• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7#     http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
15"""Functional tests for BiasAdd."""
16
17from __future__ import absolute_import
18from __future__ import division
19from __future__ import print_function
20
21import numpy as np
22
23from tensorflow.python.framework import constant_op
24from tensorflow.python.framework import dtypes
25from tensorflow.python.framework import test_util
26from tensorflow.python.ops import array_ops
27from tensorflow.python.ops import gradient_checker
28from tensorflow.python.ops import gradients_impl
29from tensorflow.python.ops import nn_ops
30import tensorflow.python.ops.nn_grad  # pylint: disable=unused-import
31from tensorflow.python.platform import test
32
33
34class BiasAddTestBase(test.TestCase):
35
36  def _npBias(self, inputs, bias):
37    assert len(bias.shape) == 1
38    assert inputs.shape[-1] == bias.shape[0]
39    return inputs + bias.reshape(([1] *
40                                  (len(inputs.shape) - 1)) + [bias.shape[0]])
41
42  def testNpBias(self):
43    self.assertAllClose(
44        np.array([[11, 22, 33], [41, 52, 63]]),
45        self._npBias(
46            np.array([[10, 20, 30], [40, 50, 60]]), np.array([1, 2, 3])))
47
48  def _testBias(self, np_inputs, np_bias, use_gpu=False):
49    np_val = self._npBias(np_inputs, np_bias)
50    with self.cached_session(use_gpu=use_gpu):
51      tf_val = nn_ops.bias_add(np_inputs, np_bias).eval()
52    self.assertAllCloseAccordingToType(np_val, tf_val)
53
54  def _AtLeast3d(self, np_value):
55    # fill the input value to at least 3-dimension
56    if np_value.ndim < 3:
57      return np.reshape(np_value, (1,) * (3 - np_value.ndim) + np_value.shape)
58    return np_value
59
60  def _NHWCToNCHW(self, np_value):
61    # fill the input value to at least 3-dimension
62    np_value = self._AtLeast3d(np_value)
63    # move the last dimension to second
64    np_dim = list(range(np_value.ndim))
65    np_dim_new = list(np_dim[0:1]) + list(np_dim[-1:]) + list(np_dim[1:-1])
66    return np.transpose(np_value, np_dim_new)
67
68  def _NCHWToNHWC(self, np_value):
69    assert len(np_value.shape) >= 3
70    np_dim = list(range(np_value.ndim))
71    # move the second dimension to the last
72    np_dim_new = list(np_dim[0:1]) + list(np_dim[2:]) + list(np_dim[1:2])
73    return np.transpose(np_value, np_dim_new)
74
75  def _testBiasNCHW(self, np_inputs, np_bias, use_gpu):
76    np_val = self._npBias(np_inputs, np_bias)
77    np_inputs = self._NHWCToNCHW(np_inputs)
78    with self.cached_session(use_gpu=use_gpu):
79      tf_val = nn_ops.bias_add(np_inputs, np_bias, data_format="NCHW").eval()
80    tf_val = self._NCHWToNHWC(tf_val)
81    self.assertAllCloseAccordingToType(self._AtLeast3d(np_val), tf_val)
82
83  def _testAll(self, np_inputs, np_bias):
84    self._testBias(np_inputs, np_bias, use_gpu=False)
85    self._testBiasNCHW(np_inputs, np_bias, use_gpu=False)
86    if np_inputs.dtype in [np.float16, np.float32, np.float64, np.int32]:
87      self._testBias(np_inputs, np_bias, use_gpu=True)
88      self._testBiasNCHW(np_inputs, np_bias, use_gpu=True)
89
90  @test_util.run_deprecated_v1
91  def testInputDims(self):
92    with self.assertRaises(ValueError):
