/
ctc.py
407 lines (342 loc) · 15.6 KB
/
ctc.py
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
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
import collections
import numpy
import six
import chainer
from chainer.backends import cuda
from chainer import function
from chainer import utils
from chainer.utils import type_check
def _logsumexp(a, xp, axis=None):
vmax = xp.amax(a, axis=axis, keepdims=True)
if xp is numpy:
vmax += xp.log(xp.sum(xp.exp(a - vmax),
axis=axis, keepdims=True, dtype=a.dtype))
else:
_logsumexp_impl = cuda.reduce(
'T x, T vmax', 'T y',
'exp(x - vmax)', 'a + b', 'y += log(a)', '0',
'logsumexp_impl')
_logsumexp_impl(a, vmax, vmax, axis=axis, keepdims=True)
return xp.squeeze(vmax, axis=axis)
def _softmax(x, xp):
val = xp.exp(x - xp.amax(x, axis=2, keepdims=True))
val /= xp.sum(val, axis=2, keepdims=True)
return val
def _label_to_path(labels, blank_symbol, xp):
path = xp.full((len(labels), labels.shape[1] * 2 + 1),
blank_symbol, dtype=numpy.int32)
path[:, 1::2] = labels
return path
def _flip_path(path, path_length, xp):
"""Flips label sequence.
This function rotates a label sequence and flips it.
``path[b, t]`` stores a label at time ``t`` in ``b``-th batch.
The rotated matrix ``r`` is defined as
``r[b, t] = path[b, t + path_length[b]]``
.. ::
a b c d . . a b c d d c b a .
e f . . . -> . . . e f -> f e . . .
g h i j k g h i j k k j i h g
"""
n_batch, n_label = path.shape
rotate = (xp.arange(n_label) + path_length[:, None]) % n_label
return path[xp.arange(n_batch, dtype='i')[:, None],
rotate][:, ::-1]
def _flip_label_probability(y, input_length, xp):
"""Flips a label probability matrix.
This function rotates a label probability matrix and flips it.
``y[i, b, l]`` stores log probability of label ``l`` at ``i``-th
input in ``b``-th batch.
The rotated matrix ``r`` is defined as
``r[i, b, l] = y[i + input_length[b], b, l]``
"""
seq, n_batch, n_vocab = y.shape
rotate = (xp.arange(seq, dtype='i')[:, None] + input_length) % seq
return y[
rotate[:, :, None],
xp.arange(n_batch, dtype='i')[None, :, None],
xp.arange(n_vocab, dtype='i')[None, None, :]][::-1]
def _flip_path_probability(prob, input_length, path_length, xp):
"""Flips a path probability matrix.
This function returns a path probability matrix and flips it.
``prob[i, b, t]`` stores log probability at ``i``-th input and
at time ``t`` in a output sequence in ``b``-th batch.
The rotated matrix ``r`` is defined as
``r[i, j, k] = prob[i + input_length[j], j, k + path_length[j]]``
"""
seq, n_batch, n_label = prob.shape
rotate_input = (xp.arange(seq, dtype='i')[:, None] + input_length) % seq
rotate_label = (
xp.arange(n_label, dtype='i') + path_length[:, None]) % n_label
return prob[
rotate_input[:, :, None],
xp.arange(n_batch, dtype='i')[None, :, None],
rotate_label][::-1, :, ::-1]
class ConnectionistTemporalClassification(function.Function):
"""The implementation of Connectionist Temporal Classfication loss functions.
To make it usable for real-world cases, this class has two policies below.
1. This class computes forward and backward variables in the log domain.
2. This class applies the softmax function to inputs. The Backward
values of CTC loss is often overflows. This is avoided by computing
backward values before the activation function is applied.
