ops
@@ -406,9 +405,9 @@ def streaming_precision(predictions, labels, ignore_mask=None,
 
   Raises:
     ValueError: If the dimensions of `predictions` and `labels` don't match or
-      if `weight` is not `None` and its shape doesn't match `predictions` or
-      if either `metrics_collections` or `updates_collections` are not
-      a list or tuple.
+      if `ignore_mask` is not `None` and its shape doesn't match `predictions`
+      or if either `metrics_collections` or `updates_collections` are not a list
+      or tuple.
   """
   with variable_scope.variable_op_scope(
       [predictions, labels], name, 'precision'):
@@ -485,9 +484,9 @@ def streaming_recall(predictions, labels, ignore_mask=None,
 
   Raises:
     ValueError: If the dimensions of `predictions` and `labels` don't match or
-      if `weight` is not `None` and its shape doesn't match `predictions` or
-      if either `metrics_collections` or `updates_collections` are not
-      a list or tuple.
+      if `ignore_mask` is not `None` and its shape doesn't match `predictions`
+      or if either `metrics_collections` or `updates_collections` are not a list
+      or tuple.
   """
   with variable_scope.variable_op_scope([predictions, labels], name, 'recall'):
     predictions, labels = _remove_squeezable_dimensions(predictions, labels)
@@ -521,6 +520,139 @@ def streaming_recall(predictions, labels, ignore_mask=None,
     return recall, update_op
 
 
+def _tp_fn_tn_fp(predictions, labels, thresholds, ignore_mask=None):
+  """Computes true_positives, false_negatives, true_negatives, false_positives.
+
+  The `_tp_fn_tn_fp` function creates four local variables, `true_positives`,
+  `true_negatives`, `false_positives` and `false_negatives`.
+  `true_positive[i]` is defined` as the number of values in `predictions` above
+  `thresholds[i]` whose corresponding entry in `labels` is `True`.
+  `false_negatives[i]` is defined` as the number of values in `predictions` at
+  most `thresholds[i]` whose corresponding entry in `labels` is `True`.
+  `true_negatives[i]` is defined` as the number of values in `predictions` at
+  most `thresholds[i]` whose corresponding entry in `labels` is `False`.
+  `false_positives[i]` is defined` as the number of values in `predictions`
+  above `thresholds[i]` whose corresponding entry in `labels` is `False`.
+
+  These four variables are updated through the `update_op`.
+  The streaming behavior is that the values of the variables after a few
+  `update_op`s is the same as if the inputs had been concatenated and a single
+  `update_op` had been performed.
+
+  If `ignore_mask` is not `None`, then the increment of the variables is
+  performed using only the elements of `predictions` and `labels` whose
+  corresponding value in `ignore_mask` is `False`.
+
+  Args:
+    predictions: A floating point `Tensor` of arbitrary shape and whose values
+      are in the range `[0, 1]`.
+    labels: A binary `Tensor` whose shape matches `predictions`.
+    thresholds: A python list or tuple of float thresholds in `[0, 1]`.
+    ignore_mask: An optional, binary tensor whose size matches `predictions`.
+
+  Returns:
+    true_positive: A variable of shape [len(thresholds)].
+    false_negative: A variable of shape [len(thresholds)].
+    true_negatives: A variable of shape [len(thresholds)].
+    false_positives: A variable of shape [len(thresholds)].
+    true_positives_update_op: An operation that increments the `true_positives`.
+    false_negative_update_op: An operation that increments the `false_negative`.
+    true_negatives_update_op: An operation that increments the `true_negatives`.
+    false_positives_update_op: An operation that increments the
+      `false_positives`.
+
+  Raises:
+    ValueError: If the shape of `predictions` and `labels` do not match or if
+      `ignore_mask` is not `None` and its shape doesn't match `predictions`
+      or if either `metrics_collections` or `updates_collections` are not a list
+      or tuple.
+  """
+  predictions, labels = _remove_squeezable_dimensions(predictions, labels)
+  predictions.get_shape().assert_is_compatible_with(labels.get_shape())
+
+  num_thresholds = len(thresholds)
+
+  # Reshape predictions and labels to be column vectors
+  predictions = array_ops.reshape(predictions, [-1, 1])
+  labels = array_ops.reshape(labels, [-1, 1])
+
+  # Use static shape if known.
+  num_predictions = predictions.get_shape().as_list()[0]
+
+  # Otherwise use dynamic shape.
+  if num_predictions is None:
+    num_predictions = array_ops.shape(predictions)[0]
+  thresh_tiled = array_ops.tile(
+      array_ops.expand_dims(array_ops.constant(thresholds), [1]),
+      array_ops.pack([1, num_predictions]))
+
+  # Tile the predictions after thresholding them across different thresholds.
+  pred_tiled = math_ops.cast(
+      math_ops.greater(
+          array_ops.tile(
+              array_ops.transpose(predictions), [num_thresholds, 1]),
+          thresh_tiled),
+      dtype=dtypes.int32)
+  # Tile labels by number of thresholds
+  labels_tiled = array_ops.tile(array_ops.transpose(labels),
+                                [num_thresholds, 1])
+
+  true_positives = _create_local('true_positives', shape=[num_thresholds])
+  false_negatives = _create_local('false_negatives', shape=[num_thresholds])
+  true_negatives = _create_local('true_negatives', shape=[num_thresholds])
+  false_positives = _create_local('false_positives', shape=[num_thresholds])
+
+  is_true_positive = math_ops.to_float(
+      math_ops.logical_and(
+          math_ops.equal(labels_tiled, 1), math_ops.equal(pred_tiled, 1)))
+  is_false_negative = math_ops.to_float(
+      math_ops.logical_and(
+          math_ops.equal(labels_tiled, 1), math_ops.equal(pred_tiled, 0)))
+  is_false_positive = math_ops.to_float(
+      math_ops.logical_and(
+          math_ops.equal(labels_tiled, 0), math_ops.equal(pred_tiled, 1)))
+  is_true_negative = math_ops.to_float(
+      math_ops.logical_and(
+          math_ops.equal(labels_tiled, 0), math_ops.equal(pred_tiled, 0)))
+
+  if ignore_mask is not None:
+    ignore_mask = array_ops.reshape(ignore_mask, [-1, 1])
+    mask_tiled = array_ops.tile(array_ops.transpose(ignore_mask),
+                                [num_thresholds, 1])
+
+    labels_tiled.get_shape().assert_is_compatible_with(mask_tiled.get_shape())
+    check_ops.assert_type(mask_tiled, dtypes.bool)
+    is_true_positive = math_ops.select(
+        mask_tiled,
+        array_ops.zeros_like(labels_tiled, dtype=dtypes.float32),
+        is_true_positive)
+    is_false_negative = math_ops.select(
+        mask_tiled,
+        array_ops.zeros_like(labels_tiled, dtype=dtypes.float32),
+        is_false_negative)
+    is_false_positive = math_ops.select(
+        mask_tiled,
+        array_ops.zeros_like(labels_tiled, dtype=dtypes.float32),
+        is_false_positive)
+    is_true_negative = math_ops.select(
+        mask_tiled,
+        array_ops.zeros_like(labels_tiled, dtype=dtypes.float32),
+        is_true_negative)
+
+  true_positives_update_op = state_ops.assign_add(
+      true_positives, math_ops.reduce_sum(is_true_positive, 1))
+  false_negatives_update_op = state_ops.assign_add(
+      false_negatives, math_ops.reduce_sum(is_false_negative, 1))
+  true_negatives_update_op = state_ops.assign_add(
+      true_negatives, math_ops.reduce_sum(is_true_negative, 1))
+  false_positives_update_op = state_ops.assign_add(
+      false_positives, math_ops.reduce_sum(is_false_positive, 1))
+
+  return (true_positives, false_negatives, true_negatives, false_positives,
+          true_positives_update_op, false_negatives_update_op,
+          true_negatives_update_op, false_positives_update_op)
+
+
 def streaming_auc(predictions, labels, ignore_mask=None, num_thresholds=200,
                   metrics_collections=None, updates_collections=None,
                   name=None):
@@ -541,12 +673,12 @@ def streaming_auc(predictions, labels, ignore_mask=None, num_thresholds=200,
 
   To faciliate the estimation of the AUC over a stream of data, the function
   creates an `update_op` operation whose behavior is dependent on the value of
-  `weights`. If `weights` is None, then `update_op` increments the
+  `ignore_mask`. If `ignore_mask` is None, then `update_op` increments the
   `true_positives`, `true_negatives`, `false_positives` and `false_negatives`
   counts with the number of each found in the current `predictions` and `labels`
-  `Tensors`. If `weights` is not `None`, then the increment is performed using
-  only the elements of `predictions` and `labels` whose corresponding value
-  in `ignore_mask` is `False`. In addition to performing the updates,
+  `Tensors`. If `ignore_mask` is not `None`, then the increment is performed
+  using only the elements of `predictions` and `labels` whose corresponding
+  value in `ignore_mask` is `False`. In addition to performing the updates,
   `update_op` also returns the `auc`.
 
   Args:
@@ -563,108 +695,27 @@ def streaming_auc(predictions, labels, ignore_mask=None, num_thresholds=200,
     name: An optional variable_op_scope name.
 
   Returns:
-    auc: A tensor representing the current area-under-curve.
+    auc: A scalar tensor representing the current area-under-curve.
     update_op: An operation that increments the `true_positives`,
       `true_negatives`, `false_positives` and `false_negatives` variables
       appropriately and whose value matches `auc`.
 
   Raises:
     ValueError: If the shape of `predictions` and `labels` do not match or if
-      `weights` is not `None` and its shape doesn't match `values`
-      or if either `metrics_collections` or `updates_collections` are not a list
-      or tuple.
+      `ignore_mask` is not `None` and its shape doesn't match `predictions` or
+      if either `metrics_collections` or `updates_collections` are not a list or
+      tuple.
   """
   with variable_scope.variable_op_scope([predictions, labels], name, 'auc'):
-    predictions, labels = _remove_squeezable_dimensions(predictions, labels)
-    predictions.get_shape().assert_is_compatible_with(labels.get_shape())
+    kepsilon = 1e-7  # to account for floating point imprecisions
+    thresholds = [(i + 1) * 1.0 / (num_thresholds - 1)
+                  for i in range(num_thresholds-2)]
+    thresholds = [0.0 - kepsilon] + thresholds + [1.0 + kepsilon]
 
-    # Reshape predictions and labels to be column vectors
-    logging_ops.Assert(
-        math_ops.equal(
-            array_ops.rank(predictions), 1),
-        ['Input predictions are expected to be a rank 1 tensor. Got ',
-         array_ops.rank(predictions)])
-    logging_ops.Assert(
-        math_ops.equal(
-            array_ops.rank(labels), 1),
-        ['Input labels are expected to be a rank 1 tensor. Got ',
-         array_ops.rank(labels)])
-    predictions = array_ops.reshape(predictions, [-1, 1])
-    labels = array_ops.reshape(labels, [-1, 1])
-
-    kepsilon = 1e-7
-
-    # Use static shape if known.
-    num_predictions = predictions.get_shape().as_list()[0]
-
-    # Otherwise use dynamic shape.
-    if num_predictions is None:
-      num_predictions = array_ops.shape(predictions)[0]
-    thresh_tiled = array_ops.tile(
-        array_ops.expand_dims(
-            math_ops.linspace(0.0 - kepsilon, 1.0 + kepsilon, num_thresholds),
-            [1]),
-        array_ops.pack([1, num_predictions]))
-
-    # Tile the predictions after thresholding them across different thresholds.
-    pred_tiled = math_ops.cast(
-        math_ops.greater(
-            array_ops.tile(
-                array_ops.transpose(predictions), [num_thresholds, 1]),
-            thresh_tiled),
-        dtype=dtypes.int32)
-    # Tile labels by number of thresholds
-    labels_tiled = array_ops.tile(array_ops.transpose(labels),
-                                  [num_thresholds, 1])
-
-    true_positives = _create_local('true_positives', shape=[num_thresholds])
-    false_negatives = _create_local('false_negatives', shape=[num_thresholds])
-    true_negatives = _create_local('true_negatives', shape=[num_thresholds])
-    false_positives = _create_local('false_positives', shape=[num_thresholds])
-
-    is_true_positive = math_ops.to_float(
-        math_ops.logical_and(
-            math_ops.equal(labels_tiled, 1), math_ops.equal(pred_tiled, 1)))
-    is_false_negative = math_ops.to_float(
-        math_ops.logical_and(
-            math_ops.equal(labels_tiled, 1), math_ops.equal(pred_tiled, 0)))
-    is_false_positive = math_ops.to_float(
-        math_ops.logical_and(
-            math_ops.equal(labels_tiled, 0), math_ops.equal(pred_tiled, 1)))
-    is_true_negative = math_ops.to_float(
-        math_ops.logical_and(
-            math_ops.equal(labels_tiled, 0), math_ops.equal(pred_tiled, 0)))
-
-    if ignore_mask is not None:
-      mask_tiled = array_ops.tile(array_ops.transpose(ignore_mask),
-                                  [num_thresholds, 1])
-      labels_tiled.get_shape().assert_is_compatible_with(mask_tiled.get_shape())
-      check_ops.assert_type(mask_tiled, dtypes.bool)
-      is_true_positive = math_ops.select(
-          mask_tiled,
-          array_ops.zeros_like(labels_tiled, dtype=dtypes.float32),
-          is_true_positive)
-      is_false_negative = math_ops.select(
-          mask_tiled,
-          array_ops.zeros_like(labels_tiled, dtype=dtypes.float32),
-          is_false_negative)
-      is_false_positive = math_ops.select(
-          mask_tiled,
-          array_ops.zeros_like(labels_tiled, dtype=dtypes.float32),
-          is_false_positive)
-      is_true_negative = math_ops.select(
-          mask_tiled,
-          array_ops.zeros_like(labels_tiled, dtype=dtypes.float32),
-          is_true_negative)
-
-    true_positives_compute_op = state_ops.assign_add(
-        true_positives, math_ops.reduce_sum(is_true_positive, 1))
-    false_negatives_compute_op = state_ops.assign_add(
-        false_negatives, math_ops.reduce_sum(is_false_negative, 1))
-    true_negatives_compute_op = state_ops.assign_add(
-        true_negatives, math_ops.reduce_sum(is_true_negative, 1))
-    false_positives_compute_op = state_ops.assign_add(
-        false_positives, math_ops.reduce_sum(is_false_positive, 1))
+    (true_positives, false_negatives, true_negatives, false_positives,
+     true_positives_compute_op, false_negatives_compute_op,
+     true_negatives_compute_op, false_positives_compute_op) = _tp_fn_tn_fp(
+         predictions, labels, thresholds, ignore_mask)
 
     epsilon = 1.0e-6
     assert array_ops.squeeze(
@@ -698,6 +749,149 @@ def streaming_auc(predictions, labels, ignore_mask=None, num_thresholds=200,
     return auc, update_op
 
