File size: 9,832 Bytes
97b6013
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
# Copyright 2017 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.
# ==============================================================================

"""Utilities and helper functions."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import numpy as np
import tensorflow as tf


def make_finite(t, replacement):
  """Replaces non-finite tensor values with the replacement value."""
  return tf.where(tf.is_finite(t), t, replacement)


def asinh(x):
  """Computes the inverse hyperbolic sine function (in tensorflow)."""
  return tf.log(x + tf.sqrt(1. + x ** 2))


def affine(inputs, output_size, scope="Affine", scale=0.1, vec_mean=0.,
           include_bias=True, bias_init=0., random_seed=None):
  """Computes an affine function of the inputs.

  Creates or recalls tensorflow variables "Matrix" and "Bias"
  to generate an affine operation on the input.

  If the inputs are a list of tensors, they are concatenated together.

  Initial weights for the matrix are drawn from a Gaussian with zero
  mean and standard deviation that is the given scale divided by the
  square root of the input dimension. Initial weights for the bias are
  set to zero.

  Args:
    inputs: List of tensors with shape (batch_size, input_size)
    output_size: Size (dimension) of the output
    scope: Variable scope for these parameters (default: "Affine")
    scale: Initial weight scale for the matrix parameters (default: 0.1),
      this constant is divided by the sqrt of the input size to get the
      std. deviation of the initial weights
    vec_mean: The mean for the random initializer
    include_bias: Whether to include the bias term
    bias_init: The initializer bias (default 0.)
    random_seed: Random seed for random initializers. (Default: None)

  Returns:
    output: Tensor with shape (batch_size, output_size)
  """

  # Concatenate the input arguments.
  x = tf.concat(inputs, 1)

  with tf.variable_scope(scope):
    input_size = x.get_shape().as_list()[1]

    sigma = scale / np.sqrt(input_size)
    rand_init = tf.random_normal_initializer(mean=vec_mean, stddev=sigma,
                                             seed=random_seed)

    matrix = tf.get_variable("Matrix", [input_size, output_size],
                             dtype=tf.float32, initializer=rand_init)

    if include_bias:
      bias = tf.get_variable("Bias", [output_size], dtype=tf.float32,
                             initializer=tf.constant_initializer(bias_init,
                                                                 tf.float32))
    else:
      bias = 0.
    output = tf.matmul(x, matrix) + bias

  return output


def project(inputs, weights, bias=0., activation=tf.identity):
  """Computes an affine or linear projection of the inputs.

  Projects the inputs onto the given weight vector and (optionally)
  adds a bias and passes the result through an activation function.

  Args:
    inputs: matrix of inputs with shape [batch_size, dim]
    weights: weight matrix with shape [dim, output_dim]
    bias: bias vector with shape [output_dim] (default: 0)
    activation: nonlinear activation function (default: tf.identity)

  Returns:
    outputs: an op which computes activation(inputs @ weights + bias)
  """
  return activation(tf.matmul(inputs, weights) + bias)


def new_mean_squared(grad_vec, decay, ms):
  """Calculates the new accumulated mean squared of the gradient.

  Args:
    grad_vec: the vector for the current gradient
    decay: the decay term
    ms: the previous mean_squared value

  Returns:
    the new mean_squared value
  """
  decay_size = decay.get_shape().num_elements()
  decay_check_ops = [
      tf.assert_less_equal(decay, 1., summarize=decay_size),
      tf.assert_greater_equal(decay, 0., summarize=decay_size)]

  with tf.control_dependencies(decay_check_ops):
    grad_squared = tf.square(grad_vec)

  # If the previous mean_squared is the 0 vector, don't use the decay and just
  # return the full grad_squared. This should only happen on the first timestep.
  decay = tf.cond(tf.reduce_all(tf.equal(ms, 0.)),
                  lambda: tf.zeros_like(decay, dtype=tf.float32), lambda: decay)

  # Update the running average of squared gradients.
  epsilon = 1e-12
  return (1. - decay) * (grad_squared + epsilon) + decay * ms


def rms_scaling(gradient, decay, ms, update_ms=True):
  """Vectorizes and scales a tensor of gradients.

  Args:
    gradient: the current gradient
    decay: the current decay value.
    ms: the previous mean squared value
    update_ms: Whether to update the mean squared value (default: True)

  Returns:
    The scaled gradient and the new ms value if update_ms is True,
    the old ms value otherwise.
  """

  # Vectorize the gradients and compute the squared gradients.
  grad_vec = tf.reshape(gradient, [-1, 1])

  if update_ms:
    ms = new_mean_squared(grad_vec, decay, ms)

  # Scale the current gradients by the RMS, squashed by the asinh function.
  scaled_gradient = asinh(grad_vec / tf.sqrt(ms + 1e-16))

  return scaled_gradient, ms


def accumulate_sparse_gradients(grad):
  """Accumulates repeated indices of a sparse gradient update.

