forked from vsl9/DeepSpeech
-
Notifications
You must be signed in to change notification settings - Fork 2
/
DeepSpeech.py
executable file
·1828 lines (1465 loc) · 80.7 KB
/
DeepSpeech.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
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#!/usr/bin/env python
# -*- coding: utf-8 -*-
from __future__ import absolute_import, division, print_function
import os
import sys
log_level_index = sys.argv.index('--log_level') + 1 if '--log_level' in sys.argv else 0
os.environ['TF_CPP_MIN_LOG_LEVEL'] = sys.argv[log_level_index] if log_level_index > 0 and log_level_index < len(sys.argv) else '3'
import datetime
import pickle
import shutil
import subprocess
import tensorflow as tf
import time
import traceback
import inspect
from six.moves import zip, range, filter, urllib, BaseHTTPServer
from tensorflow.contrib.session_bundle import exporter
from tensorflow.python.tools import freeze_graph
from threading import Thread, Lock
from util.audio import audiofile_to_input_vector
from util.feeding import DataSet, ModelFeeder
from util.gpu import get_available_gpus
from util.shared_lib import check_cupti
from util.text import sparse_tensor_value_to_texts, wer, Alphabet, ndarray_to_text
from xdg import BaseDirectory as xdg
import numpy as np
# Importer
# ========
tf.app.flags.DEFINE_string ('train_files', '', 'comma separated list of files specifying the dataset used for training. multiple files will get merged')
tf.app.flags.DEFINE_string ('dev_files', '', 'comma separated list of files specifying the dataset used for validation. multiple files will get merged')
tf.app.flags.DEFINE_string ('test_files', '', 'comma separated list of files specifying the dataset used for testing. multiple files will get merged')
tf.app.flags.DEFINE_boolean ('fulltrace', False, 'if full trace debug info should be generated during training')
# Cluster configuration
# =====================
tf.app.flags.DEFINE_string ('ps_hosts', '', 'parameter servers - comma separated list of hostname:port pairs')
tf.app.flags.DEFINE_string ('worker_hosts', '', 'workers - comma separated list of hostname:port pairs')
tf.app.flags.DEFINE_string ('job_name', 'localhost', 'job name - one of localhost (default), worker, ps')
tf.app.flags.DEFINE_integer ('task_index', 0, 'index of task within the job - worker with index 0 will be the chief')
tf.app.flags.DEFINE_integer ('replicas', -1, 'total number of replicas - if negative, its absolute value is multiplied by the number of workers')
tf.app.flags.DEFINE_integer ('replicas_to_agg', -1, 'number of replicas to aggregate - if negative, its absolute value is multiplied by the number of workers')
tf.app.flags.DEFINE_string ('coord_retries', 100, 'number of tries of workers connecting to training coordinator before failing')
tf.app.flags.DEFINE_string ('coord_host', 'localhost', 'coordination server host')
tf.app.flags.DEFINE_integer ('coord_port', 2500, 'coordination server port')
tf.app.flags.DEFINE_integer ('iters_per_worker', 1, 'number of train or inference iterations per worker before results are sent back to coordinator')
# Global Constants
# ================
tf.app.flags.DEFINE_boolean ('train', True, 'wether to train the network')
tf.app.flags.DEFINE_boolean ('test', True, 'wether to test the network')
tf.app.flags.DEFINE_integer ('epoch', 75, 'target epoch to train - if negative, the absolute number of additional epochs will be trained')
tf.app.flags.DEFINE_boolean ('use_warpctc', False, 'wether to use GPU bound Warp-CTC')
tf.app.flags.DEFINE_float ('dropout_rate', 0.05, 'dropout rate for feedforward layers')
tf.app.flags.DEFINE_float ('dropout_rate2', -1.0, 'dropout rate for layer 2 - defaults to dropout_rate')
tf.app.flags.DEFINE_float ('dropout_rate3', -1.0, 'dropout rate for layer 3 - defaults to dropout_rate')
tf.app.flags.DEFINE_float ('dropout_rate4', 0.0, 'dropout rate for layer 4 - defaults to 0.0')
tf.app.flags.DEFINE_float ('dropout_rate5', 0.0, 'dropout rate for layer 5 - defaults to 0.0')
tf.app.flags.DEFINE_float ('dropout_rate6', -1.0, 'dropout rate for layer 6 - defaults to dropout_rate')
tf.app.flags.DEFINE_float ('relu_clip', 20.0, 'ReLU clipping value for non-recurrant layers')