 
+def streaming_precision_at_thresholds(predictions, labels, thresholds,
+                                      ignore_mask=None,
+                                      metrics_collections=None,
+                                      updates_collections=None, name=None):
+  """Computes precision values for different `thresholds` on `predictions`.
+
+  The `streaming_precision_at_thresholds` function creates four local variables,
+  `true_positives`, `true_negatives`, `false_positives` and `false_negatives`
+  for various values of thresholds.
+  `precision[i]` is defined as the number of values in `predictions` above
+  `thresholds[i]` whose corresponding entry in `labels` is `True`
+  (`true_positives[i]`) divided by the number of values in `predictions`
+  above `thresholds[i]` (`true_positives[i] + false_positives[i]`).
+
+  If `ignore_mask` is not None then only values whose corresponding value in
+  `ignore_mask` is `False` are considered.
+
+  `precision` is returned along with an `update_op` whose value equals that of
+  `precision`.
+
+  Args:
+    predictions: A floating point `Tensor` of arbitrary shape and whose values
+      are in the range `[0, 1]`.
+    labels: A binary `Tensor` whose shape matches `predictions`.
+    thresholds: A python list or tuple of float thresholds in `[0, 1]`.
+    ignore_mask: An optional, binary tensor whose size matches `predictions`.
+    metrics_collections: An optional list of collections that `auc` should be
+      added to.
+    updates_collections: An optional list of collections that `update_op` should
+      be added to.
+    name: An optional variable_op_scope name.
+
+  Returns:
+    precision: A float tensor of shape [len(thresholds)].
+    update_op: An operation that increments the `true_positives`,
+      `true_negatives`, `false_positives` and `false_negatives` variables that
+      are used in the computation of `precision`.
+
+  Raises:
+    ValueError: If the shape of `predictions` and `labels` do not match or if
+      `ignore_mask` is not `None` and its shape doesn't match `predictions`
+      or if either `metrics_collections` or `updates_collections` are not a list
+      or tuple.
+  """
+  with variable_scope.variable_op_scope([predictions, labels], name,
+                                        'precision_at_thresholds'):
+    (true_positives, _, _, false_positives, true_positives_compute_op, _, _,
+     false_positives_compute_op,) = _tp_fn_tn_fp(
+         predictions, labels, thresholds, ignore_mask)
+
+    # avoid division by zero
+    epsilon = 1e-7
+    def compute_precision(name):
+      precision = math_ops.div(true_positives,
+                               epsilon + true_positives + false_positives,
+                               name='precision_' + name)
+      return precision
+
+    precision = compute_precision('value')
+    with ops.control_dependencies([true_positives_compute_op,
+                                   false_positives_compute_op]):
+      update_op = compute_precision('update_op')
+
+    if metrics_collections:
+      ops.add_to_collections(metrics_collections, precision)
+
+    if updates_collections:
+      ops.add_to_collections(updates_collections, update_op)
+
+    return precision, update_op
+
+
+def streaming_recall_at_thresholds(predictions, labels, thresholds,
+                                   ignore_mask=None, metrics_collections=None,
+                                   updates_collections=None, name=None):
+  """Computes various recall values for different `thresholds` on `predictions`.
+
+  The `streaming_recall_at_thresholds` function creates four local variables,
+  `true_positives`, `true_negatives`, `false_positives` and `false_negatives`
+  for various values of thresholds.
+  `recall[i]` is defined as the number of values in `predictions` above
+  `thresholds[i]` whose corresponding entry in `labels` is `True`
+  (`true_positives[i]`) divided by the number of True values in `labels`
+  (`true_positives[i] + false_negatives[i]`).
+
+  If `ignore_mask` is not None then only values whose corresponding value in
+  `ignore_mask` is `False` are considered.
+
+  `recall` are returned along with an `update_op` whose value equals that of
+  `recall`.
+
+  Args:
+    predictions: A floating point `Tensor` of arbitrary shape and whose values
+      are in the range `[0, 1]`.
+    labels: A binary `Tensor` whose shape matches `predictions`.
+    thresholds: A python list or tuple of float thresholds in `[0, 1]`.
+    ignore_mask: An optional, binary tensor whose size matches `predictions`.
+    metrics_collections: An optional list of collections that `auc` should be
+      added to.
+    updates_collections: An optional list of collections that `update_op` should
+      be added to.
+    name: An optional variable_op_scope name.
+
+  Returns:
+    recall: A float tensor of shape [len(thresholds)].
+    update_op: An operation that increments the `true_positives`,
+      `true_negatives`, `false_positives` and `false_negatives` variables that
+      are used in the computation of `recall`.
+
+  Raises:
+    ValueError: If the shape of `predictions` and `labels` do not match or if
+      `ignore_mask` is not `None` and its shape doesn't match `predictions`
+      or if either `metrics_collections` or `updates_collections` are not a list
+      or tuple.
+  """
+  with variable_scope.variable_op_scope([predictions, labels], name,
+                                        'recall_at_thresholds'):
+    (true_positives, false_negatives, _, _, true_positives_compute_op,
+     false_negatives_compute_op, _, _,) = _tp_fn_tn_fp(
+         predictions, labels, thresholds, ignore_mask)
+
+    # avoid division by zero
+    epsilon = 1e-7
+    def compute_recall(name):
+      recall = math_ops.div(true_positives,
+                            epsilon + true_positives + false_negatives,
+                            name='recall_' + name)
+      return recall
+
+    recall = compute_recall('value')
+    with ops.control_dependencies([true_positives_compute_op,
+                                   false_negatives_compute_op]):
+      update_op = compute_recall('update_op')
+
+    if metrics_collections:
+      ops.add_to_collections(metrics_collections, recall)
+
+    if updates_collections:
+      ops.add_to_collections(updates_collections, update_op)
+
+    return recall, update_op
+
+
 def streaming_recall_at_k(predictions, labels, k, ignore_mask=None,
                           metrics_collections=None, updates_collections=None,
                           name=None):
@@ -741,9 +935,9 @@ def streaming_recall_at_k(predictions, labels, k, ignore_mask=None,
 
   Raises:
     ValueError: If the dimensions of `predictions` and `labels` don't match or
-      if `weight` is not `None` and its shape doesn't match `predictions` or if
-      either `metrics_collections` or `updates_collections` are not a list or
-      tuple.
+      if `ignore_mask` is not `None` and its shape doesn't match `predictions`
+      or if either `metrics_collections` or `updates_collections` are not a list
+      or tuple.
   """
   in_top_k = math_ops.to_float(nn.in_top_k(predictions, labels, k))
   return streaming_mean(in_top_k, _mask_to_weights(ignore_mask),

tensorflow/contrib/metrics/python/ops/metric_ops_test.py

@@ -713,6 +713,266 @@ class StreamingAUCTest(tf.test.TestCase):
       self.assertAlmostEqual(expected_auc, auc.eval(), 2)
 
 
+class StreamingPrecisionRecallThresholdsTest(tf.test.TestCase):
+
+  def setUp(self):
+    np.random.seed(1)
+    tf.reset_default_graph()
+
+  def testMetricsCollection(self):
+    my_collection_name = '__metrics__'
+    prec, _ = tf.contrib.metrics.streaming_precision_at_thresholds(
+        predictions=tf.ones((10, 1)),
+        labels=tf.ones((10, 1)),
+        thresholds=[0, 0.5, 1.0],
+        metrics_collections=[my_collection_name])
+    rec, _ = tf.contrib.metrics.streaming_recall_at_thresholds(
+        predictions=tf.ones((10, 1)),
+        labels=tf.ones((10, 1)),
+        thresholds=[0, 0.5, 1.0],
+        metrics_collections=[my_collection_name])
+    self.assertListEqual(tf.get_collection(my_collection_name), [prec, rec])
+
+  def testUpdatesCollection(self):
+    my_collection_name = '__updates__'
+    _, precision_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+        predictions=tf.ones((10, 1)),
+        labels=tf.ones((10, 1)),
+        thresholds=[0, 0.5, 1.0],
+        updates_collections=[my_collection_name])
+    _, recall_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+        predictions=tf.ones((10, 1)),
+        labels=tf.ones((10, 1)),
+        thresholds=[0, 0.5, 1.0],
+        updates_collections=[my_collection_name])
+    self.assertListEqual(tf.get_collection(my_collection_name),
+                         [precision_op, recall_op])
+
+  def testValueTensorIsIdempotent(self):
+    predictions = tf.random_uniform((10, 3), maxval=1, dtype=tf.float32, seed=1)
+    labels = tf.random_uniform((10, 3), maxval=1, dtype=tf.int64, seed=1)
+    thresholds = [0, 0.5, 1.0]
+    prec, prec_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+        predictions, labels, thresholds)
+    rec, rec_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+        predictions, labels, thresholds)
+
+    with self.test_session() as sess:
+      sess.run(tf.initialize_local_variables())
+
+      # Run several updates, then verify idempotency.
+      sess.run([prec_op, rec_op])
+      initial_prec = prec.eval()
+      initial_rec = rec.eval()
+      for _ in range(10):
+        sess.run([prec_op, rec_op])
+        self.assertAllClose(initial_prec, prec.eval())
+        self.assertAllClose(initial_rec, rec.eval())
+
+  def testEffectivelyEquivalentShapes(self):
+    inputs = np.random.randint(0, 2, size=(100, 1))
+
+    with self.test_session() as sess:
+      predictions = tf.constant(inputs, dtype=tf.float32, shape=(100,))
+      labels = tf.constant(inputs, shape=(100, 1))
+      thresholds = [0.5]
+      prec, prec_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+          predictions, labels, thresholds)
+      rec, rec_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+          predictions, labels, thresholds)
+
+      sess.run(tf.initialize_local_variables())
+      sess.run([prec_op, rec_op])
+
+      self.assertEqual(1, prec.eval())
+      self.assertEqual(1, rec.eval())
+
+  def testAllCorrect(self):
+    inputs = np.random.randint(0, 2, size=(100, 1))
+
+    with self.test_session() as sess:
+      predictions = tf.constant(inputs, dtype=tf.float32)
+      labels = tf.constant(inputs)
+      thresholds = [0.5]
+      prec, prec_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+          predictions, labels, thresholds)
+      rec, rec_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+          predictions, labels, thresholds)
+
+      sess.run(tf.initialize_local_variables())
+      sess.run([prec_op, rec_op])
+
+      self.assertEqual(1, prec.eval())
+      self.assertEqual(1, rec.eval())
+
+  def testSomeCorrect(self):
+    with self.test_session() as sess:
+      predictions = tf.constant([1, 0, 1, 0], shape=(1, 4), dtype=tf.float32)
+      labels = tf.constant([0, 1, 1, 0], shape=(1, 4))
+      thresholds = [0.5]
+      prec, prec_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+          predictions, labels, thresholds)
+      rec, rec_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+          predictions, labels, thresholds)
+
+      sess.run(tf.initialize_local_variables())
+      sess.run([prec_op, rec_op])
+
+      self.assertAlmostEqual(0.5, prec.eval())
+      self.assertAlmostEqual(0.5, rec.eval())
+
+  def testAllIncorrect(self):
+    inputs = np.random.randint(0, 2, size=(100, 1))
+
+    with self.test_session() as sess:
+      predictions = tf.constant(inputs, dtype=tf.float32)
+      labels = tf.constant(1 - inputs, dtype=tf.float32)
+      thresholds = [0.5]
+      prec, prec_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+          predictions, labels, thresholds)
+      rec, rec_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+          predictions, labels, thresholds)
+
+      sess.run(tf.initialize_local_variables())
+      sess.run([prec_op, rec_op])
+
+      self.assertAlmostEqual(0, prec.eval())
+      self.assertAlmostEqual(0, rec.eval())
+
+  def testIgnoreMask(self):
+    with self.test_session() as sess:
+      predictions = tf.constant([[1, 0], [1, 0]], shape=(2, 2),
+                                dtype=tf.float32)
+      labels = tf.constant([[0, 1], [1, 0]], shape=(2, 2))
+      ignore_mask = tf.constant([[True, True], [False, False]], shape=(2, 2))
+      thresholds = [0.5, 1.1]
+      prec, prec_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+          predictions, labels, thresholds, ignore_mask=ignore_mask)
+      rec, rec_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+          predictions, labels, thresholds, ignore_mask=ignore_mask)
+
+      [prec_low, prec_high] = tf.split(0, 2, prec)
+      prec_low = tf.reshape(prec_low, shape=())
+      prec_high = tf.reshape(prec_high, shape=())
+      [rec_low, rec_high] = tf.split(0, 2, rec)
+      rec_low = tf.reshape(rec_low, shape=())
+      rec_high = tf.reshape(rec_high, shape=())
+
+      sess.run(tf.initialize_local_variables())
+      sess.run([prec_op, rec_op])
+
+      self.assertAlmostEqual(1.0, prec_low.eval(), places=5)
+      self.assertAlmostEqual(0.0, prec_high.eval(), places=5)
+      self.assertAlmostEqual(1.0, rec_low.eval(), places=5)
+      self.assertAlmostEqual(0.0, rec_high.eval(), places=5)
+
+  def testExtremeThresholds(self):
+    with self.test_session() as sess:
+      predictions = tf.constant([1, 0, 1, 0], shape=(1, 4), dtype=tf.float32)
+      labels = tf.constant([0, 1, 1, 1], shape=(1, 4))
+      thresholds = [-1.0, 2.0]  # lower/higher than any values
+      prec, prec_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+          predictions, labels, thresholds)
+      rec, rec_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+          predictions, labels, thresholds)
+
+      [prec_low, prec_high] = tf.split(0, 2, prec)
+      [rec_low, rec_high] = tf.split(0, 2, rec)
+
+      sess.run(tf.initialize_local_variables())
+      sess.run([prec_op, rec_op])
+
+      self.assertAlmostEqual(0.75, prec_low.eval())
+      self.assertAlmostEqual(0.0, prec_high.eval())
+      self.assertAlmostEqual(1.0, rec_low.eval())
+      self.assertAlmostEqual(0.0, rec_high.eval())
+
+  def testZeroLabelsPredictions(self):
+    with self.test_session() as sess:
+      predictions = tf.zeros([4], dtype=tf.float32)
+      labels = tf.zeros([4])
+      thresholds = [0.5]
+      prec, prec_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+          predictions, labels, thresholds)
+      rec, rec_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+          predictions, labels, thresholds)
+
+      sess.run(tf.initialize_local_variables())
+      sess.run([prec_op, rec_op])
+
+      self.assertAlmostEqual(0, prec.eval(), 6)
+      self.assertAlmostEqual(0, rec.eval(), 6)
+
+  def testWithMultipleUpdates(self):
+    num_samples = 5000
+    batch_size = 10
+    num_batches = num_samples / batch_size
+
+    # Create the labels and data.
+    labels = np.random.randint(0, 2, size=(num_samples, 1))
+    noise = np.random.normal(0.0, scale=0.2, size=(num_samples, 1))
+    predictions = 0.4 + 0.2 * labels + noise
+    predictions[predictions > 1] = 1
+    predictions[predictions < 0] = 0
+    thresholds = [0.3]
+
+    tp = 0
+    fp = 0
+    fn = 0
+    tn = 0
+    for i in range(num_samples):
+      if predictions[i] > thresholds[0]:
+        if labels[i] == 1:
+          tp += 1
+        else:
+          fp += 1
+      else:
+        if labels[i] == 1:
+          fn += 1
+        else:
+          tn += 1
+    epsilon = 1e-7
+    expected_prec = tp / (epsilon + tp + fp)
+    expected_rec = tp / (epsilon + tp + fn)
+
+    labels = labels.astype(np.float32)
+    predictions = predictions.astype(np.float32)
+
+    with self.test_session() as sess:
+      # Reshape the data so its easy to queue up:
+      predictions_batches = predictions.reshape((batch_size, num_batches))
+      labels_batches = labels.reshape((batch_size, num_batches))
+
+      # Enqueue the data:
+      predictions_queue = tf.FIFOQueue(num_batches, dtypes=tf.float32,
+                                       shapes=(batch_size,))
+      labels_queue = tf.FIFOQueue(num_batches, dtypes=tf.float32,
+                                  shapes=(batch_size,))
+
+      for i in range(int(num_batches)):
+        tf_prediction = tf.constant(predictions_batches[:, i])
+        tf_label = tf.constant(labels_batches[:, i])
+        sess.run([predictions_queue.enqueue(tf_prediction),
+                  labels_queue.enqueue(tf_label)])
+
+      tf_predictions = predictions_queue.dequeue()
+      tf_labels = labels_queue.dequeue()
+
+      prec, prec_op = tf.contrib.metrics.streaming_precision_at_thresholds(
+          tf_predictions, tf_labels, thresholds)
+      rec, rec_op = tf.contrib.metrics.streaming_recall_at_thresholds(
+          tf_predictions, tf_labels, thresholds)
+
+      sess.run(tf.initialize_local_variables())
+      for _ in range(int(num_samples / batch_size)):
+        sess.run([prec_op, rec_op])
+      # Since this is only approximate, we can't expect a 6 digits match.
+      # Although with higher number of samples/thresholds we should see the
+      # accuracy improving
+      self.assertAlmostEqual(expected_prec, prec.eval(), 2)
+      self.assertAlmostEqual(expected_rec, rec.eval(), 2)
+
+
 class StreamingRecallAtKTest(tf.test.TestCase):
 
