experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

makes gradient_checker_v2 work with sparse tensor reshape.

Browse Source 
PiperOrigin-RevId: 301667976
Change-Id: I133c918c8a5703b40d4b5a69ead4bc0ec74f27b8
This commit is contained in:
A. Unique TensorFlower 2020-03-18 14:23:48 -07:00 committed by TensorFlower Gardener
parent 5041af7879
commit 5b00c31c4b
2 changed files with 19 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tensorflow/python/ops/gradient_checker_v2.py
View File

@ -174,7 +174,6 @@ def _compute_theoretical_jacobian(f, y_shape, y_dtype, xs, param):
dy_data_flat[row] = 1 dy_data_flat[row] = 1
grad = _to_numpy(grad_fn(dy_data, *xs)[0]) grad = _to_numpy(grad_fn(dy_data, *xs)[0])
grad = _eval_indexed_slices(grad) grad = _eval_indexed_slices(grad)
dy_data_flat[row] = 0
if isinstance(grad, ops.IndexedSlicesValue): if isinstance(grad, ops.IndexedSlicesValue):
for i, v in zip(grad.indices, grad.values): for i, v in zip(grad.indices, grad.values):
c_begin = i * x_val_size c_begin = i * x_val_size
@ -182,6 +181,9 @@ def _compute_theoretical_jacobian(f, y_shape, y_dtype, xs, param):
jacobian[row, c_begin:c_end] += v.flat jacobian[row, c_begin:c_end] += v.flat
elif grad is not None: elif grad is not None:
jacobian[row, :] = grad.ravel().view(jacobian.dtype) jacobian[row, :] = grad.ravel().view(jacobian.dtype)
# This reset of `dy_data_flat` needs to happen after `grad` is copied to
# `jacobian` because `grad` and `dy_data_flat` may share memory.
dy_data_flat[row] = 0
# If the output is empty, run the gradients at least once and make sure # If the output is empty, run the gradients at least once and make sure
# they produce zeros. # they produce zeros.

16
tensorflow/python/ops/gradient_checker_v2_test.py
View File

@ -23,6 +23,7 @@ import numpy as np
from tensorflow.python.eager import backprop from tensorflow.python.eager import backprop
from tensorflow.python.framework import constant_op from tensorflow.python.framework import constant_op
from tensorflow.python.framework import dtypes from tensorflow.python.framework import dtypes
from tensorflow.python.framework import sparse_tensor
from tensorflow.python.framework import test_util from tensorflow.python.framework import test_util
from tensorflow.python.ops import array_ops from tensorflow.python.ops import array_ops
from tensorflow.python.ops import custom_gradient from tensorflow.python.ops import custom_gradient
@ -30,6 +31,7 @@ from tensorflow.python.ops import \
gradient_checker_v2 as gradient_checker gradient_checker_v2 as gradient_checker
from tensorflow.python.ops import math_ops from tensorflow.python.ops import math_ops
from tensorflow.python.ops import nn_ops from tensorflow.python.ops import nn_ops
from tensorflow.python.ops import sparse_ops
# needs this to register gradient for SoftmaxCrossEntropyWithLogits: # needs this to register gradient for SoftmaxCrossEntropyWithLogits:
import tensorflow.python.ops.nn_grad # pylint: disable=unused-import import tensorflow.python.ops.nn_grad # pylint: disable=unused-import
from tensorflow.python.platform import test from tensorflow.python.platform import test
@ -46,6 +48,20 @@ def _random_complex(shape, dtype):
@test_util.run_all_in_graph_and_eager_modes @test_util.run_all_in_graph_and_eager_modes
class GradientCheckerTest(test.TestCase): class GradientCheckerTest(test.TestCase):
def testSparseTensorReshape(self):
x = constant_op.constant(2.0, shape=(2,))
def sparse_tensor_reshape(values):
sparse = sparse_tensor.SparseTensor(
indices=[[0, 0], [1, 2]], values=values, dense_shape=[3, 4])
sparse = sparse_ops.sparse_reshape(sparse, shape=(12,))
return sparse.values
error = gradient_checker.max_error(
*gradient_checker.compute_gradient(sparse_tensor_reshape, [x]))
self.assertLess(error, 1e-4)
def testWithStaticShape(self): def testWithStaticShape(self):
size = (2, 3) size = (2, 3)
constant = constant_op.constant(2.0, shape=size, name="const") constant = constant_op.constant(2.0, shape=size, name="const")