# Adam optimizer (http://arxiv.org/abs/1412.6980) parameters
tf.app.flags.DEFINE_float ('beta1', 0.9, 'beta 1 parameter of Adam optimizer')
tf.app.flags.DEFINE_float ('beta2', 0.999, 'beta 2 parameter of Adam optimizer')
tf.app.flags.DEFINE_float ('epsilon', 1e-8, 'epsilon parameter of Adam optimizer')
tf.app.flags.DEFINE_float ('learning_rate', 0.001, 'learning rate of Adam optimizer')
# Batch sizes
tf.app.flags.DEFINE_integer ('train_batch_size', 1, 'number of elements in a training batch')
tf.app.flags.DEFINE_integer ('dev_batch_size', 1, 'number of elements in a validation batch')
tf.app.flags.DEFINE_integer ('test_batch_size', 1, 'number of elements in a test batch')
# Sample limits
tf.app.flags.DEFINE_integer ('limit_train', 0, 'maximum number of elements to use from train set - 0 means no limit')
tf.app.flags.DEFINE_integer ('limit_dev', 0, 'maximum number of elements to use from validation set- 0 means no limit')
tf.app.flags.DEFINE_integer ('limit_test', 0, 'maximum number of elements to use from test set- 0 means no limit')
# Step widths
tf.app.flags.DEFINE_integer ('display_step', 0, 'number of epochs we cycle through before displaying detailed progress - 0 means no progress display')
tf.app.flags.DEFINE_integer ('validation_step', 0, 'number of epochs we cycle through before validating the model - a detailed progress report is dependent on "--display_step" - 0 means no validation steps')
# Checkpointing
tf.app.flags.DEFINE_string ('checkpoint_dir', '', 'directory in which checkpoints are stored - defaults to directory "deepspeech/checkpoints" within user\'s data home specified by the XDG Base Directory Specification')
tf.app.flags.DEFINE_integer ('checkpoint_secs', 600, 'checkpoint saving interval in seconds')
tf.app.flags.DEFINE_integer ('max_to_keep', 5, 'number of checkpoint files to keep - default value is 5')
# Exporting
tf.app.flags.DEFINE_string ('export_dir', '', 'directory in which exported models are stored - if omitted, the model won\'t get exported')
tf.app.flags.DEFINE_integer ('export_version', 1, 'version number of the exported model')
tf.app.flags.DEFINE_boolean ('remove_export', False, 'wether to remove old exported models')
tf.app.flags.DEFINE_boolean ('use_seq_length', True, 'have sequence_length in the exported graph (will make tfcompile unhappy)')
# Reporting
tf.app.flags.DEFINE_integer ('log_level', 1, 'log level for console logs - 0: INFO, 1: WARN, 2: ERROR, 3: FATAL')
tf.app.flags.DEFINE_boolean ('log_traffic', False, 'log cluster transaction and traffic information during debug logging')
tf.app.flags.DEFINE_string ('wer_log_pattern', '', 'pattern for machine readable global logging of WER progress; has to contain %%s, %%s and %%f for the set name, the date and the float respectively; example: "GLOBAL LOG: logwer(\'12ade231\', %%s, %%s, %%f)" would result in some entry like "GLOBAL LOG: logwer(\'12ade231\', \'train\', \'2017-05-18T03:09:48-0700\', 0.05)"; if omitted (default), there will be no logging')
tf.app.flags.DEFINE_boolean ('log_placement', False, 'wether to log device placement of the operators to the console')
tf.app.flags.DEFINE_integer ('report_count', 10, 'number of phrases with lowest WER (best matching) to print out during a WER report')
tf.app.flags.DEFINE_string ('summary_dir', '', 'target directory for TensorBoard summaries - defaults to directory "deepspeech/summaries" within user\'s data home specified by the XDG Base Directory Specification')
tf.app.flags.DEFINE_integer ('summary_secs', 0, 'interval in seconds for saving TensorBoard summaries - if 0, no summaries will be written')
# Geometry
tf.app.flags.DEFINE_integer ('n_hidden', 2048, 'layer width to use when initialising layers')
# Initialization
tf.app.flags.DEFINE_integer ('random_seed', 4567, 'default random seed that is used to initialize variables')
tf.app.flags.DEFINE_float ('default_stddev', 0.046875, 'default standard deviation to use when initialising weights and biases')
# Early Stopping
tf.app.flags.DEFINE_boolean ('early_stop', True, 'enable early stopping mechanism over validation dataset. Make sure that dev FLAG is enabled for this to work')