   def setUp(self):

tensorflow/core/framework/types.cc

@@ -304,6 +304,7 @@ int DataTypeSize(DataType dt) {
     return sizeof(T);
   switch (dt) {
     TF_CALL_POD_TYPES(CASE);
+    TF_CALL_QUANTIZED_TYPES(CASE);
     default:
       return 0;
   }

tensorflow/core/graph/graph_def_builder.h

@@ -50,10 +50,10 @@ namespace tensorflow {
 // You call it like:
 //   GraphDefBuilder b;
 //   using namespace ::tensorflow::ops;  // NOLINT(build/namespaces)
-//   Node* a = Const(7, b.opts());
+//   Node* na = Const(7, b.opts());
 //   // Note: WithName() returns a copy, opts is unchanged.
-//   Node* b = Const(5, b.opts().WithName("control-input"));
-//   Node* c = Identity(a, b.opts().WithControlInput(b));
+//   Node* nb = Const(5, b.opts().WithName("control-input"));
+//   Node* nc = Identity(na, b.opts().WithControlInput(nb));
 //   GraphDef graph_def;
 //   Status status = b.ToGraphDef(&graph_def);
 //   if (!status.ok()) { /* Handle error */ }

tensorflow/core/kernels/random_op_test.cc

@@ -25,12 +25,12 @@ namespace tensorflow {
 
 Tensor VecShape(int64 v) {
   if (v >= std::numeric_limits<int32>::max()) {
-    Tensor shape(DT_INT32, TensorShape({1}));
+    Tensor shape(DT_INT64, TensorShape({1}));
     shape.vec<int64>()(0) = v;
     return shape;
   } else {
-    Tensor shape(DT_INT64, TensorShape({1}));
-    shape.vec<int64>()(0) = v;
+    Tensor shape(DT_INT32, TensorShape({1}));
+    shape.vec<int32>()(0) = v;
     return shape;
   }
 }

tensorflow/core/kernels/tensor_array_ops.cc

@@ -31,7 +31,6 @@ limitations under the License.
 #include "tensorflow/core/kernels/concat_lib.h"
 #include "tensorflow/core/kernels/split_lib.h"
 #include "tensorflow/core/kernels/tensor_array.h"
-#include "tensorflow/core/lib/core/refcount.h"
 #include "tensorflow/core/lib/core/errors.h"
 #include "tensorflow/core/lib/core/refcount.h"
 #include "tensorflow/core/lib/strings/strcat.h"
@@ -235,7 +234,7 @@ class TensorArrayGradOp : public TensorArrayCreationOp {
     }
 
     auto creator = [this, tensor_array, array_size,
-                    tensor_array_output_handle](TensorArray** ret) {
+                    tensor_array_output_handle](TensorArray** ret) -> Status {
       *ret = new TensorArray(
           tensor_array->ElemType(), *tensor_array_output_handle, array_size,
           false /* dynamic_size */, true /* multiple_writes_aggregate */,

tensorflow/core/platform/types.h

@@ -20,7 +20,7 @@ limitations under the License.
 #include "tensorflow/core/platform/platform.h"
 
 // Include appropriate platform-dependent implementations
-#if defined(PLATFORM_GOOGLE)
+#if defined(PLATFORM_GOOGLE) || defined(GOOGLE_INTEGRAL_TYPES)
 #include "tensorflow/core/platform/google/integral_types.h"
 #elif defined(PLATFORM_POSIX) || defined(PLATFORM_POSIX_ANDROID) || \
     defined(PLATFORM_GOOGLE_ANDROID)

tensorflow/examples/skflow/digits.py

@@ -49,7 +49,7 @@ def conv_model(X, y):
     features = tf.reshape(features, [-1, 12])
     return learn.models.logistic_regression(features, y)
 
-val_monitor = monitors.ValidationMonitor(X_val, y_val, n_classes=10, print_steps=50)
+val_monitor = monitors.ValidationMonitor(X_val, y_val, every_n_steps=50)
 # Create a classifier, train and predict.
 classifier = learn.TensorFlowEstimator(model_fn=conv_model, n_classes=10,
                                         steps=1000, learning_rate=0.05,

tensorflow/examples/skflow/iris_val_based_early_stopping.py

@@ -30,8 +30,7 @@ X_train, X_test, y_train, y_test = train_test_split(iris.data,
 X_train, X_val, y_train, y_val = train_test_split(X_train, y_train,
                                                   test_size=0.2, random_state=42)
 val_monitor = learn.monitors.ValidationMonitor(X_val, y_val,
-                                                early_stopping_rounds=200,
-                                                n_classes=3)
+                                               early_stopping_rounds=200)
 
 # classifier with early stopping on training data
 classifier1 = learn.TensorFlowDNNClassifier(hidden_units=[10, 20, 10],

tensorflow/g3doc/api_docs/python/contrib.distributions.md

@@ -219,10 +219,17 @@ Generate `n` samples.
 Base class for continuous probability distributions.
 
 `ContinuousDistribution` defines the API for the likelihood functions `pdf`
-and `log_pdf` of continuous probability distributions.
+and `log_pdf` of continuous probability distributions, and a property
+`is_reparameterized` (returning `True` or `False`) which describes
+whether the samples of this distribution are calculated in a differentiable
+way from a non-parameterized distribution.  For example, the `Normal`
+distribution with parameters `mu` and `sigma` is reparameterized as
 
-Subclasses must override both `pdf` and `log_pdf` but one can call this base
-class's implementation.
+```Normal(mu, sigma) = sigma * Normal(0, 1) + mu```
+
+Subclasses must override `pdf` and `log_pdf` but one can call this base
+class's implementation.  They must also override the `is_reparameterized`
+property.
 
 See `BaseDistribution` for more information on the API for probability
 distributions.
@@ -300,6 +307,13 @@ Same meaning as `batch_shape`. May be only partially defined.
 Same meaning as `event_shape`. May be only partially defined.
 
 
+- - -
+
+#### `tf.contrib.distributions.ContinuousDistribution.is_reparameterized` {#ContinuousDistribution.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.ContinuousDistribution.log_cdf(value, name='log_cdf')` {#ContinuousDistribution.log_cdf}
@@ -609,6 +623,13 @@ where digamma(alpha) is the digamma function.
 
 
 
+- - -
+
+#### `tf.contrib.distributions.Chi2.is_reparameterized` {#Chi2.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Chi2.log_cdf(x, name='log_cdf')` {#Chi2.log_cdf}
@@ -798,6 +819,13 @@ where digamma(alpha) is the digamma function.
 
 
 
+- - -
+
+#### `tf.contrib.distributions.Exponential.is_reparameterized` {#Exponential.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Exponential.lam` {#Exponential.lam}
@@ -871,20 +899,20 @@ Log pdf of observations in `x` under these Gamma distribution(s).
 
 #### `tf.contrib.distributions.Exponential.sample(n, seed=None, name=None)` {#Exponential.sample}
 
-Generate `n` samples.
+Sample `n` observations from the Exponential Distributions.
 
 ##### Args:
 
 
-*  <b>`n`</b>: scalar. Number of samples to draw from each distribution.
-*  <b>`seed`</b>: Python integer seed for RNG
-*  <b>`name`</b>: name to give to the op.
+*  <b>`n`</b>: `Scalar`, type int32, the number of observations to sample.
+*  <b>`seed`</b>: Python integer, the random seed.
+*  <b>`name`</b>: The name to give this op.
 
 ##### Returns:
 
 
-*  <b>`samples`</b>: a `Tensor` of shape `(n,) + self.batch_shape + self.event_shape`
-      with values of type `self.dtype`.
+*  <b>`samples`</b>: `[n, ...]`, a `Tensor` of `n` samples for each
+    of the distributions determined by the hyperparameters.
 
 
 - - -
@@ -1025,6 +1053,13 @@ where digamma(alpha) is the digamma function.
 
 
 
+- - -
+
+#### `tf.contrib.distributions.Gamma.is_reparameterized` {#Gamma.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Gamma.log_cdf(x, name='log_cdf')` {#Gamma.log_cdf}
@@ -1226,6 +1261,13 @@ The entropy of Gaussian distribution(s).
 *  <b>`entropy`</b>: tensor of dtype `dtype`, the entropy.
 
 
+- - -
+
+#### `tf.contrib.distributions.Gaussian.is_reparameterized` {#Gaussian.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Gaussian.log_cdf(x, name=None)` {#Gaussian.log_cdf}
@@ -1465,6 +1507,13 @@ The entropy of Student t distribution(s).
 
 
 
+- - -
+
+#### `tf.contrib.distributions.StudentT.is_reparameterized` {#StudentT.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.StudentT.log_cdf(value, name='log_cdf')` {#StudentT.log_cdf}
@@ -1697,6 +1746,13 @@ The entropy of Uniform distribution(s).
 
 
 
+- - -
+
+#### `tf.contrib.distributions.Uniform.is_reparameterized` {#Uniform.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Uniform.log_cdf(x, name='log_cdf')` {#Uniform.log_cdf}
@@ -1913,6 +1969,13 @@ The entropies of these Multivariate Normals.
 *  <b>`entropy`</b>: tensor of dtype `dtype`, the entropies.
 
 
+- - -
+
+#### `tf.contrib.distributions.MultivariateNormal.is_reparameterized` {#MultivariateNormal.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.MultivariateNormal.log_pdf(x, name=None)` {#MultivariateNormal.log_pdf}

tensorflow/g3doc/api_docs/python/contrib.metrics.md

@@ -242,9 +242,9 @@ values in `ignore_mask` are `False`. In addition to performing the updates,
 
 
 *  <b>`ValueError`</b>: If the dimensions of `predictions` and `labels` don't match or
-    if `weight` is not `None` and its shape doesn't match `predictions` or
-    if either `metrics_collections` or `updates_collections` are not
-    a list or tuple.
+    if `ignore_mask` is not `None` and its shape doesn't match `predictions`
+    or if either `metrics_collections` or `updates_collections` are not a list
+    or tuple.
 
 
 - - -
@@ -295,9 +295,9 @@ values in `ignore_mask` are `False`. In addition to performing the updates,
 
 
 *  <b>`ValueError`</b>: If the dimensions of `predictions` and `labels` don't match or
-    if `weight` is not `None` and its shape doesn't match `predictions` or
-    if either `metrics_collections` or `updates_collections` are not
-    a list or tuple.
+    if `ignore_mask` is not `None` and its shape doesn't match `predictions`
+    or if either `metrics_collections` or `updates_collections` are not a list
+    or tuple.
 
 
 - - -
@@ -321,12 +321,12 @@ numbers of thresholds more closely approximating the true AUC.
 
 To faciliate the estimation of the AUC over a stream of data, the function
 creates an `update_op` operation whose behavior is dependent on the value of
-`weights`. If `weights` is None, then `update_op` increments the
+`ignore_mask`. If `ignore_mask` is None, then `update_op` increments the
 `true_positives`, `true_negatives`, `false_positives` and `false_negatives`
 counts with the number of each found in the current `predictions` and `labels`
-`Tensors`. If `weights` is not `None`, then the increment is performed using
-only the elements of `predictions` and `labels` whose corresponding value
-in `ignore_mask` is `False`. In addition to performing the updates,
+`Tensors`. If `ignore_mask` is not `None`, then the increment is performed
+using only the elements of `predictions` and `labels` whose corresponding
+value in `ignore_mask` is `False`. In addition to performing the updates,
 `update_op` also returns the `auc`.
 
 ##### Args:
@@ -347,7 +347,7 @@ in `ignore_mask` is `False`. In addition to performing the updates,
 ##### Returns:
 
 
-*  <b>`auc`</b>: A tensor representing the current area-under-curve.
+*  <b>`auc`</b>: A scalar tensor representing the current area-under-curve.
 *  <b>`update_op`</b>: An operation that increments the `true_positives`,
     `true_negatives`, `false_positives` and `false_negatives` variables
     appropriately and whose value matches `auc`.
@@ -356,9 +356,9 @@ in `ignore_mask` is `False`. In addition to performing the updates,
 
 
 *  <b>`ValueError`</b>: If the shape of `predictions` and `labels` do not match or if
-    `weights` is not `None` and its shape doesn't match `values`
-    or if either `metrics_collections` or `updates_collections` are not a list
-    or tuple.
+    `ignore_mask` is not `None` and its shape doesn't match `predictions` or
+    if either `metrics_collections` or `updates_collections` are not a list or
+    tuple.
 
 
 - - -
@@ -411,9 +411,9 @@ recall value.
 
 
 *  <b>`ValueError`</b>: If the dimensions of `predictions` and `labels` don't match or
-    if `weight` is not `None` and its shape doesn't match `predictions` or if
-    either `metrics_collections` or `updates_collections` are not a list or
-    tuple.
+    if `ignore_mask` is not `None` and its shape doesn't match `predictions`
+    or if either `metrics_collections` or `updates_collections` are not a list
+    or tuple.
 
 
 - - -

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.distributions.Chi2.md

@@ -100,6 +100,13 @@ where digamma(alpha) is the digamma function.
 
 
 
+- - -
+
+#### `tf.contrib.distributions.Chi2.is_reparameterized` {#Chi2.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Chi2.log_cdf(x, name='log_cdf')` {#Chi2.log_cdf}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.distributions.ContinuousDistribution.md

@@ -1,10 +1,17 @@
 Base class for continuous probability distributions.
 
 `ContinuousDistribution` defines the API for the likelihood functions `pdf`
-and `log_pdf` of continuous probability distributions.
+and `log_pdf` of continuous probability distributions, and a property
+`is_reparameterized` (returning `True` or `False`) which describes
+whether the samples of this distribution are calculated in a differentiable
+way from a non-parameterized distribution.  For example, the `Normal`
+distribution with parameters `mu` and `sigma` is reparameterized as
 
-Subclasses must override both `pdf` and `log_pdf` but one can call this base
-class's implementation.
+```Normal(mu, sigma) = sigma * Normal(0, 1) + mu```
+
+Subclasses must override `pdf` and `log_pdf` but one can call this base
+class's implementation.  They must also override the `is_reparameterized`
+property.
 
 See `BaseDistribution` for more information on the API for probability
 distributions.
@@ -82,6 +89,13 @@ Same meaning as `batch_shape`. May be only partially defined.
 Same meaning as `event_shape`. May be only partially defined.
 