"""
def __init__(self, blank_symbol, reduce='mean'):
self.blank_symbol = blank_symbol
self.zero_padding = -10000000000.0
if reduce not in ('mean', 'no'):
raise ValueError(
"only 'mean' and 'no' are valid "
"for 'reduce', but '%s' is given" % reduce)
self.reduce = reduce
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 4)
input_length_type, label_length_type, t_type, x_type = in_types
type_check.expect(
input_length_type.dtype == numpy.int32,
input_length_type.ndim == 1,
label_length_type.dtype == numpy.int32,
label_length_type.ndim == 1,
t_type.ndim == 2,
t_type.dtype == numpy.int32,
x_type.ndim == 3,
x_type.dtype == numpy.float32,
)
n_batch = x_type.shape[1]
type_check.expect(
t_type.shape[0] == n_batch,
input_length_type.shape[0] == n_batch,
label_length_type.shape[0] == n_batch,
)
def log_matrix(self, x, xp):
if xp == numpy:
res = numpy.ma.log(x).filled(fill_value=self.zero_padding)
else:
create_recurrence_relation = cuda.cupy.ElementwiseKernel(
'T x, T e', 'T y',
'y = x == 0 ? e : log(x)',
'create_recurrence_relation')
res = create_recurrence_relation(x, self.zero_padding)
return res.astype(numpy.float32)
# path probablity to label probability
def label_probability(self, label_size, path, path_length,
multiply_seq, xp):
seq_length = len(multiply_seq)
n_batch = len(path)
dtype = multiply_seq.dtype
ret = xp.zeros((seq_length, n_batch, label_size), dtype)
if xp == numpy:
for b in six.moves.range(len(path)):
target_path = path[b, :path_length[b]]
chars = {c for c in target_path}
for c in chars:
ret[:, b, c] = xp.sum(
multiply_seq[:, b, 0:path_length[b]]
[:, target_path == c], axis=1)
else:
cuda.cupy.ElementwiseKernel(
'T prob, I path, I path_length, I max_path_length',
'raw T cum_prob',
'''
I t = i % max_path_length;
if (t < path_length) {
int n_batch = cum_prob.shape()[1];
I s = i / (max_path_length * n_batch);
I b = (i - s * (max_path_length * n_batch))
/ max_path_length;
int ind[] = {s, b, path};
atomicAdd(&cum_prob[ind], prob);
}
''', 'ctc_label_prob_sum'
)(multiply_seq, path, path_length[:, None], path.shape[1], ret)
return ret
def _computes_transition(
self, prev_prob, path, path_length, cum_prob, y):
xp = cuda.get_array_module(prev_prob)
if xp == numpy:
n_batch, max_path_length = path.shape
mat = xp.full(
(3, n_batch, max_path_length), self.zero_padding, 'f')
mat[0, :, :] = prev_prob
mat[1, :, 1:] = prev_prob[:, :-1]
mat[2, :, 2:] = prev_prob[:, :-2]
# disable transition between the same symbols
# (including blank-to-blank)
same_transition = (path[:, :-2] == path[:, 2:])
mat[2, :, 2:][same_transition] = self.zero_padding
prob = _logsumexp(mat, xp, axis=0)
outside = xp.arange(max_path_length) >= path_length[:, None]
prob[outside] = self.zero_padding
cum_prob += prob
batch_index = xp.arange(n_batch, dtype='i')
prob += y[batch_index[:, None], path]
else:
prob = xp.empty_like(prev_prob)
cuda.elementwise(
'raw T prob, raw I path, I path_length, T zero, raw T y',
'T z, T cum_prob',
'''
int length = prob.shape()[1];
int b = i / length;
int t = i - b * length;
if (t >= path_length) {
z = zero;
cum_prob += zero;
return;
}
int ind1[] = {b, t};
int ind2[] = {b, t - 1};
int ind3[] = {b, t - 2};
float f1 = prob[ind1];
float f2 = (0 <= t - 1) ? prob[ind2] : zero;
float f3 = (0 <= t - 2 && path[ind3] != path[ind1]) ?
prob[ind3] : zero;
// calculates log-sum-exp
float m = max(f1, max(f2, f3));
z = m + log(exp(f1 - m) + exp(f2 - m) + exp(f3 - m));
cum_prob += z;
int y_ind[] = {b, path[ind1]};
z += y[y_ind];
''', 'ctc_transition'
)(prev_prob, path, path_length[:, None], self.zero_padding, y,
prob, cum_prob)
return prob
def calc_trans(self, yseq, input_length,
label, label_length, path, path_length, xp):
max_input_length, n_batch, n_unit = yseq.shape
max_label_length = label.shape[1]
max_path_length = path.shape[1]
assert label.shape == (n_batch, max_label_length), label.shape
assert path.shape == (n_batch, max_label_length * 2 + 1)
forward_prob = xp.full(
(n_batch, max_path_length), self.zero_padding, dtype='f')
forward_prob[:, 0] = 0
backward_prob = forward_prob
batch_index = xp.arange(n_batch, dtype='i')
seq_index = xp.arange(len(yseq), dtype='i')
prob = yseq[seq_index[:, None, None], batch_index[:, None], path]