# This parameter is irrespective of the time taken by single epoch to complete and checkpoint saving intervals.
# It is possible that early stopping is triggered far after the best checkpoint is already replaced by checkpoint saving interval mechanism.
# One has to align the parameters (earlystop_nsteps, checkpoint_secs) accordingly as per the time taken by an epoch on different datasets.
tf.app.flags.DEFINE_integer ('earlystop_nsteps', 4, 'number of steps to consider for early stopping. Loss is not stored in the checkpoint so when checkpoint is revived it starts the loss calculation from start at that point')
tf.app.flags.DEFINE_float ('estop_mean_thresh', 0.5, 'mean threshold for loss to determine the condition if early stopping is required')
tf.app.flags.DEFINE_float ('estop_std_thresh', 0.5, 'standard deviation threshold for loss to determine the condition if early stopping is required')
# Decoder
tf.app.flags.DEFINE_string ('decoder_library_path', 'native_client/libctc_decoder_with_kenlm.so', 'path to the libctc_decoder_with_kenlm.so library containing the decoder implementation.')
tf.app.flags.DEFINE_string ('alphabet_config_path', 'data/alphabet.txt', 'path to the configuration file specifying the alphabet used by the network. See the comment in data/alphabet.txt for a description of the format.')
tf.app.flags.DEFINE_string ('lm_binary_path', 'data/lm/lm.binary', 'path to the language model binary file created with KenLM')
tf.app.flags.DEFINE_string ('lm_trie_path', 'data/lm/trie', 'path to the language model trie file created with native_client/generate_trie')
tf.app.flags.DEFINE_integer ('beam_width', 1024, 'beam width used in the CTC decoder when building candidate transcriptions')
tf.app.flags.DEFINE_float ('lm_weight', 1.75, 'the alpha hyperparameter of the CTC decoder. Language Model weight.')
tf.app.flags.DEFINE_float ('word_count_weight', 1.00, 'the beta hyperparameter of the CTC decoder. Word insertion weight (penalty).')
tf.app.flags.DEFINE_float ('valid_word_count_weight', 1.00, 'valid word insertion weight. This is used to lessen the word insertion penalty when the inserted word is part of the vocabulary.')
# Inference mode
tf.app.flags.DEFINE_string ('one_shot_infer', '', 'one-shot inference mode: specify a wav file and the script will load the checkpoint and perform inference on it. Disables training, testing and exporting.')
for var in ['b1', 'h1', 'b2', 'h2', 'b3', 'h3', 'b5', 'h5', 'b6', 'h6']:
tf.app.flags.DEFINE_float('%s_stddev' % var, None, 'standard deviation to use when initialising %s' % var)
FLAGS = tf.app.flags.FLAGS
def initialize_globals():
# ps and worker hosts required for p2p cluster setup
FLAGS.ps_hosts = list(filter(len, FLAGS.ps_hosts.split(',')))
FLAGS.worker_hosts = list(filter(len, FLAGS.worker_hosts.split(',')))
# Determine, if we are the chief worker
global is_chief
is_chief = len(FLAGS.worker_hosts) == 0 or (FLAGS.task_index == 0 and FLAGS.job_name == 'worker')
# Initializing and starting the training coordinator
global COORD
COORD = TrainingCoordinator()
COORD.start()
# The absolute number of computing nodes - regardless of cluster or single mode
global num_workers
num_workers = max(1, len(FLAGS.worker_hosts))