 
+- - -
+
+#### `tf.contrib.distributions.ContinuousDistribution.is_reparameterized` {#ContinuousDistribution.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.ContinuousDistribution.log_cdf(value, name='log_cdf')` {#ContinuousDistribution.log_cdf}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.distributions.Exponential.md

@@ -93,6 +93,13 @@ where digamma(alpha) is the digamma function.
 
 
 
+- - -
+
+#### `tf.contrib.distributions.Exponential.is_reparameterized` {#Exponential.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Exponential.lam` {#Exponential.lam}
@@ -166,20 +173,20 @@ Log pdf of observations in `x` under these Gamma distribution(s).
 
 #### `tf.contrib.distributions.Exponential.sample(n, seed=None, name=None)` {#Exponential.sample}
 
-Generate `n` samples.
+Sample `n` observations from the Exponential Distributions.
 
 ##### Args:
 
 
-*  <b>`n`</b>: scalar. Number of samples to draw from each distribution.
-*  <b>`seed`</b>: Python integer seed for RNG
-*  <b>`name`</b>: name to give to the op.
+*  <b>`n`</b>: `Scalar`, type int32, the number of observations to sample.
+*  <b>`seed`</b>: Python integer, the random seed.
+*  <b>`name`</b>: The name to give this op.
 
 ##### Returns:
 
 
-*  <b>`samples`</b>: a `Tensor` of shape `(n,) + self.batch_shape + self.event_shape`
-      with values of type `self.dtype`.
+*  <b>`samples`</b>: `[n, ...]`, a `Tensor` of `n` samples for each
+    of the distributions determined by the hyperparameters.
 
 
 - - -

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.distributions.Gamma.md

@@ -124,6 +124,13 @@ where digamma(alpha) is the digamma function.
 
 
 
+- - -
+
+#### `tf.contrib.distributions.Gamma.is_reparameterized` {#Gamma.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Gamma.log_cdf(x, name='log_cdf')` {#Gamma.log_cdf}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.distributions.Gaussian.md

@@ -105,6 +105,13 @@ The entropy of Gaussian distribution(s).
 *  <b>`entropy`</b>: tensor of dtype `dtype`, the entropy.
 
 
+- - -
+
+#### `tf.contrib.distributions.Gaussian.is_reparameterized` {#Gaussian.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Gaussian.log_cdf(x, name=None)` {#Gaussian.log_cdf}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.distributions.MultivariateNormal.md

@@ -125,6 +125,13 @@ The entropies of these Multivariate Normals.
 *  <b>`entropy`</b>: tensor of dtype `dtype`, the entropies.
 
 
+- - -
+
+#### `tf.contrib.distributions.MultivariateNormal.is_reparameterized` {#MultivariateNormal.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.MultivariateNormal.log_pdf(x, name=None)` {#MultivariateNormal.log_pdf}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.distributions.StudentT.md

@@ -137,6 +137,13 @@ The entropy of Student t distribution(s).
 
 
 
+- - -
+
+#### `tf.contrib.distributions.StudentT.is_reparameterized` {#StudentT.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.StudentT.log_cdf(value, name='log_cdf')` {#StudentT.log_cdf}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.distributions.Uniform.md

@@ -126,6 +126,13 @@ The entropy of Uniform distribution(s).
 
 
 
+- - -
+
+#### `tf.contrib.distributions.Uniform.is_reparameterized` {#Uniform.is_reparameterized}
+
+
+
+
 - - -
 
 #### `tf.contrib.distributions.Uniform.log_cdf(x, name='log_cdf')` {#Uniform.log_cdf}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.BaseEstimator.md

@@ -12,7 +12,7 @@ Parameters:
   model_dir: Directory to save model parameters, graph and etc.
 - - -
 
-#### `tf.contrib.learn.BaseEstimator.__init__(model_dir=None)` {#BaseEstimator.__init__}
+#### `tf.contrib.learn.BaseEstimator.__init__(model_dir=None, config=None)` {#BaseEstimator.__init__}
 
 
 
@@ -49,7 +49,7 @@ Evaluates given model with provided evaluation data.
 
 - - -
 
-#### `tf.contrib.learn.BaseEstimator.fit(x, y, steps, batch_size=32, monitor=None)` {#BaseEstimator.fit}
+#### `tf.contrib.learn.BaseEstimator.fit(x, y, steps, batch_size=32, monitors=None)` {#BaseEstimator.fit}
 
 Trains a model given training data X and y.
 
@@ -64,8 +64,8 @@ Trains a model given training data X and y.
      (class labels in classification, real numbers in regression).
 *  <b>`steps`</b>: number of steps to train model for.
 *  <b>`batch_size`</b>: minibatch size to use on the input, defaults to 32.
-*  <b>`monitor`</b>: monitor object to print training progress and invoke
-           early stopping.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
 
 ##### Returns:
 
@@ -92,7 +92,14 @@ params : mapping of string to any
 
 - - -
 
-#### `tf.contrib.learn.BaseEstimator.partial_fit(x, y, steps=1, batch_size=32, monitor=None)` {#BaseEstimator.partial_fit}
+#### `tf.contrib.learn.BaseEstimator.model_dir` {#BaseEstimator.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.BaseEstimator.partial_fit(x, y, steps=1, batch_size=32, monitors=None)` {#BaseEstimator.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -115,8 +122,8 @@ to converge, and you want to split up training into subparts.
     (class label in classification, real numbers in regression).
 *  <b>`steps`</b>: number of steps to train model for.
 *  <b>`batch_size`</b>: minibatch size to use on the input, defaults to 32.
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke
-           early stopping.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
 
 ##### Returns:
 
@@ -151,7 +158,7 @@ Returns prediction probabilities for given features (classification).
 
 
 *  <b>`x`</b>: features.
-*  <b>`batch_size`</b>: OVerride default batch size.
+*  <b>`batch_size`</b>: Override default batch size.
 
 ##### Returns:
 
@@ -176,7 +183,7 @@ self
 
 - - -
 
-#### `tf.contrib.learn.BaseEstimator.train(input_fn, steps, monitor=None)` {#BaseEstimator.train}
+#### `tf.contrib.learn.BaseEstimator.train(input_fn, steps, monitors=None)` {#BaseEstimator.train}
 
 Trains a model given input builder function.
 
@@ -186,8 +193,8 @@ Trains a model given input builder function.
 *  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
             dict and Tensor.
 *  <b>`steps`</b>: number of steps to train model for.
-*  <b>`monitor`</b>: monitor object to print training progress and invoke
-           early stopping.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
 
 ##### Returns:
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.Estimator.md

@@ -15,9 +15,10 @@ Parameters:
                (like tf.train.GradientDescentOptimizer).
   clip_gradients: clip_norm value for call to `clip_by_global_norm`. None
                   denotes no gradient clipping.
+  config: Configuration object.
 - - -
 
-#### `tf.contrib.learn.Estimator.__init__(model_fn=None, model_dir=None, classification=True, learning_rate=0.01, optimizer='SGD', clip_gradients=None)` {#Estimator.__init__}
+#### `tf.contrib.learn.Estimator.__init__(model_fn=None, model_dir=None, classification=True, learning_rate=0.1, optimizer='Adagrad', clip_gradients=None, config=None)` {#Estimator.__init__}
 
 
 
@@ -54,7 +55,7 @@ Evaluates given model with provided evaluation data.
 
 - - -
 
-#### `tf.contrib.learn.Estimator.fit(x, y, steps, batch_size=32, monitor=None)` {#Estimator.fit}
+#### `tf.contrib.learn.Estimator.fit(x, y, steps, batch_size=32, monitors=None)` {#Estimator.fit}
 
 Trains a model given training data X and y.
 
@@ -69,8 +70,8 @@ Trains a model given training data X and y.
      (class labels in classification, real numbers in regression).
 *  <b>`steps`</b>: number of steps to train model for.
 *  <b>`batch_size`</b>: minibatch size to use on the input, defaults to 32.
-*  <b>`monitor`</b>: monitor object to print training progress and invoke
-           early stopping.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
 
 ##### Returns:
 
@@ -97,7 +98,14 @@ params : mapping of string to any
 
 - - -
 
-#### `tf.contrib.learn.Estimator.partial_fit(x, y, steps=1, batch_size=32, monitor=None)` {#Estimator.partial_fit}
+#### `tf.contrib.learn.Estimator.model_dir` {#Estimator.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.Estimator.partial_fit(x, y, steps=1, batch_size=32, monitors=None)` {#Estimator.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -120,8 +128,8 @@ to converge, and you want to split up training into subparts.
     (class label in classification, real numbers in regression).
 *  <b>`steps`</b>: number of steps to train model for.
 *  <b>`batch_size`</b>: minibatch size to use on the input, defaults to 32.
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke
-           early stopping.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
 
 ##### Returns:
 
@@ -156,7 +164,7 @@ Returns prediction probabilities for given features (classification).
 
 
 *  <b>`x`</b>: features.
-*  <b>`batch_size`</b>: OVerride default batch size.
+*  <b>`batch_size`</b>: Override default batch size.
 
 ##### Returns:
 
@@ -181,7 +189,7 @@ self
 
 - - -
 
-#### `tf.contrib.learn.Estimator.train(input_fn, steps, monitor=None)` {#Estimator.train}
+#### `tf.contrib.learn.Estimator.train(input_fn, steps, monitors=None)` {#Estimator.train}
 
 Trains a model given input builder function.
 
@@ -191,8 +199,8 @@ Trains a model given input builder function.
 *  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
             dict and Tensor.
 *  <b>`steps`</b>: number of steps to train model for.
-*  <b>`monitor`</b>: monitor object to print training progress and invoke
-           early stopping.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
 
 ##### Returns:
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.TensorFlowClassifier.md

@@ -15,7 +15,14 @@ Returns weights of the linear classifier.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowClassifier.fit(X, y, monitor=None, logdir=None)` {#TensorFlowClassifier.fit}
+#### `tf.contrib.learn.TensorFlowClassifier.evaluate(x=None, y=None, input_fn=None, steps=None)` {#TensorFlowClassifier.evaluate}
+
+See base class.
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowClassifier.fit(x, y, steps=None, monitors=None, logdir=None)` {#TensorFlowClassifier.fit}
 
 Builds a neural network model given provided `model_fn` and training
 data X and y.
@@ -29,7 +36,7 @@ To restart learning, create new estimator.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -37,8 +44,10 @@ To restart learning, create new estimator.
   iterator that returns array of targets. The training target values
   (class labels in classification, real numbers in regression).
 
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke early
-    stopping
+*  <b>`steps`</b>: int, number of steps to train.
+         If None or 0, train for `self.steps`.
+*  <b>`monitors`</b>: List of `BaseMonitor` objects to print training progress and
+    invoke early stopping.
 *  <b>`logdir`</b>: the directory to save the log file that can be used for
   optional visualization.
 
@@ -110,7 +119,14 @@ Returns list of all variable names in this model.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowClassifier.partial_fit(X, y)` {#TensorFlowClassifier.partial_fit}
+#### `tf.contrib.learn.TensorFlowClassifier.model_dir` {#TensorFlowClassifier.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowClassifier.partial_fit(x, y)` {#TensorFlowClassifier.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -125,7 +141,7 @@ to converge, and you want to split up training into subparts.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -140,7 +156,7 @@ to converge, and you want to split up training into subparts.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowClassifier.predict(X, axis=1, batch_size=None)` {#TensorFlowClassifier.predict}
+#### `tf.contrib.learn.TensorFlowClassifier.predict(x, axis=1, batch_size=None)` {#TensorFlowClassifier.predict}
 
 Predict class or regression for X.
 
@@ -151,7 +167,7 @@ returned.
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`axis`</b>: Which axis to argmax for classification.
     By default axis 1 (next after batch) is used.
     Use 2 for sequence predictions.
@@ -168,14 +184,14 @@ returned.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowClassifier.predict_proba(X, batch_size=None)` {#TensorFlowClassifier.predict_proba}
+#### `tf.contrib.learn.TensorFlowClassifier.predict_proba(x, batch_size=None)` {#TensorFlowClassifier.predict_proba}
 
 Predict class probability of the input samples X.
 
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`batch_size`</b>: If test set is too big, use batch size to split
     it into mini batches. By default the batch_size member variable is used.
 
@@ -260,6 +276,26 @@ Returns
 self
 
 
+- - -
+
+#### `tf.contrib.learn.TensorFlowClassifier.train(input_fn, steps, monitors=None)` {#TensorFlowClassifier.train}
+
+Trains a model given input builder function.
+
+##### Args:
+
+
+*  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
+            dict and Tensor.
+*  <b>`steps`</b>: number of steps to train model for.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
+
+##### Returns:
+
+  Returns self.
+
+
 - - -
 
 #### `tf.contrib.learn.TensorFlowClassifier.weights_` {#TensorFlowClassifier.weights_}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.TensorFlowDNNClassifier.md

@@ -38,7 +38,14 @@ Returns bias of the DNN's bias layers.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowDNNClassifier.fit(X, y, monitor=None, logdir=None)` {#TensorFlowDNNClassifier.fit}
+#### `tf.contrib.learn.TensorFlowDNNClassifier.evaluate(x=None, y=None, input_fn=None, steps=None)` {#TensorFlowDNNClassifier.evaluate}
+
+See base class.
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowDNNClassifier.fit(x, y, steps=None, monitors=None, logdir=None)` {#TensorFlowDNNClassifier.fit}
 
 Builds a neural network model given provided `model_fn` and training
 data X and y.
@@ -52,7 +59,7 @@ To restart learning, create new estimator.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -60,8 +67,10 @@ To restart learning, create new estimator.
   iterator that returns array of targets. The training target values
   (class labels in classification, real numbers in regression).
 
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke early
-    stopping
+*  <b>`steps`</b>: int, number of steps to train.
+         If None or 0, train for `self.steps`.
+*  <b>`monitors`</b>: List of `BaseMonitor` objects to print training progress and
+    invoke early stopping.
 *  <b>`logdir`</b>: the directory to save the log file that can be used for
   optional visualization.
 
@@ -133,7 +142,14 @@ Returns list of all variable names in this model.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowDNNClassifier.partial_fit(X, y)` {#TensorFlowDNNClassifier.partial_fit}
+#### `tf.contrib.learn.TensorFlowDNNClassifier.model_dir` {#TensorFlowDNNClassifier.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowDNNClassifier.partial_fit(x, y)` {#TensorFlowDNNClassifier.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -148,7 +164,7 @@ to converge, and you want to split up training into subparts.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -163,7 +179,7 @@ to converge, and you want to split up training into subparts.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowDNNClassifier.predict(X, axis=1, batch_size=None)` {#TensorFlowDNNClassifier.predict}
+#### `tf.contrib.learn.TensorFlowDNNClassifier.predict(x, axis=1, batch_size=None)` {#TensorFlowDNNClassifier.predict}
 
 Predict class or regression for X.
 
@@ -174,7 +190,7 @@ returned.
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`axis`</b>: Which axis to argmax for classification.
     By default axis 1 (next after batch) is used.
     Use 2 for sequence predictions.
@@ -191,14 +207,14 @@ returned.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowDNNClassifier.predict_proba(X, batch_size=None)` {#TensorFlowDNNClassifier.predict_proba}
+#### `tf.contrib.learn.TensorFlowDNNClassifier.predict_proba(x, batch_size=None)` {#TensorFlowDNNClassifier.predict_proba}
 
 Predict class probability of the input samples X.
 
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`batch_size`</b>: If test set is too big, use batch size to split
     it into mini batches. By default the batch_size member variable is used.
 