93      nn_ops.bias_add([1, 2], [1])
94
95  @test_util.run_deprecated_v1
96  def testBiasVec(self):
97    with self.assertRaises(ValueError):
98      nn_ops.bias_add(
99          array_ops.reshape([1, 2], shape=[1, 2]),
100          array_ops.reshape([1, 2], shape=[1, 2]))
101
102  @test_util.run_deprecated_v1
103  def testBiasInputsMatch(self):
104    with self.assertRaises(ValueError):
105      nn_ops.bias_add(
106          array_ops.reshape([1, 2], shape=[1, 2]),
107          array_ops.reshape([1], shape=[1]))
108
109  @test_util.run_deprecated_v1
110  def testIntTypes(self):
111    for t in [np.int8, np.int16, np.int32, np.int64]:
112      self._testAll(
113          np.array([[10, 20, 30], [40, 50, 60]]).astype(t),
114          np.array([1, 2, 3]).astype(t))
115
116  @test_util.run_deprecated_v1
117  def testFloatTypes(self):
118    for t in [np.float16, np.float32, np.float64]:
119      self._testAll(
120          np.random.rand(4, 3, 3).astype(t),
121          np.random.rand(3).astype(t))
122
123  @test_util.run_deprecated_v1
124  def test4DFloatTypes(self):
125    for t in [np.float16, np.float32, np.float64]:
126      self._testAll(
127          np.random.rand(4, 3, 2, 3).astype(t),
128          np.random.rand(3).astype(t))
129      self._testAll(
130          np.random.rand(2048, 4, 4, 4).astype(t),
131          np.random.rand(4).astype(t))
132      self._testAll(
133          np.random.rand(4, 4, 4, 2048).astype(t),
134          np.random.rand(2048).astype(t))
135
136  @test_util.run_deprecated_v1
137  def test5DFloatTypes(self):
138    for t in [np.float16, np.float32, np.float64]:
139      self._testAll(
140          np.random.rand(4, 3, 2, 3, 4).astype(t),
141          np.random.rand(4).astype(t))
142
143  def _testGradient(self, np_input, bias, dtype, data_format, use_gpu):
144    with self.cached_session(use_gpu=use_gpu):
145      if data_format == "NCHW":
146        np_input = self._NHWCToNCHW(np_input)
147      input_tensor = constant_op.constant(
148          np_input, shape=np_input.shape, dtype=dtype)
149      bias_tensor = constant_op.constant(bias, shape=bias.shape, dtype=dtype)
150      output_tensor = nn_ops.bias_add(
151          input_tensor, bias_tensor, data_format=data_format)
152      tensor_jacob_t, tensor_jacob_n = gradient_checker.compute_gradient(
153          input_tensor, np_input.shape, output_tensor, np_input.shape)
154      bias_jacob_t, bias_jacob_n = gradient_checker.compute_gradient(
155          bias_tensor, bias.shape, output_tensor, np_input.shape)
156
157      # Test gradient of BiasAddGrad
158      bias_add_grad = gradients_impl.gradients(
159          nn_ops.l2_loss(output_tensor), bias_tensor)[0]
160      grad_jacob_t, grad_jacob_n = gradient_checker.compute_gradient(
161          output_tensor, np_input.shape, bias_add_grad, bias.shape)
162
163      if dtype == np.float16:
164        # Compare fp16 analytical gradients to fp32 numerical gradients,
165        # since fp16 numerical gradients are too imprecise unless great
166        # care is taken with choosing the inputs and the delta. This is
167        # a weaker, but pragmatic, check (in particular, it does not test
168        # the op itself, only its gradient).