# Create a cluster from the parameter server and worker hosts.
global cluster
cluster = tf.train.ClusterSpec({'ps': FLAGS.ps_hosts, 'worker': FLAGS.worker_hosts})
# If replica numbers are negative, we multiply their absolute values with the number of workers
if FLAGS.replicas < 0:
FLAGS.replicas = num_workers * -FLAGS.replicas
if FLAGS.replicas_to_agg < 0:
FLAGS.replicas_to_agg = num_workers * -FLAGS.replicas_to_agg
# The device path base for this node
global worker_device
worker_device = '/job:%s/task:%d' % (FLAGS.job_name, FLAGS.task_index)
# This node's CPU device
global cpu_device
cpu_device = worker_device + '/cpu:0'
# This node's available GPU devices
global available_devices
available_devices = [worker_device + gpu for gpu in get_available_gpus()]
# If there is no GPU available, we fall back to CPU based operation
if 0 == len(available_devices):
available_devices = [cpu_device]
# Set default dropout rates
if FLAGS.dropout_rate2 < 0:
FLAGS.dropout_rate2 = FLAGS.dropout_rate
if FLAGS.dropout_rate3 < 0:
FLAGS.dropout_rate3 = FLAGS.dropout_rate
if FLAGS.dropout_rate6 < 0:
FLAGS.dropout_rate6 = FLAGS.dropout_rate
global dropout_rates
dropout_rates = [ FLAGS.dropout_rate,
FLAGS.dropout_rate2,
FLAGS.dropout_rate3,
FLAGS.dropout_rate4,
FLAGS.dropout_rate5,
FLAGS.dropout_rate6 ]
global no_dropout
no_dropout = [ 0.0 ] * 6
# Set default checkpoint dir
if len(FLAGS.checkpoint_dir) == 0:
FLAGS.checkpoint_dir = xdg.save_data_path(os.path.join('deepspeech','checkpoints'))
# Set default summary dir
if len(FLAGS.summary_dir) == 0:
FLAGS.summary_dir = xdg.save_data_path(os.path.join('deepspeech','summaries'))
# Standard session configuration that'll be used for all new sessions.
global session_config
session_config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=FLAGS.log_placement)
global alphabet
alphabet = Alphabet(os.path.abspath(FLAGS.alphabet_config_path))
# Geometric Constants
# ===================
# For an explanation of the meaning of the geometric constants, please refer to
# doc/Geometry.md
# Number of MFCC features
global n_input
n_input = 26 # TODO: Determine this programatically from the sample rate
# The number of frames in the context
global n_context
n_context = 9 # TODO: Determine the optimal value using a validation data set
# Number of units in hidden layers
global n_hidden
n_hidden = FLAGS.n_hidden
global n_hidden_1
n_hidden_1 = n_hidden
global n_hidden_2
n_hidden_2 = n_hidden
global n_hidden_5
n_hidden_5 = n_hidden
# LSTM cell state dimension
global n_cell_dim
n_cell_dim = n_hidden
# The number of units in the third layer, which feeds in to the LSTM
global n_hidden_3
n_hidden_3 = 2 * n_cell_dim
# The number of characters in the target language plus one
global n_character
n_character = alphabet.size() + 1 # +1 for CTC blank label
# The number of units in the sixth layer
global n_hidden_6
n_hidden_6 = n_character
# Assign default values for standard deviation
for var in ['b1', 'h1', 'b2', 'h2', 'b3', 'h3', 'b5', 'h5', 'b6', 'h6']:
val = getattr(FLAGS, '%s_stddev' % var)
if val is None:
setattr(FLAGS, '%s_stddev' % var, FLAGS.default_stddev)