@@ -283,6 +299,26 @@ Returns
 self
 
 
+- - -
+
+#### `tf.contrib.learn.TensorFlowDNNClassifier.train(input_fn, steps, monitors=None)` {#TensorFlowDNNClassifier.train}
+
+Trains a model given input builder function.
+
+##### Args:
+
+
+*  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
+            dict and Tensor.
+*  <b>`steps`</b>: number of steps to train model for.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
+
+##### Returns:
+
+  Returns self.
+
+
 - - -
 
 #### `tf.contrib.learn.TensorFlowDNNClassifier.weights_` {#TensorFlowDNNClassifier.weights_}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.TensorFlowDNNRegressor.md

@@ -38,7 +38,14 @@ Returns bias of the DNN's bias layers.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowDNNRegressor.fit(X, y, monitor=None, logdir=None)` {#TensorFlowDNNRegressor.fit}
+#### `tf.contrib.learn.TensorFlowDNNRegressor.evaluate(x=None, y=None, input_fn=None, steps=None)` {#TensorFlowDNNRegressor.evaluate}
+
+See base class.
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowDNNRegressor.fit(x, y, steps=None, monitors=None, logdir=None)` {#TensorFlowDNNRegressor.fit}
 
 Builds a neural network model given provided `model_fn` and training
 data X and y.
@@ -52,7 +59,7 @@ To restart learning, create new estimator.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -60,8 +67,10 @@ To restart learning, create new estimator.
   iterator that returns array of targets. The training target values
   (class labels in classification, real numbers in regression).
 
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke early
-    stopping
+*  <b>`steps`</b>: int, number of steps to train.
+         If None or 0, train for `self.steps`.
+*  <b>`monitors`</b>: List of `BaseMonitor` objects to print training progress and
+    invoke early stopping.
 *  <b>`logdir`</b>: the directory to save the log file that can be used for
   optional visualization.
 
@@ -133,7 +142,14 @@ Returns list of all variable names in this model.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowDNNRegressor.partial_fit(X, y)` {#TensorFlowDNNRegressor.partial_fit}
+#### `tf.contrib.learn.TensorFlowDNNRegressor.model_dir` {#TensorFlowDNNRegressor.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowDNNRegressor.partial_fit(x, y)` {#TensorFlowDNNRegressor.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -148,7 +164,7 @@ to converge, and you want to split up training into subparts.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -163,7 +179,7 @@ to converge, and you want to split up training into subparts.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowDNNRegressor.predict(X, axis=1, batch_size=None)` {#TensorFlowDNNRegressor.predict}
+#### `tf.contrib.learn.TensorFlowDNNRegressor.predict(x, axis=1, batch_size=None)` {#TensorFlowDNNRegressor.predict}
 
 Predict class or regression for X.
 
@@ -174,7 +190,7 @@ returned.
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`axis`</b>: Which axis to argmax for classification.
     By default axis 1 (next after batch) is used.
     Use 2 for sequence predictions.
@@ -191,14 +207,14 @@ returned.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowDNNRegressor.predict_proba(X, batch_size=None)` {#TensorFlowDNNRegressor.predict_proba}
+#### `tf.contrib.learn.TensorFlowDNNRegressor.predict_proba(x, batch_size=None)` {#TensorFlowDNNRegressor.predict_proba}
 
 Predict class probability of the input samples X.
 
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`batch_size`</b>: If test set is too big, use batch size to split
     it into mini batches. By default the batch_size member variable is used.
 
@@ -287,6 +303,26 @@ Returns
 self
 
 
+- - -
+
+#### `tf.contrib.learn.TensorFlowDNNRegressor.train(input_fn, steps, monitors=None)` {#TensorFlowDNNRegressor.train}
+
+Trains a model given input builder function.
+
+##### Args:
+
+
+*  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
+            dict and Tensor.
+*  <b>`steps`</b>: number of steps to train model for.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
+
+##### Returns:
+
+  Returns self.
+
+
 - - -
 
 #### `tf.contrib.learn.TensorFlowDNNRegressor.weights_` {#TensorFlowDNNRegressor.weights_}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.TensorFlowEstimator.md

@@ -38,7 +38,14 @@ Parameters:
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowEstimator.fit(X, y, monitor=None, logdir=None)` {#TensorFlowEstimator.fit}
+#### `tf.contrib.learn.TensorFlowEstimator.evaluate(x=None, y=None, input_fn=None, steps=None)` {#TensorFlowEstimator.evaluate}
+
+See base class.
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowEstimator.fit(x, y, steps=None, monitors=None, logdir=None)` {#TensorFlowEstimator.fit}
 
 Builds a neural network model given provided `model_fn` and training
 data X and y.
@@ -52,7 +59,7 @@ To restart learning, create new estimator.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -60,8 +67,10 @@ To restart learning, create new estimator.
   iterator that returns array of targets. The training target values
   (class labels in classification, real numbers in regression).
 
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke early
-    stopping
+*  <b>`steps`</b>: int, number of steps to train.
+         If None or 0, train for `self.steps`.
+*  <b>`monitors`</b>: List of `BaseMonitor` objects to print training progress and
+    invoke early stopping.
 *  <b>`logdir`</b>: the directory to save the log file that can be used for
   optional visualization.
 
@@ -133,7 +142,14 @@ Returns list of all variable names in this model.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowEstimator.partial_fit(X, y)` {#TensorFlowEstimator.partial_fit}
+#### `tf.contrib.learn.TensorFlowEstimator.model_dir` {#TensorFlowEstimator.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowEstimator.partial_fit(x, y)` {#TensorFlowEstimator.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -148,7 +164,7 @@ to converge, and you want to split up training into subparts.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -163,7 +179,7 @@ to converge, and you want to split up training into subparts.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowEstimator.predict(X, axis=1, batch_size=None)` {#TensorFlowEstimator.predict}
+#### `tf.contrib.learn.TensorFlowEstimator.predict(x, axis=1, batch_size=None)` {#TensorFlowEstimator.predict}
 
 Predict class or regression for X.
 
@@ -174,7 +190,7 @@ returned.
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`axis`</b>: Which axis to argmax for classification.
     By default axis 1 (next after batch) is used.
     Use 2 for sequence predictions.
@@ -191,14 +207,14 @@ returned.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowEstimator.predict_proba(X, batch_size=None)` {#TensorFlowEstimator.predict_proba}
+#### `tf.contrib.learn.TensorFlowEstimator.predict_proba(x, batch_size=None)` {#TensorFlowEstimator.predict_proba}
 
 Predict class probability of the input samples X.
 
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`batch_size`</b>: If test set is too big, use batch size to split
     it into mini batches. By default the batch_size member variable is used.
 
@@ -256,3 +272,23 @@ Returns
 self
 
 
+- - -
+
+#### `tf.contrib.learn.TensorFlowEstimator.train(input_fn, steps, monitors=None)` {#TensorFlowEstimator.train}
+
+Trains a model given input builder function.
+
+##### Args:
+
+
+*  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
+            dict and Tensor.
+*  <b>`steps`</b>: number of steps to train model for.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
+
+##### Returns:
+
+  Returns self.
+
+

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.TensorFlowLinearClassifier.md

@@ -15,7 +15,14 @@ Returns weights of the linear classifier.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowLinearClassifier.fit(X, y, monitor=None, logdir=None)` {#TensorFlowLinearClassifier.fit}
+#### `tf.contrib.learn.TensorFlowLinearClassifier.evaluate(x=None, y=None, input_fn=None, steps=None)` {#TensorFlowLinearClassifier.evaluate}
+
+See base class.
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowLinearClassifier.fit(x, y, steps=None, monitors=None, logdir=None)` {#TensorFlowLinearClassifier.fit}
 
 Builds a neural network model given provided `model_fn` and training
 data X and y.
@@ -29,7 +36,7 @@ To restart learning, create new estimator.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -37,8 +44,10 @@ To restart learning, create new estimator.
   iterator that returns array of targets. The training target values
   (class labels in classification, real numbers in regression).
 
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke early
-    stopping
+*  <b>`steps`</b>: int, number of steps to train.
+         If None or 0, train for `self.steps`.
+*  <b>`monitors`</b>: List of `BaseMonitor` objects to print training progress and
+    invoke early stopping.
 *  <b>`logdir`</b>: the directory to save the log file that can be used for
   optional visualization.
 
@@ -110,7 +119,14 @@ Returns list of all variable names in this model.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowLinearClassifier.partial_fit(X, y)` {#TensorFlowLinearClassifier.partial_fit}
+#### `tf.contrib.learn.TensorFlowLinearClassifier.model_dir` {#TensorFlowLinearClassifier.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowLinearClassifier.partial_fit(x, y)` {#TensorFlowLinearClassifier.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -125,7 +141,7 @@ to converge, and you want to split up training into subparts.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -140,7 +156,7 @@ to converge, and you want to split up training into subparts.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowLinearClassifier.predict(X, axis=1, batch_size=None)` {#TensorFlowLinearClassifier.predict}
+#### `tf.contrib.learn.TensorFlowLinearClassifier.predict(x, axis=1, batch_size=None)` {#TensorFlowLinearClassifier.predict}
 
 Predict class or regression for X.
 
@@ -151,7 +167,7 @@ returned.
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`axis`</b>: Which axis to argmax for classification.
     By default axis 1 (next after batch) is used.
     Use 2 for sequence predictions.
@@ -168,14 +184,14 @@ returned.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowLinearClassifier.predict_proba(X, batch_size=None)` {#TensorFlowLinearClassifier.predict_proba}
+#### `tf.contrib.learn.TensorFlowLinearClassifier.predict_proba(x, batch_size=None)` {#TensorFlowLinearClassifier.predict_proba}
 
 Predict class probability of the input samples X.
 
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`batch_size`</b>: If test set is too big, use batch size to split
     it into mini batches. By default the batch_size member variable is used.
 
@@ -260,6 +276,26 @@ Returns
 self
 
 
+- - -
+
+#### `tf.contrib.learn.TensorFlowLinearClassifier.train(input_fn, steps, monitors=None)` {#TensorFlowLinearClassifier.train}
+
+Trains a model given input builder function.
+
+##### Args:
+
+
+*  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
+            dict and Tensor.
+*  <b>`steps`</b>: number of steps to train model for.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
+
+##### Returns:
+
+  Returns self.
+
+
 - - -
 
 #### `tf.contrib.learn.TensorFlowLinearClassifier.weights_` {#TensorFlowLinearClassifier.weights_}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.TensorFlowLinearRegressor.md

@@ -15,7 +15,14 @@ Returns bias of the linear regression.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowLinearRegressor.fit(X, y, monitor=None, logdir=None)` {#TensorFlowLinearRegressor.fit}
+#### `tf.contrib.learn.TensorFlowLinearRegressor.evaluate(x=None, y=None, input_fn=None, steps=None)` {#TensorFlowLinearRegressor.evaluate}
+
+See base class.
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowLinearRegressor.fit(x, y, steps=None, monitors=None, logdir=None)` {#TensorFlowLinearRegressor.fit}
 
 Builds a neural network model given provided `model_fn` and training
 data X and y.
@@ -29,7 +36,7 @@ To restart learning, create new estimator.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -37,8 +44,10 @@ To restart learning, create new estimator.
   iterator that returns array of targets. The training target values
   (class labels in classification, real numbers in regression).
 
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke early
-    stopping
+*  <b>`steps`</b>: int, number of steps to train.
+         If None or 0, train for `self.steps`.
+*  <b>`monitors`</b>: List of `BaseMonitor` objects to print training progress and
+    invoke early stopping.
 *  <b>`logdir`</b>: the directory to save the log file that can be used for
   optional visualization.
 
@@ -110,7 +119,14 @@ Returns list of all variable names in this model.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowLinearRegressor.partial_fit(X, y)` {#TensorFlowLinearRegressor.partial_fit}
+#### `tf.contrib.learn.TensorFlowLinearRegressor.model_dir` {#TensorFlowLinearRegressor.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowLinearRegressor.partial_fit(x, y)` {#TensorFlowLinearRegressor.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -125,7 +141,7 @@ to converge, and you want to split up training into subparts.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -140,7 +156,7 @@ to converge, and you want to split up training into subparts.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowLinearRegressor.predict(X, axis=1, batch_size=None)` {#TensorFlowLinearRegressor.predict}
+#### `tf.contrib.learn.TensorFlowLinearRegressor.predict(x, axis=1, batch_size=None)` {#TensorFlowLinearRegressor.predict}
 
 Predict class or regression for X.
 
@@ -151,7 +167,7 @@ returned.
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`axis`</b>: Which axis to argmax for classification.
     By default axis 1 (next after batch) is used.
     Use 2 for sequence predictions.
@@ -168,14 +184,14 @@ returned.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowLinearRegressor.predict_proba(X, batch_size=None)` {#TensorFlowLinearRegressor.predict_proba}
+#### `tf.contrib.learn.TensorFlowLinearRegressor.predict_proba(x, batch_size=None)` {#TensorFlowLinearRegressor.predict_proba}
 
 Predict class probability of the input samples X.
 
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`batch_size`</b>: If test set is too big, use batch size to split
     it into mini batches. By default the batch_size member variable is used.
 
@@ -264,6 +280,26 @@ Returns
 self
 
 
+- - -
+
+#### `tf.contrib.learn.TensorFlowLinearRegressor.train(input_fn, steps, monitors=None)` {#TensorFlowLinearRegressor.train}
+
+Trains a model given input builder function.
+
+##### Args:
+
+
+*  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
+            dict and Tensor.
+*  <b>`steps`</b>: number of steps to train model for.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
+
+##### Returns:
+
+  Returns self.
+
+
 - - -
 
 #### `tf.contrib.learn.TensorFlowLinearRegressor.weights_` {#TensorFlowLinearRegressor.weights_}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.TensorFlowRNNClassifier.md

@@ -48,7 +48,14 @@ Returns bias of the rnn layer.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRNNClassifier.fit(X, y, monitor=None, logdir=None)` {#TensorFlowRNNClassifier.fit}
+#### `tf.contrib.learn.TensorFlowRNNClassifier.evaluate(x=None, y=None, input_fn=None, steps=None)` {#TensorFlowRNNClassifier.evaluate}
+
+See base class.
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowRNNClassifier.fit(x, y, steps=None, monitors=None, logdir=None)` {#TensorFlowRNNClassifier.fit}
 
 Builds a neural network model given provided `model_fn` and training
 data X and y.
@@ -62,7 +69,7 @@ To restart learning, create new estimator.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -70,8 +77,10 @@ To restart learning, create new estimator.
   iterator that returns array of targets. The training target values
   (class labels in classification, real numbers in regression).
 
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke early
-    stopping
+*  <b>`steps`</b>: int, number of steps to train.
+         If None or 0, train for `self.steps`.
+*  <b>`monitors`</b>: List of `BaseMonitor` objects to print training progress and
+    invoke early stopping.
 *  <b>`logdir`</b>: the directory to save the log file that can be used for
   optional visualization.
 
@@ -143,7 +152,14 @@ Returns list of all variable names in this model.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRNNClassifier.partial_fit(X, y)` {#TensorFlowRNNClassifier.partial_fit}
+#### `tf.contrib.learn.TensorFlowRNNClassifier.model_dir` {#TensorFlowRNNClassifier.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowRNNClassifier.partial_fit(x, y)` {#TensorFlowRNNClassifier.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -158,7 +174,7 @@ to converge, and you want to split up training into subparts.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -173,7 +189,7 @@ to converge, and you want to split up training into subparts.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRNNClassifier.predict(X, axis=1, batch_size=None)` {#TensorFlowRNNClassifier.predict}
+#### `tf.contrib.learn.TensorFlowRNNClassifier.predict(x, axis=1, batch_size=None)` {#TensorFlowRNNClassifier.predict}
 
 Predict class or regression for X.
 