  Args:
    grad: a tf.IndexedSlices gradient

  Returns:
    grad_indices: unique indices
    grad_values: gradient values corresponding to the indices
  """

  grad_indices, grad_segments = tf.unique(grad.indices)
  grad_values = tf.unsorted_segment_sum(grad.values, grad_segments,
                                        tf.shape(grad_indices)[0])
  return grad_indices, grad_values


def slice_tensor(dense_tensor, indices, head_dims):
  """Extracts slices from a partially flattened dense tensor.

  indices is assumed to index into the first dimension of head_dims.
  dense_tensor is assumed to have a shape [D_0, D_1, ...] such that
  prod(head_dims) == D_0. This function will extract slices along the
  first_dimension of head_dims.

  Example:

  Consider a tensor with shape head_dims = [100, 2] and a dense_tensor with
  shape [200, 3]. Note that the first dimension of dense_tensor equals the
  product of head_dims. This function will reshape dense_tensor such that
  its shape is now [100, 2, 3] (i.e. the first dimension became head-dims)
  and then slice it along the first dimension. After slicing, the slices will
  have their initial dimensions flattened just as they were in dense_tensor
  (e.g. if there are 4 indices, the return value will have a shape of [4, 3]).

  Args:
    dense_tensor: a N-D dense tensor. Shape: [D_0, D_1, ...]
    indices: a 1-D integer tensor. Shape: [K]
    head_dims: True dimensions of the dense_tensor's first dimension.

  Returns:
    Extracted slices. Shape [K, D_1, ...]
  """

  tail_dims = tf.shape(dense_tensor)[1:]
  dense_tensor = tf.reshape(dense_tensor,
                            tf.concat([head_dims, tail_dims], 0))

  slices = tf.gather(dense_tensor, indices)
  # NOTE(siege): This kills the shape annotation.
  return tf.reshape(slices, tf.concat([[-1], tail_dims], 0))


def stack_tensor(slices, indices, dense_tensor, head_dims):
  """Reconsititutes a tensor from slices and corresponding indices.

  This is an inverse operation to slice_tensor. Missing slices are set to 0.

  Args:
    slices: a tensor. Shape [K, D_1, ...]
    indices: a 1-D integer tensor. Shape: [K]
    dense_tensor: the original tensor the slices were taken
      from. Shape: [D_0, D_1, ...]
    head_dims: True dimensions of the dense_tensor's first dimension.

  Returns:
    Reconsituted tensor. Shape: [D_0, D_1, ...]
  """
  # NOTE(siege): This cast shouldn't be necessary.
  indices = tf.cast(indices, tf.int32)

  tail_dims = tf.shape(dense_tensor)[1:]
  dense_shape = tf.concat([head_dims, tail_dims], 0)

  slices = tf.reshape(slices, tf.concat([[-1], dense_shape[1:]], 0))
  indices = tf.expand_dims(indices, -1)

  return tf.reshape(tf.scatter_nd(indices, slices, dense_shape),
                    tf.shape(dense_tensor))


def update_slices(slices, indices, dense_tensor, head_dims):
  """Reconstitutes a tensor from slices and corresponding indices.

  Like _stack_tensor, but instead of setting missing slices to 0, sets them to
  what they were in the original tensor. The return value is reshaped to be
  the same as dense_tensor.

  Args:
    slices: a tensor. Shape [K, D_1, ...]
    indices: a 1-D integer tensor. Shape: [K]
    dense_tensor: the original tensor the slices were taken
      from. Shape: [D_0, D_1, ...]
    head_dims: True dimensions of the dense_tensor's first dimension.

  Returns:
    Reconsituted tensor. Shape: [D_0, D_1, ...]
  """
  # NOTE(siege): This cast shouldn't be necessary.
  indices = tf.cast(indices, tf.int32)

  tail_dims = tf.shape(dense_tensor)[1:]
  dense_shape = tf.concat([head_dims, tail_dims], 0)

  update_mask_vals = tf.fill(tf.shape(indices), 1)
  reshaped_indices = tf.expand_dims(indices, -1)
  update_mask = tf.equal(
      tf.scatter_nd(reshaped_indices, update_mask_vals, head_dims[:1]), 1)

  reshaped_dense_slices = tf.reshape(
      stack_tensor(slices, indices, dense_tensor, head_dims), dense_shape)
  reshaped_dense_tensor = tf.reshape(dense_tensor, dense_shape)

  return tf.reshape(
      tf.where(update_mask, reshaped_dense_slices, reshaped_dense_tensor),
      tf.shape(dense_tensor))