# Queues that are used to gracefully stop parameter servers.
# Each queue stands for one ps. A finishing worker sends a token to each queue befor joining/quitting.
# Each ps will dequeue as many tokens as there are workers before joining/quitting.
# This ensures parameter servers won't quit, if still required by at least one worker and
# also won't wait forever (like with a standard `server.join()`).
global done_queues
done_queues = []
for i, ps in enumerate(FLAGS.ps_hosts):
# Queues are hosted by their respective owners
with tf.device('/job:ps/task:%d' % i):
done_queues.append(tf.FIFOQueue(1, tf.int32, shared_name=('queue%i' % i)))
# Placeholder to pass in the worker's index as token
global token_placeholder
token_placeholder = tf.placeholder(tf.int32)
# Enqueue operations for each parameter server
global done_enqueues
done_enqueues = [queue.enqueue(token_placeholder) for queue in done_queues]
# Dequeue operations for each parameter server
global done_dequeues
done_dequeues = [queue.dequeue() for queue in done_queues]
if len(FLAGS.one_shot_infer) > 0:
FLAGS.train = False
FLAGS.test = False
FLAGS.export_dir = ''
if not os.path.exists(FLAGS.one_shot_infer):
log_error('Path specified in --one_shot_infer is not a valid file.')
exit(1)
# Logging functions
# =================
def prefix_print(prefix, message):
print(prefix + ('\n' + prefix).join(message.split('\n')))
def log_debug(message):
if FLAGS.log_level == 0:
prefix_print('D ', message)
def log_traffic(message):
if FLAGS.log_traffic:
log_debug(message)
def log_info(message):
if FLAGS.log_level <= 1:
prefix_print('I ', message)
def log_warn(message):
if FLAGS.log_level <= 2:
prefix_print('W ', message)
def log_error(message):
if FLAGS.log_level <= 3:
prefix_print('E ', message)
# Graph Creation
# ==============
def variable_on_worker_level(name, shape, initializer):
r'''
Next we concern ourselves with graph creation.
However, before we do so we must introduce a utility function ``variable_on_worker_level()``
used to create a variable in CPU memory.
'''
# Use the /cpu:0 device on worker_device for scoped operations
if len(FLAGS.ps_hosts) == 0:
device = worker_device
else:
device = tf.train.replica_device_setter(worker_device=worker_device, cluster=cluster)
with tf.device(device):
# Create or get apropos variable
var = tf.get_variable(name=name, shape=shape, initializer=initializer)
return var
def BiRNN(batch_x, seq_length, dropout):
r'''
That done, we will define the learned variables, the weights and biases,
within the method ``BiRNN()`` which also constructs the neural network.
The variables named ``hn``, where ``n`` is an integer, hold the learned weight variables.
The variables named ``bn``, where ``n`` is an integer, hold the learned bias variables.
In particular, the first variable ``h1`` holds the learned weight matrix that
converts an input vector of dimension ``n_input + 2*n_input*n_context``
to a vector of dimension ``n_hidden_1``.
Similarly, the second variable ``h2`` holds the weight matrix converting
an input vector of dimension ``n_hidden_1`` to one of dimension ``n_hidden_2``.
The variables ``h3``, ``h5``, and ``h6`` are similar.
Likewise, the biases, ``b1``, ``b2``..., hold the biases for the various layers.
'''
# Input shape: [batch_size, n_steps, n_input + 2*n_input*n_context]
batch_x_shape = tf.shape(batch_x)
# Reshaping `batch_x` to a tensor with shape `[n_steps*batch_size, n_input + 2*n_input*n_context]`.
# This is done to prepare the batch for input into the first layer which expects a tensor of rank `2`.
# Permute n_steps and batch_size
batch_x = tf.transpose(batch_x, [1, 0, 2])
# Reshape to prepare input for first layer
batch_x = tf.reshape(batch_x, [-1, n_input + 2*n_input*n_context]) # (n_steps*batch_size, n_input + 2*n_input*n_context)