@@ -184,7 +200,7 @@ returned.
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`axis`</b>: Which axis to argmax for classification.
     By default axis 1 (next after batch) is used.
     Use 2 for sequence predictions.
@@ -201,14 +217,14 @@ returned.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRNNClassifier.predict_proba(X, batch_size=None)` {#TensorFlowRNNClassifier.predict_proba}
+#### `tf.contrib.learn.TensorFlowRNNClassifier.predict_proba(x, batch_size=None)` {#TensorFlowRNNClassifier.predict_proba}
 
 Predict class probability of the input samples X.
 
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`batch_size`</b>: If test set is too big, use batch size to split
     it into mini batches. By default the batch_size member variable is used.
 
@@ -293,6 +309,26 @@ Returns
 self
 
 
+- - -
+
+#### `tf.contrib.learn.TensorFlowRNNClassifier.train(input_fn, steps, monitors=None)` {#TensorFlowRNNClassifier.train}
+
+Trains a model given input builder function.
+
+##### Args:
+
+
+*  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
+            dict and Tensor.
+*  <b>`steps`</b>: number of steps to train model for.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
+
+##### Returns:
+
+  Returns self.
+
+
 - - -
 
 #### `tf.contrib.learn.TensorFlowRNNClassifier.weights_` {#TensorFlowRNNClassifier.weights_}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.TensorFlowRNNRegressor.md

@@ -48,7 +48,14 @@ Returns bias of the rnn layer.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRNNRegressor.fit(X, y, monitor=None, logdir=None)` {#TensorFlowRNNRegressor.fit}
+#### `tf.contrib.learn.TensorFlowRNNRegressor.evaluate(x=None, y=None, input_fn=None, steps=None)` {#TensorFlowRNNRegressor.evaluate}
+
+See base class.
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowRNNRegressor.fit(x, y, steps=None, monitors=None, logdir=None)` {#TensorFlowRNNRegressor.fit}
 
 Builds a neural network model given provided `model_fn` and training
 data X and y.
@@ -62,7 +69,7 @@ To restart learning, create new estimator.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -70,8 +77,10 @@ To restart learning, create new estimator.
   iterator that returns array of targets. The training target values
   (class labels in classification, real numbers in regression).
 
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke early
-    stopping
+*  <b>`steps`</b>: int, number of steps to train.
+         If None or 0, train for `self.steps`.
+*  <b>`monitors`</b>: List of `BaseMonitor` objects to print training progress and
+    invoke early stopping.
 *  <b>`logdir`</b>: the directory to save the log file that can be used for
   optional visualization.
 
@@ -143,7 +152,14 @@ Returns list of all variable names in this model.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRNNRegressor.partial_fit(X, y)` {#TensorFlowRNNRegressor.partial_fit}
+#### `tf.contrib.learn.TensorFlowRNNRegressor.model_dir` {#TensorFlowRNNRegressor.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowRNNRegressor.partial_fit(x, y)` {#TensorFlowRNNRegressor.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -158,7 +174,7 @@ to converge, and you want to split up training into subparts.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -173,7 +189,7 @@ to converge, and you want to split up training into subparts.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRNNRegressor.predict(X, axis=1, batch_size=None)` {#TensorFlowRNNRegressor.predict}
+#### `tf.contrib.learn.TensorFlowRNNRegressor.predict(x, axis=1, batch_size=None)` {#TensorFlowRNNRegressor.predict}
 
 Predict class or regression for X.
 
@@ -184,7 +200,7 @@ returned.
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`axis`</b>: Which axis to argmax for classification.
     By default axis 1 (next after batch) is used.
     Use 2 for sequence predictions.
@@ -201,14 +217,14 @@ returned.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRNNRegressor.predict_proba(X, batch_size=None)` {#TensorFlowRNNRegressor.predict_proba}
+#### `tf.contrib.learn.TensorFlowRNNRegressor.predict_proba(x, batch_size=None)` {#TensorFlowRNNRegressor.predict_proba}
 
 Predict class probability of the input samples X.
 
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`batch_size`</b>: If test set is too big, use batch size to split
     it into mini batches. By default the batch_size member variable is used.
 
@@ -297,6 +313,26 @@ Returns
 self
 
 
+- - -
+
+#### `tf.contrib.learn.TensorFlowRNNRegressor.train(input_fn, steps, monitors=None)` {#TensorFlowRNNRegressor.train}
+
+Trains a model given input builder function.
+
+##### Args:
+
+
+*  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
+            dict and Tensor.
+*  <b>`steps`</b>: number of steps to train model for.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
+
+##### Returns:
+
+  Returns self.
+
+
 - - -
 
 #### `tf.contrib.learn.TensorFlowRNNRegressor.weights_` {#TensorFlowRNNRegressor.weights_}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.TensorFlowRegressor.md

@@ -15,7 +15,14 @@ Returns bias of the linear regression.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRegressor.fit(X, y, monitor=None, logdir=None)` {#TensorFlowRegressor.fit}
+#### `tf.contrib.learn.TensorFlowRegressor.evaluate(x=None, y=None, input_fn=None, steps=None)` {#TensorFlowRegressor.evaluate}
+
+See base class.
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowRegressor.fit(x, y, steps=None, monitors=None, logdir=None)` {#TensorFlowRegressor.fit}
 
 Builds a neural network model given provided `model_fn` and training
 data X and y.
@@ -29,7 +36,7 @@ To restart learning, create new estimator.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -37,8 +44,10 @@ To restart learning, create new estimator.
   iterator that returns array of targets. The training target values
   (class labels in classification, real numbers in regression).
 
-*  <b>`monitor`</b>: Monitor object to print training progress and invoke early
-    stopping
+*  <b>`steps`</b>: int, number of steps to train.
+         If None or 0, train for `self.steps`.
+*  <b>`monitors`</b>: List of `BaseMonitor` objects to print training progress and
+    invoke early stopping.
 *  <b>`logdir`</b>: the directory to save the log file that can be used for
   optional visualization.
 
@@ -110,7 +119,14 @@ Returns list of all variable names in this model.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRegressor.partial_fit(X, y)` {#TensorFlowRegressor.partial_fit}
+#### `tf.contrib.learn.TensorFlowRegressor.model_dir` {#TensorFlowRegressor.model_dir}
+
+
+
+
+- - -
+
+#### `tf.contrib.learn.TensorFlowRegressor.partial_fit(x, y)` {#TensorFlowRegressor.partial_fit}
 
 Incremental fit on a batch of samples.
 
@@ -125,7 +141,7 @@ to converge, and you want to split up training into subparts.
 ##### Args:
 
 
-*  <b>`X`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
+*  <b>`x`</b>: matrix or tensor of shape [n_samples, n_features...]. Can be
   iterator that returns arrays of features. The training input
   samples for fitting the model.
 
@@ -140,7 +156,7 @@ to converge, and you want to split up training into subparts.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRegressor.predict(X, axis=1, batch_size=None)` {#TensorFlowRegressor.predict}
+#### `tf.contrib.learn.TensorFlowRegressor.predict(x, axis=1, batch_size=None)` {#TensorFlowRegressor.predict}
 
 Predict class or regression for X.
 
@@ -151,7 +167,7 @@ returned.
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`axis`</b>: Which axis to argmax for classification.
     By default axis 1 (next after batch) is used.
     Use 2 for sequence predictions.
@@ -168,14 +184,14 @@ returned.
 
 - - -
 
-#### `tf.contrib.learn.TensorFlowRegressor.predict_proba(X, batch_size=None)` {#TensorFlowRegressor.predict_proba}
+#### `tf.contrib.learn.TensorFlowRegressor.predict_proba(x, batch_size=None)` {#TensorFlowRegressor.predict_proba}
 
 Predict class probability of the input samples X.
 
 ##### Args:
 
 
-*  <b>`X`</b>: array-like matrix, [n_samples, n_features...] or iterator.
+*  <b>`x`</b>: array-like matrix, [n_samples, n_features...] or iterator.
 *  <b>`batch_size`</b>: If test set is too big, use batch size to split
     it into mini batches. By default the batch_size member variable is used.
 
@@ -264,6 +280,26 @@ Returns
 self
 
 
+- - -
+
+#### `tf.contrib.learn.TensorFlowRegressor.train(input_fn, steps, monitors=None)` {#TensorFlowRegressor.train}
+
+Trains a model given input builder function.
+
+##### Args:
+
+
+*  <b>`input_fn`</b>: Input builder function, returns tuple of dicts or
+            dict and Tensor.
+*  <b>`steps`</b>: number of steps to train model for.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+            inside the training loop.
+
+##### Returns:
+
+  Returns self.
+
+
 - - -
 
 #### `tf.contrib.learn.TensorFlowRegressor.weights_` {#TensorFlowRegressor.weights_}

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.learn.train.md

@@ -1,4 +1,4 @@
-### `tf.contrib.learn.train(graph, output_dir, train_op, loss_op, global_step_tensor=None, init_op=None, init_fn=None, log_every_steps=10, supervisor_is_chief=True, supervisor_master='', supervisor_save_model_secs=600, supervisor_save_summaries_secs=10, feed_fn=None, max_steps=None, fail_on_nan_loss=True)` {#train}
+### `tf.contrib.learn.train(graph, output_dir, train_op, loss_op, global_step_tensor=None, init_op=None, init_feed_dict=None, init_fn=None, log_every_steps=10, supervisor_is_chief=True, supervisor_master='', supervisor_save_model_secs=600, supervisor_save_summaries_steps=100, feed_fn=None, max_steps=None, fail_on_nan_loss=True, monitors=None)` {#train}
 
 Train a model.
 
@@ -27,6 +27,8 @@ program is terminated with exit code 1.
     one is extracted from the graph using the same logic as in `Supervisor`.
 *  <b>`init_op`</b>: An op that initializes the graph. If `None`, use `Supervisor`'s
     default.
+*  <b>`init_feed_dict`</b>: A dictionary that maps `Tensor` objects to feed values.
+    This feed dictionary will be used when `init_op` is evaluated.
 *  <b>`init_fn`</b>: Optional callable passed to Supervisor to initialize the model.
 *  <b>`log_every_steps`</b>: Output logs regularly. The logs contain timing data and the
     current loss.
@@ -35,13 +37,15 @@ program is terminated with exit code 1.
 *  <b>`supervisor_master`</b>: The master string to use when preparing the session.
 *  <b>`supervisor_save_model_secs`</b>: Save a checkpoint every
     `supervisor_save_model_secs` seconds when training.
-*  <b>`supervisor_save_summaries_secs`</b>: Save summaries every
-    `supervisor_save_summaries_secs` seconds when training.
+*  <b>`supervisor_save_summaries_steps`</b>: Save summaries every
+    `supervisor_save_summaries_steps` seconds when training.
 *  <b>`feed_fn`</b>: A function that is called every iteration to produce a `feed_dict`
     passed to `session.run` calls. Optional.
 *  <b>`max_steps`</b>: Train until `global_step_tensor` evaluates to this value.
 *  <b>`fail_on_nan_loss`</b>: If true, raise `NanLossDuringTrainingError` if `loss_op`
     evaluates to `NaN`. If false, continue training as if nothing happened.
+*  <b>`monitors`</b>: List of `BaseMonitor` subclass instances. Used for callbacks
+    inside the training loop.
 
 ##### Returns:
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.metrics.streaming_auc.md

@@ -17,12 +17,12 @@ numbers of thresholds more closely approximating the true AUC.
 
 To faciliate the estimation of the AUC over a stream of data, the function
 creates an `update_op` operation whose behavior is dependent on the value of
-`weights`. If `weights` is None, then `update_op` increments the
+`ignore_mask`. If `ignore_mask` is None, then `update_op` increments the
 `true_positives`, `true_negatives`, `false_positives` and `false_negatives`
 counts with the number of each found in the current `predictions` and `labels`
-`Tensors`. If `weights` is not `None`, then the increment is performed using
-only the elements of `predictions` and `labels` whose corresponding value
-in `ignore_mask` is `False`. In addition to performing the updates,
+`Tensors`. If `ignore_mask` is not `None`, then the increment is performed
+using only the elements of `predictions` and `labels` whose corresponding
+value in `ignore_mask` is `False`. In addition to performing the updates,
 `update_op` also returns the `auc`.
 
 ##### Args:
@@ -43,7 +43,7 @@ in `ignore_mask` is `False`. In addition to performing the updates,
 ##### Returns:
 
 
-*  <b>`auc`</b>: A tensor representing the current area-under-curve.
+*  <b>`auc`</b>: A scalar tensor representing the current area-under-curve.
 *  <b>`update_op`</b>: An operation that increments the `true_positives`,
     `true_negatives`, `false_positives` and `false_negatives` variables
     appropriately and whose value matches `auc`.
@@ -52,7 +52,7 @@ in `ignore_mask` is `False`. In addition to performing the updates,
 
 
 *  <b>`ValueError`</b>: If the shape of `predictions` and `labels` do not match or if
-    `weights` is not `None` and its shape doesn't match `values`
-    or if either `metrics_collections` or `updates_collections` are not a list
-    or tuple.
+    `ignore_mask` is not `None` and its shape doesn't match `predictions` or
+    if either `metrics_collections` or `updates_collections` are not a list or
+    tuple.
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.metrics.streaming_precision.md

@@ -44,7 +44,7 @@ values in `ignore_mask` are `False`. In addition to performing the updates,
 
 
 *  <b>`ValueError`</b>: If the dimensions of `predictions` and `labels` don't match or
-    if `weight` is not `None` and its shape doesn't match `predictions` or
-    if either `metrics_collections` or `updates_collections` are not
-    a list or tuple.
+    if `ignore_mask` is not `None` and its shape doesn't match `predictions`
+    or if either `metrics_collections` or `updates_collections` are not a list
+    or tuple.
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.metrics.streaming_recall.md

@@ -44,7 +44,7 @@ values in `ignore_mask` are `False`. In addition to performing the updates,
 
 
 *  <b>`ValueError`</b>: If the dimensions of `predictions` and `labels` don't match or
-    if `weight` is not `None` and its shape doesn't match `predictions` or
-    if either `metrics_collections` or `updates_collections` are not
-    a list or tuple.
+    if `ignore_mask` is not `None` and its shape doesn't match `predictions`
+    or if either `metrics_collections` or `updates_collections` are not a list
+    or tuple.
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.contrib.metrics.streaming_recall_at_k.md

@@ -46,7 +46,7 @@ recall value.
 
 
 *  <b>`ValueError`</b>: If the dimensions of `predictions` and `labels` don't match or
-    if `weight` is not `None` and its shape doesn't match `predictions` or if
-    either `metrics_collections` or `updates_collections` are not a list or
-    tuple.
+    if `ignore_mask` is not `None` and its shape doesn't match `predictions`
+    or if either `metrics_collections` or `updates_collections` are not a list
+    or tuple.
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.make_template.md

@@ -1,4 +1,4 @@
-### `tf.make_template(name_, func_, **kwargs)` {#make_template}
+### `tf.make_template(name_, func_, create_scope_now_=False, **kwargs)` {#make_template}
 
 Given an arbitrary function, wrap it so that it does variable sharing.
 