# forward computation.
for i, y in enumerate(yseq):
forward_prob = self._computes_transition(
forward_prob, path, path_length, prob[i], y)
r_path = _flip_path(path, path_length, xp)
yseq_inv = _flip_label_probability(yseq, input_length, xp)
prob = _flip_path_probability(prob, input_length, path_length, xp)
for i, y_inv in enumerate(yseq_inv):
backward_prob = self._computes_transition(
backward_prob, r_path, path_length, prob[i], y_inv)
return _flip_path_probability(prob, input_length, path_length, xp)
def forward(self, inputs):
xp = cuda.get_array_module(inputs[0])
self.input_length, label_length, t, xs = inputs
if chainer.is_debug():
assert len(xs) >= xp.max(self.input_length)
assert t.shape[1] >= xp.max(label_length)
self.path_length = 2 * label_length + 1
self.yseq = _softmax(xs, xp)
log_yseq = self.log_matrix(self.yseq, xp)
self.path = _label_to_path(t, self.blank_symbol, xp)
self.prob_trans = self.calc_trans(
log_yseq, self.input_length, t,
label_length, self.path, self.path_length, xp)
loss = -_logsumexp(self.prob_trans[0], xp, axis=1)
if self.reduce == 'mean':
loss = utils.force_array(xp.mean(loss))
return loss,
def backward(self, inputs, grad_output):
xp = cuda.get_array_module(inputs[0])
batch_size = len(inputs[2])
total_probability = _logsumexp(self.prob_trans[0], xp, axis=1)
label_prob = self.label_probability(
self.yseq.shape[2], self.path, self.path_length,
xp.exp(self.prob_trans - total_probability[:, None]), xp)
self.yseq -= label_prob
if self.reduce == 'mean':
self.yseq *= grad_output[0] / batch_size
else:
self.yseq *= grad_output[0][..., None]
# mask
self.yseq *= (
xp.arange(len(self.yseq))[:, None] < self.input_length)[..., None]
return None, None, None, self.yseq
def connectionist_temporal_classification(
x, t, blank_symbol, input_length=None, label_length=None,
reduce='mean'):
"""Connectionist Temporal Classification loss function.
Connectionist Temporal Classification(CTC) [Graves2006]_ is a loss function
of sequence labeling where the alignment between the inputs and target is
unknown. See also [Graves2012]_
The output is a variable whose value depends on the value of
the option ``reduce``. If it is ``'no'``, it holds the samplewise
loss values. If it is ``'mean'``, it takes the mean of loss values.
Args:
x (list or tuple of :class:`~chainer.Variable`):
A list of unnormalized probabilities for labels.
Each element of ``x``, ``x[i]`` is a :class:`~chainer.Variable`
object, which has shape ``(B, V)``, where ``B``
is the batch size and ``V`` is the number of labels.
The softmax of ``x[i]`` represents the probabilities of the labels
at time ``i``.
t (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`):
A matrix including expected label sequences.
Its shape is ``(B, M)``, where ``B`` is the batch size and ``M`` is
the maximum length of the label sequences.
All elements in ``t`` must be less than ``V``, the number of
labels.
blank_symbol (int): Index of blank_symbol.
This value must be non-negative.
input_length (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray` or ``None``):
Length of sequence for each of mini batch ``x`` (optional).
Its shape must be ``(B,)``.
If the ``input_length`` is omitted or ``None``, it assumes that
all of ``x`` is valid input.
label_length (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray` or ``None``):
Length of sequence for each of mini batch ``t`` (optional).
Its shape must be ``(B,)``.
If the ``label_length`` is omitted or ``None``, it assumes that
all of ``t`` is valid input.
reduce (str): Reduction option. Its value must be either
``'mean'`` or ``'no'``. Otherwise,
:class:`ValueError` is raised.
Returns:
~chainer.Variable:
A variable holding a scalar value of the CTC loss.
If ``reduce`` is ``'no'``, the output variable holds array
whose shape is `(B,)` where `B` is the number of samples.
If it is ``'mean'``, it holds a scalar.
.. note::
You need to input ``x`` without applying to activation functions(e.g.
softmax function), because this function applies softmax functions
to ``x`` before calculating CTC loss to avoid numerical limitations.
You also need to apply softmax function to forwarded values before you
decode it.
.. note::
This function is differentiable only by ``x``.
.. note::
This function supports (batch, sequence, 1-dimensional input)-data.
.. [Graves2006] Alex Graves, Santiago Fernandez,\
Faustino Gomez, Jurgen Schmidhuber,\
`Connectionist Temporal Classification: Labelling Unsegmented\
Sequence Data with Recurrent Neural Networks\
<ftp://ftp.idsia.ch/pub/juergen/icml2006.pdf>`_
.. [Graves2012] Alex Graves,\
`Supervised Sequence Labelling with Recurrent Neural Networks\
<https://www.cs.toronto.edu/~graves/preprint.pdf>`_
"""
if not isinstance(x, collections.Sequence):
raise TypeError('x must be a list of Variables')
if not isinstance(blank_symbol, int):
raise TypeError('blank_symbol must be non-negative integer.')
assert 0 <= blank_symbol < x[0].shape[1]
# This implementation only supports 1-dimensional data.
# TODO(jnishi): Support d(>1)-dimentinal inputs.
assert x[0].ndim == 2
xp = cuda.get_array_module(x[0])
if input_length is None:
input_length = xp.full(len(x[0]), len(x), dtype=numpy.int32)
if label_length is None:
label_length = xp.full(len(t), t.shape[1], dtype=numpy.int32)
return ConnectionistTemporalClassification(blank_symbol, reduce)(
input_length, label_length, t, chainer.functions.stack(x))