# The next three blocks will pass `batch_x` through three hidden layers with
# clipped RELU activation and dropout.
# 1st layer
b1 = variable_on_worker_level('b1', [n_hidden_1], tf.random_normal_initializer(stddev=FLAGS.b1_stddev))
h1 = variable_on_worker_level('h1', [n_input + 2*n_input*n_context, n_hidden_1], tf.contrib.layers.xavier_initializer(uniform=False))
layer_1 = tf.minimum(tf.nn.relu(tf.add(tf.matmul(batch_x, h1), b1)), FLAGS.relu_clip)
layer_1 = tf.nn.dropout(layer_1, (1.0 - dropout[0]))
# 2nd layer
b2 = variable_on_worker_level('b2', [n_hidden_2], tf.random_normal_initializer(stddev=FLAGS.b2_stddev))
h2 = variable_on_worker_level('h2', [n_hidden_1, n_hidden_2], tf.random_normal_initializer(stddev=FLAGS.h2_stddev))
layer_2 = tf.minimum(tf.nn.relu(tf.add(tf.matmul(layer_1, h2), b2)), FLAGS.relu_clip)
layer_2 = tf.nn.dropout(layer_2, (1.0 - dropout[1]))
# 3rd layer
b3 = variable_on_worker_level('b3', [n_hidden_3], tf.random_normal_initializer(stddev=FLAGS.b3_stddev))
h3 = variable_on_worker_level('h3', [n_hidden_2, n_hidden_3], tf.random_normal_initializer(stddev=FLAGS.h3_stddev))
layer_3 = tf.minimum(tf.nn.relu(tf.add(tf.matmul(layer_2, h3), b3)), FLAGS.relu_clip)
layer_3 = tf.nn.dropout(layer_3, (1.0 - dropout[2]))
# Now we create the forward and backward LSTM units.
# Both of which have inputs of length `n_cell_dim` and bias `1.0` for the forget gate of the LSTM.
# Forward direction cell: (if else required for TF 1.0 and 1.1 compat)
lstm_fw_cell = tf.contrib.rnn.BasicLSTMCell(n_cell_dim, forget_bias=1.0, state_is_tuple=True) \
if 'reuse' not in inspect.getargspec(tf.contrib.rnn.BasicLSTMCell.__init__).args else \
tf.contrib.rnn.BasicLSTMCell(n_cell_dim, forget_bias=1.0, state_is_tuple=True, reuse=tf.get_variable_scope().reuse)
lstm_fw_cell = tf.contrib.rnn.DropoutWrapper(lstm_fw_cell,
input_keep_prob=1.0 - dropout[3],
output_keep_prob=1.0 - dropout[3],
seed=FLAGS.random_seed)
# Backward direction cell: (if else required for TF 1.0 and 1.1 compat)
lstm_bw_cell = tf.contrib.rnn.BasicLSTMCell(n_cell_dim, forget_bias=1.0, state_is_tuple=True) \
if 'reuse' not in inspect.getargspec(tf.contrib.rnn.BasicLSTMCell.__init__).args else \
tf.contrib.rnn.BasicLSTMCell(n_cell_dim, forget_bias=1.0, state_is_tuple=True, reuse=tf.get_variable_scope().reuse)
lstm_bw_cell = tf.contrib.rnn.DropoutWrapper(lstm_bw_cell,
input_keep_prob=1.0 - dropout[4],
output_keep_prob=1.0 - dropout[4],
seed=FLAGS.random_seed)
# `layer_3` is now reshaped into `[n_steps, batch_size, 2*n_cell_dim]`,
# as the LSTM BRNN expects its input to be of shape `[max_time, batch_size, input_size]`.
layer_3 = tf.reshape(layer_3, [-1, batch_x_shape[0], n_hidden_3])