@@ -68,10 +68,14 @@ with tf.variable_scope(vs, reuse=True):
   w2 = scale_by_y2(input2)
 ```
 
-Note: The full variable scope is captured at the time of the first call.
+Depending on the value of `create_scope_now_`, the full variable scope may be
+captured either at the time of first call or at the time of construction. If
+this option is set to True, then all Tensors created by repeated calls to the
+template will have an extra trailing _N+1 to their name, as the first time the
+scope is entered in the Template constructor no Tensors are created.
 
-Note: `name_` and `func_` have a following underscore to reduce the likelihood
-of collisions with kwargs.
+Note: `name_`, `func_` and `create_scope_now_` have a trailing underscore to
+reduce the likelihood of collisions with kwargs.
 
 ##### Args:
 
@@ -79,14 +83,20 @@ of collisions with kwargs.
 *  <b>`name_`</b>: A name for the scope created by this template. If necessary, the name
     will be made unique by appending `_N` to the name.
 *  <b>`func_`</b>: The function to wrap.
+*  <b>`create_scope_now_`</b>: Boolean controlling whether the scope should be created
+    when the template is constructed or when the template is called. Default
+    is False, meaning the scope is created when the template is called.
 *  <b>`**kwargs`</b>: Keyword arguments to apply to `func_`.
 
 ##### Returns:
 
-  A function that will enter a `variable_scope` before calling `func_`. The
-  first time it is called, it will create a non-reusing scope so that the
-  variables will be unique.  On each subsequent call, it will reuse those
-  variables.
+  A function to encapsulate a set of variables which should be created once
+  and reused. An enclosing scope will created, either where `make_template`
+  is called, or wherever the result is called, depending on the value of
+  `create_scope_now_`. Regardless of the value, the first time the template
+  is called it will enter the scope with no reuse, and call `func_` to create
+  variables, which are guaranteed to be unique. All subsequent calls will
+  re-enter the scope and reuse those variables.
 
 ##### Raises:
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.merge_all_summaries.md

@@ -11,6 +11,6 @@ Merges all summaries collected in the default graph.
 ##### Returns:
 
   If no summaries were collected, returns None.  Otherwise returns a scalar
-  `Tensor` of type`string` containing the serialized `Summary` protocol
+  `Tensor` of type `string` containing the serialized `Summary` protocol
   buffer resulting from the merging.
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.nn.sufficient_statistics.md

@@ -1,4 +1,4 @@
-### `tf.nn.sufficient_statistics(x, axes, shift=True, keep_dims=False, name=None)` {#sufficient_statistics}
+### `tf.nn.sufficient_statistics(x, axes, shift=False, keep_dims=False, name=None)` {#sufficient_statistics}
 
 Calculate the sufficient statistics for the mean and variance of `x`.
 
@@ -6,6 +6,9 @@ These sufficient statistics are computed using the one pass algorithm on
 an input that's optionally shifted using the value of the 1st element in `x`.
 See:
 https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Computing_shifted_data
+Unfortunately, in some cases using a random individual sample as the shift
+value leads experimentally to very poor numerical stability, so it is disabled
+by default. The one-pass approach might have to be revised accordingly.
 
 ##### Args:
 

tensorflow/g3doc/api_docs/python/functions_and_classes/tf.train.write_graph.md

@@ -1,6 +1,6 @@
 ### `tf.train.write_graph(graph_def, logdir, name, as_text=True)` {#write_graph}
 
-Writes a graph proto on disk.
+Writes a graph proto to a file.
 
 The graph is written as a binary proto unless `as_text` is `True`.
 
@@ -14,7 +14,8 @@ tf.train.write_graph(sess.graph_def, '/tmp/my-model', 'train.pbtxt')
 
 
 *  <b>`graph_def`</b>: A `GraphDef` protocol buffer.
-*  <b>`logdir`</b>: Directory where to write the graph.
+*  <b>`logdir`</b>: Directory where to write the graph. This can refer to remote
+    filesystems, such as Google Cloud Storage (GCS).
 *  <b>`name`</b>: Filename for the graph.
 *  <b>`as_text`</b>: If `True`, writes the graph as an ASCII proto.
 

tensorflow/g3doc/api_docs/python/nn.md

@@ -824,7 +824,7 @@ convolutional neural networks (NIPS 2012)]
 
 - - -
 
-### `tf.nn.sufficient_statistics(x, axes, shift=True, keep_dims=False, name=None)` {#sufficient_statistics}
+### `tf.nn.sufficient_statistics(x, axes, shift=False, keep_dims=False, name=None)` {#sufficient_statistics}
 
 Calculate the sufficient statistics for the mean and variance of `x`.
 
@@ -832,6 +832,9 @@ These sufficient statistics are computed using the one pass algorithm on
 an input that's optionally shifted using the value of the 1st element in `x`.
 See:
 https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Computing_shifted_data
+Unfortunately, in some cases using a random individual sample as the shift
+value leads experimentally to very poor numerical stability, so it is disabled
+by default. The one-pass approach might have to be revised accordingly.
 
 ##### Args:
 

tensorflow/g3doc/api_docs/python/state_ops.md

@@ -1395,7 +1395,7 @@ Returns the current variable scope.
 
 - - -
 
-### `tf.make_template(name_, func_, **kwargs)` {#make_template}
+### `tf.make_template(name_, func_, create_scope_now_=False, **kwargs)` {#make_template}
 
 Given an arbitrary function, wrap it so that it does variable sharing.
 
@@ -1465,10 +1465,14 @@ with tf.variable_scope(vs, reuse=True):
   w2 = scale_by_y2(input2)
 ```
 
-Note: The full variable scope is captured at the time of the first call.
+Depending on the value of `create_scope_now_`, the full variable scope may be
+captured either at the time of first call or at the time of construction. If
+this option is set to True, then all Tensors created by repeated calls to the
+template will have an extra trailing _N+1 to their name, as the first time the
+scope is entered in the Template constructor no Tensors are created.
 
-Note: `name_` and `func_` have a following underscore to reduce the likelihood
-of collisions with kwargs.
+Note: `name_`, `func_` and `create_scope_now_` have a trailing underscore to
+reduce the likelihood of collisions with kwargs.
 
 ##### Args:
 
@@ -1476,14 +1480,20 @@ of collisions with kwargs.
 *  <b>`name_`</b>: A name for the scope created by this template. If necessary, the name
     will be made unique by appending `_N` to the name.
 *  <b>`func_`</b>: The function to wrap.
+*  <b>`create_scope_now_`</b>: Boolean controlling whether the scope should be created
+    when the template is constructed or when the template is called. Default
+    is False, meaning the scope is created when the template is called.
 *  <b>`**kwargs`</b>: Keyword arguments to apply to `func_`.
 
 ##### Returns:
 
-  A function that will enter a `variable_scope` before calling `func_`. The
-  first time it is called, it will create a non-reusing scope so that the
-  variables will be unique.  On each subsequent call, it will reuse those
-  variables.
+  A function to encapsulate a set of variables which should be created once
+  and reused. An enclosing scope will created, either where `make_template`
+  is called, or wherever the result is called, depending on the value of
+  `create_scope_now_`. Regardless of the value, the first time the template
+  is called it will enter the scope with no reuse, and call `func_` to create
+  variables, which are guaranteed to be unique. All subsequent calls will
+  re-enter the scope and reuse those variables.
 
 ##### Raises:
 

tensorflow/g3doc/api_docs/python/train.md

@@ -3078,7 +3078,7 @@ Merges all summaries collected in the default graph.
 ##### Returns:
 
   If no summaries were collected, returns None.  Otherwise returns a scalar
-  `Tensor` of type`string` containing the serialized `Summary` protocol
+  `Tensor` of type `string` containing the serialized `Summary` protocol
   buffer resulting from the merging.
 
 
@@ -3346,7 +3346,7 @@ global_step: 10
 
 ### `tf.train.write_graph(graph_def, logdir, name, as_text=True)` {#write_graph}
 
-Writes a graph proto on disk.
+Writes a graph proto to a file.
 
 The graph is written as a binary proto unless `as_text` is `True`.
 
@@ -3360,7 +3360,8 @@ tf.train.write_graph(sess.graph_def, '/tmp/my-model', 'train.pbtxt')
 
 
 *  <b>`graph_def`</b>: A `GraphDef` protocol buffer.
-*  <b>`logdir`</b>: Directory where to write the graph.
+*  <b>`logdir`</b>: Directory where to write the graph. This can refer to remote
+    filesystems, such as Google Cloud Storage (GCS).
 *  <b>`name`</b>: Filename for the graph.
 *  <b>`as_text`</b>: If `True`, writes the graph as an ASCII proto.
 

tensorflow/g3doc/how_tos/style_guide.md

@@ -92,8 +92,8 @@ creates a part of the graph and returns output tensors.
  If operation needs to save some `Tensor`s to Graph collections,
  put the arguments with names of the collections right before `name` argument.
 
-* Tensor arguments should be either a single tensor or an iterable of tensors.
- *Not a "Tensor or list of Tensors".*
+* Tensor arguments should be either a single tensor or an iterable of tensors,
+  not both.  E.g. a "Tensor or list of Tensors" is too broad.
 
 * Operations that take tensors as arguments should call `convert_to_tensor`
  to convert non-tensor inputs into tensors if they are using C++ operations.

tensorflow/python/BUILD

@@ -15,6 +15,7 @@ load("//tensorflow:tensorflow.bzl", "cuda_py_test")
 load("//tensorflow:tensorflow.bzl", "cuda_py_tests")
 load("//tensorflow/core:platform/default/build_config.bzl", "tf_proto_library")
 load("//tensorflow/core:platform/default/build_config.bzl", "tf_proto_library_py")
+load("//tensorflow/core:platform/default/build_config.bzl", "tf_additional_lib_deps")
 
 py_library(
     name = "python",
@@ -983,6 +984,7 @@ tf_py_wrap_cc(
         "lib/io/py_record_writer.i",
         "platform/base.i",
         "platform/numpy.i",
+        "training/saver_io.i",
         "training/server_lib.i",
         "util/port.i",
         "util/py_checkpoint_reader.i",
@@ -994,12 +996,12 @@ tf_py_wrap_cc(
         ":py_record_writer_lib",
         ":python_op_gen",
         ":tf_session_helper",
-        "//tensorflow/core:lib",
-        "//tensorflow/core/distributed_runtime:server_lib",
         "//tensorflow/core/distributed_runtime/rpc:grpc_server_lib",
         "//tensorflow/core/distributed_runtime/rpc:grpc_session",
+        "//tensorflow/core:lib",
+        "//tensorflow/core/distributed_runtime:server_lib",
         "//util/python:python_headers",
-    ],
+    ] + tf_additional_lib_deps(),
 )
 
 py_library(

tensorflow/python/kernel_tests/control_flow_ops_py_test.py

@@ -1256,17 +1256,50 @@ class ControlFlowTest(tf.test.TestCase):
 
       def break_run_twice(ix):
         def _break():
-          assert not ran_once[ix]
           ran_once[ix] = True
           return tf.constant(ix)
         return _break
 
       # Should not fail - each conditional gets called exactly once
+      # except default.  Default gets called twice: once to create an
+      # empty output and once for the actual cond switch.
       r6 = tf.case([(x < y, break_run_twice(0)), (x > y, break_run_twice(1))],
-                   default=break_run_twice(2))
+                   default=lambda: tf.constant(2))
 
       self.assertAllEqual(r6.eval(), 0)
 
+  def testCaseSideEffects(self):
+    with self.test_session() as sess:
+      v0 = tf.Variable(-1)
+      v1 = tf.Variable(-1)
+      v2 = tf.Variable(-1)
+
+      a = lambda: control_flow_ops.with_dependencies([tf.assign(v0, 0)], 0)
+      b = lambda: control_flow_ops.with_dependencies([tf.assign(v1, 1)], 1)
+      c = lambda: control_flow_ops.with_dependencies([tf.assign(v2, 2)], 2)
+
+      x = tf.constant(1)
+      y = tf.constant(2)
+
+      r0 = tf.case(((x < y, a), (x > y, b)), default=c, exclusive=True)
+      r1 = tf.case(((x > y, a), (x < y, b)), default=c, exclusive=True)
+      r2 = tf.case(((x > y, a), (x > y, b)), default=c, exclusive=True)
+
+      tf.initialize_all_variables().run()
+      self.assertAllEqual(sess.run([v0, v1, v2]), [-1] * 3)
+      self.assertEqual(2, r2.eval())
+      self.assertAllEqual(sess.run([v0, v1, v2]), [-1, -1, 2])
+
+      tf.initialize_all_variables().run()
+      self.assertAllEqual(sess.run([v0, v1, v2]), [-1] * 3)
+      self.assertEqual(1, r1.eval())
+      self.assertAllEqual(sess.run([v0, v1, v2]), [-1, 1, -1])
+
+      tf.initialize_all_variables().run()
+      self.assertAllEqual(sess.run([v0, v1, v2]), [-1] * 3)
+      self.assertEqual(0, r0.eval())
+      self.assertAllEqual(sess.run([v0, v1, v2]), [0, -1, -1])
+
   def testOneOpCond(self):
     with self.test_session():
       v = tf.Variable(0)

tensorflow/python/kernel_tests/template_test.py

@@ -206,6 +206,27 @@ class TemplateTest(tf.test.TestCase):
     self.assertEqual("s1/nested_1/dummy:0", v1.name)
     self.assertEqual("s1_2/nested_1/dummy:0", v3.name)
 
+  def test_immediate_scope_creation(self):
+    # Create templates in scope a then call in scope b. make_template should
+    # capture the scope the first time it is called, and make_immediate_template
+    # should capture the scope at construction time.
+    with tf.variable_scope("ctor_scope"):
+      tmpl_immed = template.make_template(
+          "a", var_scoped_function, True)  # create scope here
+      tmpl_defer = template.make_template(
+          "b", var_scoped_function, False)  # default: create scope at __call__
+    with tf.variable_scope("call_scope"):
+      inner_imm_var = tmpl_immed()
+      inner_defer_var = tmpl_defer()
+    outer_imm_var = tmpl_immed()
+    outer_defer_var = tmpl_defer()
+
+    self.assertNotEqual(inner_imm_var, inner_defer_var)
+    self.assertEqual(outer_imm_var, inner_imm_var)
+    self.assertEqual(outer_defer_var, inner_defer_var)
+
+    self.assertEqual("ctor_scope/a/dummy:0", inner_imm_var.name)
+    self.assertEqual("call_scope/b/dummy:0", inner_defer_var.name)
 
 if __name__ == "__main__":
   tf.test.main()

tensorflow/python/lib/core/strings.i

@@ -69,8 +69,8 @@ bool _BytesToStringPiece(PyObject* obj, tensorflow::StringPiece* result) {
   $1 = &temp;
 }
 
