Minor eager mode improvements

Mainly doesn't get shape when getting rank will suffice.

Before:
entry {
  name: "EagerLinearRegressionBenchmark.eager_train_cpu"
  iters: 2000
  wall_time: 1.54617094994
  extras {
    key: "examples_per_sec"
    value {
      double_value: 82785.1538701
    }
  }
}

After:
entry {
  name: "EagerLinearRegressionBenchmark.eager_train_cpu"
  iters: 2000
  wall_time: 1.38547611237
  extras {
    key: "examples_per_sec"
    value {
      double_value: 92387.0132856
    }
  }
}
PiperOrigin-RevId: 204001196

tensorflow/python/eager/pywrap_tfe_src.cc

@@ -1898,10 +1898,10 @@ PyObject* RecordGradient(PyObject* op_name, PyObject* inputs, PyObject* attrs,
 
 void MaybeWatchVariable(PyObject* input) {
   DCHECK(CheckResourceVariable(input));
-  DCHECK(PyObject_HasAttrString(input, "trainable"));
+  DCHECK(PyObject_HasAttrString(input, "_trainable"));
 
   tensorflow::Safe_PyObjectPtr trainable(
-      PyObject_GetAttrString(input, "trainable"));
+      PyObject_GetAttrString(input, "_trainable"));
   if (trainable.get() == Py_False) return;
   TFE_Py_TapeSetWatchVariable(input);
 }

tensorflow/python/framework/common_shapes.py

@@ -28,6 +28,18 @@ from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import tensor_util
 
 
+def has_fully_defined_shape(tensor):
+  """Returns true if tensor has a fully defined shape."""
+  return isinstance(tensor, ops.EagerTensor) or tensor.shape.is_fully_defined()
+
+
+def rank(tensor):
+  """Return a rank if it is a tensor, else return None."""
+  if isinstance(tensor, ops.Tensor):
+    return tensor._rank()  # pylint: disable=protected-access
+  return None
+
+
 def scalar_shape(unused_op):
   """Shape function for ops that output a scalar value."""
   return [tensor_shape.scalar()]

tensorflow/python/keras/engine/base_layer.py

@@ -685,8 +685,8 @@ class Layer(checkpointable.CheckpointableBase):
 
     # Handle Keras mask propagation from previous layer to current layer.
     previous_mask = None
-    if (not hasattr(self, '_compute_previous_mask') or
-        self._compute_previous_mask):
+    if build_graph and (not hasattr(self, '_compute_previous_mask') or
+                        self._compute_previous_mask):
       previous_mask = collect_previous_mask(inputs)
       if not hasattr(self, '_call_fn_args'):
         self._call_fn_args = self._no_dependency(
@@ -726,6 +726,7 @@ class Layer(checkpointable.CheckpointableBase):
         if all(hasattr(x, 'shape') for x in input_list):
           input_shapes = nest.map_structure(lambda x: x.shape, inputs)
         self.build(input_shapes)
+        self.built = True
 
       # Check input assumptions set after layer building, e.g. input shape.
       if build_graph or in_deferred_mode:
@@ -761,8 +762,6 @@ class Layer(checkpointable.CheckpointableBase):
       if in_deferred_mode or build_graph and have_all_keras_metadata(inputs):
         inputs, outputs = self._set_connectivity_metadata_(
             inputs, outputs, args, kwargs)
-
-      self.built = True
       if context.executing_eagerly():
         return outputs
 

tensorflow/python/keras/layers/core.py

@@ -26,6 +26,7 @@ import warnings
 import numpy as np
 
 from tensorflow.python.eager import context
+from tensorflow.python.framework import common_shapes
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_shape
 from tensorflow.python.keras import activations
@@ -929,13 +930,13 @@ class Dense(Layer):
 
   def call(self, inputs):
     inputs = ops.convert_to_tensor(inputs, dtype=self.dtype)
-    shape = inputs.get_shape().as_list()
-    if len(shape) > 2:
+    rank = common_shapes.rank(inputs)
+    if rank > 2:
       # Broadcasting is required for the inputs.
-      outputs = standard_ops.tensordot(inputs, self.kernel, [[len(shape) - 1],
-                                                             [0]])
+      outputs = standard_ops.tensordot(inputs, self.kernel, [[rank - 1], [0]])
       # Reshape the output back to the original ndim of the input.
       if not context.executing_eagerly():
+        shape = inputs.get_shape().as_list()
         output_shape = shape[:-1] + [self.units]
         outputs.set_shape(output_shape)
     else:

tensorflow/python/ops/math_ops.py

@@ -651,6 +651,9 @@ def cast(x, dtype, name=None):
     TypeError: If `x` cannot be cast to the `dtype`.
   """
   base_type = dtypes.as_dtype(dtype).base_dtype
+  if isinstance(x,
+                (ops.Tensor, _resource_variable_type)) and base_type == x.dtype:
+    return x
   with ops.name_scope(name, "Cast", [x]) as name:
     if isinstance(x, sparse_tensor.SparseTensor):
       values_cast = cast(x.values, base_type, name=name)
@@ -1222,8 +1225,9 @@ def _ReductionDims(x, axis, reduction_indices):
     return axis
   else:
     # Fast path: avoid creating Rank and Range ops if ndims is known.
-    if isinstance(x, ops.Tensor) and x._rank() is not None:  # pylint: disable=protected-access
-      return constant_op.constant(np.arange(x._rank()), dtype=dtypes.int32)  # pylint: disable=protected-access
+    rank = common_shapes.rank(x)
+    if rank is not None:
+      return constant_op.constant(np.arange(rank), dtype=dtypes.int32)
     if (isinstance(x, sparse_tensor.SparseTensor) and
         x.dense_shape.get_shape().is_fully_defined()):
       rank = x.dense_shape.get_shape()[0].value  # sparse.dense_shape is 1-D.
@@ -1234,8 +1238,8 @@ def _ReductionDims(x, axis, reduction_indices):
 
 
 def _may_reduce_to_scalar(keepdims, axis, reduction_indices, output):
-  """Set a reduction's output's shape to be a scalar if we are certain."""
-  if (not output.shape.is_fully_defined()) and (not keepdims) and (
+  """Set a reduction's output shape to be a scalar if we are certain."""
+  if not common_shapes.has_fully_defined_shape(output) and (not keepdims) and (
       axis is None) and (reduction_indices is None):
     output.set_shape(())
   return output