169        input_tensor = constant_op.constant(
170            np_input, shape=np_input.shape, dtype=np.float32)
171        bias_tensor = constant_op.constant(
172            bias, shape=bias.shape, dtype=np.float32)
173        output_tensor = nn_ops.bias_add(
174            input_tensor, bias_tensor, data_format=data_format)
175        _, tensor_jacob_n = gradient_checker.compute_gradient(
176            input_tensor, np_input.shape, output_tensor, np_input.shape)
177        _, bias_jacob_n = gradient_checker.compute_gradient(
178            bias_tensor, bias.shape, output_tensor, np_input.shape)
179
180        bias_add_grad = gradients_impl.gradients(
181            nn_ops.l2_loss(output_tensor), bias_tensor)[0]
182        _, grad_jacob_n = gradient_checker.compute_gradient(
183            output_tensor, np_input.shape, bias_add_grad, bias.shape)
184
185      threshold = 5e-3
186      if dtype == dtypes.float64:
187        threshold = 1e-10
188      self.assertAllClose(tensor_jacob_t, tensor_jacob_n, threshold, threshold)
189      self.assertAllClose(bias_jacob_t, bias_jacob_n, threshold, threshold)
190      self.assertAllClose(grad_jacob_t, grad_jacob_n, threshold, threshold)
191
192  @test_util.run_deprecated_v1
193  def testGradientTensor2D(self):
194    for (data_format, use_gpu) in ("NHWC", False), ("NHWC", True):
195      for dtype in (dtypes.float16, dtypes.float32, dtypes.float64):
196        np_input = np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0],
197                            dtype=dtype.as_numpy_dtype).reshape(3, 2)
198        bias = np.array([1.3, 2.4], dtype=dtype.as_numpy_dtype)
199        self._testGradient(np_input, bias, dtype, data_format, use_gpu)
200
201  @test_util.run_deprecated_v1
202  def testGradientTensor3D(self):
203    for (data_format, use_gpu) in [("NHWC", False), ("NHWC", True),
204                                   ("NCHW", False), ("NCHW", True)]:
205      for dtype in (dtypes.float16, dtypes.float32, dtypes.float64):
206        np_input = np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0],
207                            dtype=dtype.as_numpy_dtype).reshape(1, 3, 2)
208        bias = np.array([1.3, 2.4], dtype=dtype.as_numpy_dtype)
209        self._testGradient(np_input, bias, dtype, data_format, use_gpu)
210
211  @test_util.run_deprecated_v1
212  def testGradientTensor4D(self):
213    for (data_format, use_gpu) in [("NHWC", False)]:
214      for dtype in (dtypes.float16, dtypes.float32, dtypes.float64):
215        np_input = np.arange(
216            1.0, 49.0,
217            dtype=dtype.as_numpy_dtype).reshape([2, 3, 4, 2]).astype(np.float32)
218        bias = np.array([1.3, 2.4], dtype=dtype.as_numpy_dtype)
219        self._testGradient(np_input, bias, dtype, data_format, use_gpu)
220        np_input = np.arange(
221            1.0, 513.0,
222            dtype=dtype.as_numpy_dtype).reshape([64, 2, 2,
223                                                 2]).astype(np.float32)
224        self._testGradient(np_input, bias, dtype, data_format, use_gpu)
225        np_input = np.arange(
226            1.0, 513.0,
227            dtype=dtype.as_numpy_dtype).reshape([2, 2, 2,
228                                                 64]).astype(np.float32)
229        self._testGradient(np_input,
230                           np.random.rand(64).astype(dtype.as_numpy_dtype),
231                           dtype, data_format, use_gpu)
232
233  @test_util.run_deprecated_v1
234  def testGradientTensor5D(self):
235    for (data_format, use_gpu) in [("NHWC", False), ("NHWC", True),
236                                   ("NCHW", False), ("NCHW", True)]:
237      for dtype in (dtypes.float16, dtypes.float32, dtypes.float64):
238        np_input = np.arange(
239            1.0, 49.0,
240            dtype=dtype.as_numpy_dtype).reshape([1, 2, 3, 4,
241                                                 2]).astype(np.float32)
242        bias = np.array([1.3, 2.4], dtype=dtype.as_numpy_dtype)
243        self._testGradient(np_input, bias, dtype, data_format, use_gpu)
244
245  @test_util.run_deprecated_v1
246  def testEmpty(self):
247    np.random.seed(7)
248    for shape in (0, 0), (2, 0), (0, 2), (4, 3, 0), (4, 0, 3), (0, 4, 3):
249      self._testAll(np.random.randn(*shape), np.random.randn(shape[-1]))
250
251  @test_util.run_deprecated_v1
252  def testEmptyGradient(self):
253    for (data_format, use_gpu) in ("NHWC", False), ("NHWC", True):
254      for shape in (0, 0), (2, 0), (0, 2):
255        self._testGradient(
256            np.random.randn(*shape), np.random.randn(shape[-1]), dtypes.float64,
257            data_format, use_gpu)
258
259    for (data_format, use_gpu) in [("NHWC", False), ("NHWC", True),
260                                   ("NCHW", False), ("NCHW", True)]:
261      for shape in (4, 3, 0), (4, 0, 3), (0, 4, 3):
262        self._testGradient(
263            np.random.randn(*shape), np.random.randn(shape[-1]), dtypes.float64,
264            data_format, use_gpu)
265