-// C++ functions returning tensorflow::StringPiece will simply return bytes in Python,
-// or None if the StringPiece contained a NULL pointer.
+// C++ functions returning tensorflow::StringPiece will simply return bytes in
+// Python, or None if the StringPiece contained a NULL pointer.
 %typemap(out) tensorflow::StringPiece {
   if ($1.data()) {
     $result = PyBytes_FromStringAndSize($1.data(), $1.size());
@@ -79,3 +79,30 @@ bool _BytesToStringPiece(PyObject* obj, tensorflow::StringPiece* result) {
     $result = Py_None;
   }
 }
+
+// Converts a C++ string vector to a Python string list.
+%typemap(out) std::vector<string> {
+  const int size = $1.size();
+  auto temp_string_list = tensorflow::make_safe(PyList_New(size));
+  if (!temp_string_list) {
+    SWIG_fail;
+  }
+  tensorflow::Safe_PyObjectVector converted;
+  converted.reserve(size);
+  for (const string& op : $1) {
+%#if PY_MAJOR_VERSION >= 3
+    PyObject* py_str = PyUnicode_FromStringAndSize(op.data(), op.size());
+%#else
+    PyObject* py_str = PyString_FromStringAndSize(op.data(), op.size());
+%#endif
+    if (!py_str) {
+      SWIG_fail;
+    }
+    converted.emplace_back(tensorflow::make_safe(py_str));
+  }
+  for (int i = 0; i < converted.size(); ++i) {
+    PyList_SET_ITEM(temp_string_list.get(), i, converted[i].release());
+  }
+  $result = temp_string_list.release();
+}
+

tensorflow/python/ops/control_flow_ops.py

@@ -1985,29 +1985,64 @@ def case(pred_fn_pairs, default, exclusive=False, name="case"):
     for i, p in enumerate(preds):
       with ops.name_scope("not_%d" % i):
         not_preds.append(math_ops.logical_not(p))
-    and_not_preds = [constant_op.constant(True, name="and_not_true")]
-    for i, notp in enumerate(not_preds[:-1]):
+    and_not_preds = [constant_op.constant(True, name="always_true")]
+    for i, notp in enumerate(not_preds):
       with ops.name_scope("and_not_%d" % i):
         and_not_preds.append(math_ops.logical_and(and_not_preds[-1], notp))
 
     # preds = [p1, p2, p3]
     # fns = [f1, f2, f3]
     # not_preds = [~p1, ~p2, ~p3]
-    # case_preds = [p1 & True,
+    # and_not_preds = [True, ~p1, ~p1 & ~p2, ~p1 & ~p2 & ~p3]
+    # case_preds = [p1,
     #               p2 & ~p1,
-    #               p3 & ~p1 & ~ p2]
+    #               p3 & ~p2 & ~p1,
+    #              ~p3 & ~p2 & ~p1]
+
     case_preds = []
-    for i, (p, and_not_p_prev) in enumerate(zip(preds, and_not_preds)):
+    for i, (p, and_not_p_prev) in enumerate(zip(preds, and_not_preds[:-1])):
       with ops.name_scope("case_%d" % i):
         case_preds.append(math_ops.logical_and(p, and_not_p_prev))
+    with ops.name_scope("case_none_are_true"):
+      case_preds.append(and_not_preds[-1])
+
+    # Create an empty tensor, or list, with the right type and shape
+    with ops.name_scope("case_create_empty"):
+      dummy_value = default()
+      def _correct_empty(v):
+        if isinstance(v, ops.Operation):
+          return no_op()
+        elif v.dtype == dtypes.string:
+          return array_ops.constant("")
+        else:
+          return array_ops.constant(v.dtype.as_numpy_dtype())
+
+      if isinstance(dummy_value, collections.Sequence):
+        dummy_type = type(dummy_value)
+        empty = lambda: dummy_type(_correct_empty(v) for v in dummy_value)
+      else:
+        empty = lambda: _correct_empty(dummy_value)
+
+    # case_sequence = [
+    #   cond(~p3 & ~p2 & ~p1, default, empty),
+    #   cond(p3 & ~p2 & ~p1, f3, lambda: case_sequence[0]),
+    #   cond(p2 & ~p1, f2, lambda: case_sequence[1]),
+    #   cond(p1, f1, lambda: case_sequence[2])
+    # ]
+    #
+    # And the return value will be case_sequence[-1]
+    def _build_case():
+      all_fns = [fn for fn in fns]
+      all_fns.append(default)
+      prev_case = None
+      for i, (cp, fn) in enumerate(list(zip(case_preds, all_fns))[::-1]):
+        prev_case = cond(
+            cp, fn,
+            empty if i == 0 else lambda: prev_case,
+            name="If_%d" % i)
+      return prev_case
 
-    # case_sequence = [cond(p3 & ..., f3, default),
-    #                  cond(p2 & ..., f2, lambda: case_sequence[0]),
-    #                  ...
-    #                  cond(p1 & True, f1, lambda: case_sequence[i-1])]
-    # and prev_case_seq will loop from case_sequence[0] to case_sequence[-1]
     if exclusive:
-      # TODO(ebrevdo): Add Where() for DT_BOOL, replace with Size(Where(preds))
       preds_c = array_ops.pack(preds, name="preds_c")
       num_true_conditions = math_ops.reduce_sum(
           math_ops.cast(preds_c, dtypes.int32), name="num_true_conds")
@@ -2022,21 +2057,11 @@ def case(pred_fn_pairs, default, exclusive=False, name="case"):
       with ops.control_dependencies([
           logging_ops.Assert(condition=at_most_one_true_condition,
                              data=error_msg, summarize=len(preds))]):
-        prev_case_seq = None
-        for i, (cp, fn) in enumerate(list(zip(case_preds, fns))[::-1]):
-          prev_case_seq = cond(
-              cp, fn,
-              default if i == 0 else lambda: prev_case_seq,
-              name="If_%d" % i)
+        case_seq = _build_case()
     else:
-      prev_case_seq = None
-      for i, (cp, fn) in enumerate(list(zip(case_preds, fns))[::-1]):
-        prev_case_seq = cond(
-            cp, fn,
-            default if i == 0 else lambda: prev_case_seq,
-            name="If_%d" % i)
+      case_seq = _build_case()
 
-    return prev_case_seq
+    return case_seq
 
 
 ops.RegisterShape("Enter")(common_shapes.unchanged_shape)

tensorflow/python/ops/logging_ops.py

@@ -261,7 +261,7 @@ def merge_all_summaries(key=ops.GraphKeys.SUMMARIES):
 
   Returns:
     If no summaries were collected, returns None.  Otherwise returns a scalar
-    `Tensor` of type`string` containing the serialized `Summary` protocol
+    `Tensor` of type `string` containing the serialized `Summary` protocol
     buffer resulting from the merging.
   """
   summary_ops = ops.get_collection(key)
@@ -271,6 +271,30 @@ def merge_all_summaries(key=ops.GraphKeys.SUMMARIES):
     return merge_summary(summary_ops)
 
 
+def get_summary_op():
+  """Returns a single Summary op that would run all summaries.
+
+  Either existing one from `SUMMARY_OP` collection or merges all existing
+  summaries.
+
+  Returns:
+    If no summaries were collected, returns None. Otherwise returns a scalar
+    `Tensor` of type `string` containing the serialized `Summary` protocol
+    buffer resulting from the merging.
+  """
+  summary_op = ops.get_collection(ops.GraphKeys.SUMMARY_OP)
+  if summary_op is not None:
+    if summary_op:
+      summary_op = summary_op[0]
+    else:
+      summary_op = None
+  if summary_op is None:
+    summary_op = merge_all_summaries()
+    if summary_op is not None:
+      ops.add_to_collection(ops.GraphKeys.SUMMARY_OP, summary_op)
+  return summary_op
+
+
 def scalar_summary(tags, values, collections=None, name=None):
   """Outputs a `Summary` protocol buffer with scalar values.
 

tensorflow/python/ops/nn.py

@@ -587,13 +587,16 @@ def separable_conv2d(input, depthwise_filter, pointwise_filter, strides,
                          padding="VALID", name=name)
 
 
-def sufficient_statistics(x, axes, shift=True, keep_dims=False, name=None):
+def sufficient_statistics(x, axes, shift=False, keep_dims=False, name=None):
   """Calculate the sufficient statistics for the mean and variance of `x`.
 
   These sufficient statistics are computed using the one pass algorithm on
   an input that's optionally shifted using the value of the 1st element in `x`.
   See:
   https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Computing_shifted_data
+  Unfortunately, in some cases using a random individual sample as the shift
+  value leads experimentally to very poor numerical stability, so it is disabled
+  by default. The one-pass approach might have to be revised accordingly.
 
   Args:
     x: A `Tensor`.

tensorflow/python/ops/nn_ops.py

@@ -752,6 +752,41 @@ def _calc_conv_weight_params(graph, node):
                                            filter_in_depth * filter_out_depth))
 
 
+@ops.RegisterStatistics("DepthwiseConv2dNative", "flops")
+def _calc_depthwise_conv_flops(graph, node):
+  """Calculates the compute resources needed for DepthwiseConv2dNative."""
+  input_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
+  input_shape.assert_is_fully_defined()
+  filter_shape = graph_util.tensor_shape_from_node_def_name(graph,
+                                                            node.input[1])
+  filter_shape.assert_is_fully_defined()
+  output_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
+  output_shape.assert_is_fully_defined()
+  filter_height = int(filter_shape[0])
+  filter_width = int(filter_shape[1])
+  output_count = np.prod(output_shape.as_list())
+  return ops.OpStats("flops", (output_count * filter_height * filter_width * 2))
+
+
+@ops.RegisterStatistics("DepthwiseConv2dNative", "weight_parameters")
+def _calc_depthwise_conv_weight_params(graph, node):
+  """Calculates the on-disk size of the weights for DepthwiseConv2dNative."""
+  input_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
+  input_shape.assert_is_fully_defined()
+  filter_shape = graph_util.tensor_shape_from_node_def_name(graph,
+                                                            node.input[1])
+  filter_shape.assert_is_fully_defined()
+  output_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
+  output_shape.assert_is_fully_defined()
+  filter_height = int(filter_shape[0])
+  filter_width = int(filter_shape[1])
+  filter_in_depth = int(filter_shape[2])
+  filter_channel_multiplier = int(filter_shape[3])
+  return ops.OpStats("weight_parameters", (filter_height * filter_width *
+                                           filter_in_depth *
+                                           filter_channel_multiplier))
+
+
 @ops.RegisterShape("Conv3D")
 def _Conv3DShape(op):
   """Shape function for Conv3D."""

tensorflow/python/ops/numerics.py

@@ -57,9 +57,9 @@ def add_check_numerics_ops():
   """
   check_op = []
   # This code relies on the ordering of ops in get_operations().
-  # The consumer of a tensor always comes before that tensor's producer in
+  # The producer of a tensor always comes before that tensor's consumer in
   # this list. This is true because get_operations() returns ops in the order
-  # added, and ops can only be added once its inputs are added.
+  # added, and an op can only be added after its inputs are added.
   for op in ops.get_default_graph().get_operations():
     for output in op.outputs:
       if output.dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:

tensorflow/python/ops/template.py

@@ -29,7 +29,7 @@ from tensorflow.python.platform import tf_logging as logging
 __all__ = ["make_template"]
 
 
-def make_template(name_, func_, **kwargs):
+def make_template(name_, func_, create_scope_now_=False, **kwargs):
   """Given an arbitrary function, wrap it so that it does variable sharing.
 
   This wraps `func_` in a Template and partially evaluates it. Templates are
@@ -98,29 +98,39 @@ def make_template(name_, func_, **kwargs):
     w2 = scale_by_y2(input2)
   ```
 
-  Note: The full variable scope is captured at the time of the first call.
+  Depending on the value of `create_scope_now_`, the full variable scope may be
+  captured either at the time of first call or at the time of construction. If
+  this option is set to True, then all Tensors created by repeated calls to the
+  template will have an extra trailing _N+1 to their name, as the first time the
+  scope is entered in the Template constructor no Tensors are created.
 
-  Note: `name_` and `func_` have a following underscore to reduce the likelihood
-  of collisions with kwargs.
+  Note: `name_`, `func_` and `create_scope_now_` have a trailing underscore to
+  reduce the likelihood of collisions with kwargs.
 
   Args:
     name_: A name for the scope created by this template. If necessary, the name
       will be made unique by appending `_N` to the name.
     func_: The function to wrap.
+    create_scope_now_: Boolean controlling whether the scope should be created
+      when the template is constructed or when the template is called. Default
+      is False, meaning the scope is created when the template is called.
     **kwargs: Keyword arguments to apply to `func_`.
 
   Returns:
-    A function that will enter a `variable_scope` before calling `func_`. The
-    first time it is called, it will create a non-reusing scope so that the
-    variables will be unique.  On each subsequent call, it will reuse those
-    variables.
+    A function to encapsulate a set of variables which should be created once
+    and reused. An enclosing scope will created, either where `make_template`
+    is called, or wherever the result is called, depending on the value of
+    `create_scope_now_`. Regardless of the value, the first time the template
+    is called it will enter the scope with no reuse, and call `func_` to create
+    variables, which are guaranteed to be unique. All subsequent calls will
+    re-enter the scope and reuse those variables.
 
   Raises:
     ValueError: if the name is None.
   """
   if kwargs:
     func_ = functools.partial(func_, **kwargs)
-  return Template(name_, func_)
+  return Template(name_, func_, create_scope_now=create_scope_now_)
 
 
 def _skip_common_stack_elements(stacktrace, base_case):
@@ -137,10 +147,13 @@ class Template(object):
   Templates are functions that create variables the first time they are called
   and reuse them thereafter. See `make_template` for full documentation.
 
-  Note: The full variable scope is captured at the time of the first call.
+  Note: By default, the full variable scope is captured at the time of first
+  call. If `create_scope_now_` is passed as True to the constructor, the full
+  scope will be captured there, but no variables will created until the first
+  call.
   """
 
-  def __init__(self, name, func):
+  def __init__(self, name, func, create_scope_now=False):
     """Creates a template for the given function.
 
     Args:
@@ -148,6 +161,15 @@ class Template(object):
         name will be made unique by appending `_N` to the it (see how
         `tf.variable_op_scope` treats the `default_name` for details).
       func: The function to apply each time.
+      create_scope_now: Whether to create the scope at Template construction
+        time, rather than first call. Defaults to false. Creating the scope at
+        construction time may be more convenient if the template is to passed
+        through much lower level code, and you want to be sure of the scope
+        name without knowing exactly where it will be first called. If set to
+        True, the scope will be created in the constructor, and all subsequent
+        times in __call__, leading to a trailing numeral being added to the
+        names of all created Tensors. If set to False, the scope will be created
+        at the first call location.
 
     Raises:
       ValueError: if the name is None.
@@ -157,7 +179,14 @@ class Template(object):
     self._name = name
     if name is None:
       raise ValueError("name cannot be None.")
-    self._var_scope = None
+    if create_scope_now:
+      with variable_scope.variable_op_scope([], None, self._name) as vs:
+        self._var_scope = vs
+    else:
+      self._var_scope = None
+    # This variable keeps track of whether the template has been called yet,
+    # which is not the same as whether the scope has been created.
+    self._variables_created = False
 
   def _call_func(self, args, kwargs, check_for_new_variables):
     try:
@@ -204,12 +233,23 @@ class Template(object):
       raise
 
   def __call__(self, *args, **kwargs):
-    # Capture the name of the variable_scope here because if we capture at
-    # construction, then name_scopes would have a '_N+1' suffix.
     if self._var_scope:
-      with variable_scope.variable_scope(self._var_scope, reuse=True):
-        return self._call_func(args, kwargs, check_for_new_variables=True)
+      if self._variables_created:
+        # This is not the first visit to __call__, so variables have already
+        # been created, and we want to reuse them.
+        with variable_scope.variable_scope(self._var_scope, reuse=True):
+          return self._call_func(args, kwargs, check_for_new_variables=True)
+      else:
+        # This is the first visit to __call__, but the scope has already been
+        # created in the constructor. Set _variables_created so that subsequent
+        # calls take the if branch above.
+        self._variables_created = True
+        with variable_scope.variable_scope(self._var_scope):
+          return self._call_func(args, kwargs, check_for_new_variables=False)
     else:
+      # The scope was not created at construction time, so create it here.
+      # Subsequent calls should reuse variables.
+      self._variables_created = True
       with variable_scope.variable_op_scope([], None, self._name) as vs:
         self._var_scope = vs
         return self._call_func(args, kwargs, check_for_new_variables=False)