# Now we feed `layer_3` into the LSTM BRNN cell and obtain the LSTM BRNN output.
outputs, output_states = tf.nn.bidirectional_dynamic_rnn(cell_fw=lstm_fw_cell,
cell_bw=lstm_bw_cell,
inputs=layer_3,
dtype=tf.float32,
time_major=True,
sequence_length=seq_length)
# Reshape outputs from two tensors each of shape [n_steps, batch_size, n_cell_dim]
# to a single tensor of shape [n_steps*batch_size, 2*n_cell_dim]
outputs = tf.concat(outputs, 2)
outputs = tf.reshape(outputs, [-1, 2*n_cell_dim])
# Now we feed `outputs` to the fifth hidden layer with clipped RELU activation and dropout
b5 = variable_on_worker_level('b5', [n_hidden_5], tf.random_normal_initializer(stddev=FLAGS.b5_stddev))
h5 = variable_on_worker_level('h5', [(2 * n_cell_dim), n_hidden_5], tf.random_normal_initializer(stddev=FLAGS.h5_stddev))
layer_5 = tf.minimum(tf.nn.relu(tf.add(tf.matmul(outputs, h5), b5)), FLAGS.relu_clip)
layer_5 = tf.nn.dropout(layer_5, (1.0 - dropout[5]))
# Now we apply the weight matrix `h6` and bias `b6` to the output of `layer_5`
# creating `n_classes` dimensional vectors, the logits.
b6 = variable_on_worker_level('b6', [n_hidden_6], tf.random_normal_initializer(stddev=FLAGS.b6_stddev))
h6 = variable_on_worker_level('h6', [n_hidden_5, n_hidden_6], tf.contrib.layers.xavier_initializer(uniform=False))
layer_6 = tf.add(tf.matmul(layer_5, h6), b6)
# Finally we reshape layer_6 from a tensor of shape [n_steps*batch_size, n_hidden_6]
# to the slightly more useful shape [n_steps, batch_size, n_hidden_6].
# Note, that this differs from the input in that it is time-major.
layer_6 = tf.reshape(layer_6, [-1, batch_x_shape[0], n_hidden_6], name="logits")
# Output shape: [n_steps, batch_size, n_hidden_6]
return layer_6
if not os.path.exists(os.path.abspath(FLAGS.decoder_library_path)):
print('ERROR: The decoder library file does not exist. Make sure you have ' \
'downloaded or built the native client binaries and pass the ' \
'appropriate path to the binaries in the --decoder_library_path parameter.')
custom_op_module = tf.load_op_library(FLAGS.decoder_library_path)
def decode_with_lm(inputs, sequence_length, beam_width=100,
top_paths=1, merge_repeated=True):
decoded_ixs, decoded_vals, decoded_shapes, log_probabilities = (
custom_op_module.ctc_beam_search_decoder_with_lm(
inputs, sequence_length, beam_width=beam_width,
model_path=FLAGS.lm_binary_path, trie_path=FLAGS.lm_trie_path, alphabet_path=FLAGS.alphabet_config_path,
lm_weight=FLAGS.lm_weight, word_count_weight=FLAGS.word_count_weight, valid_word_count_weight=FLAGS.valid_word_count_weight,
top_paths=top_paths, merge_repeated=merge_repeated))
return (
[tf.SparseTensor(ix, val, shape) for (ix, val, shape)
in zip(decoded_ixs, decoded_vals, decoded_shapes)],
log_probabilities)
# Accuracy and Loss
# =================
# In accord with 'Deep Speech: Scaling up end-to-end speech recognition'
# (http://arxiv.org/abs/1412.5567),
# the loss function used by our network should be the CTC loss function
# (http://www.cs.toronto.edu/~graves/preprint.pdf).
# Conveniently, this loss function is implemented in TensorFlow.
# Thus, we can simply make use of this implementation to define our loss.
def calculate_mean_edit_distance_and_loss(model_feeder, tower, dropout):
r'''
This routine beam search decodes a mini-batch and calculates the loss and mean edit distance.
Next to total and average loss it returns the mean edit distance,
the decoded result and the batch's original Y.
'''
# Obtain the next batch of data
batch_x, batch_seq_len, batch_y = model_feeder.next_batch(tower)
# Calculate the logits of the batch using BiRNN
logits = BiRNN(batch_x, tf.to_int64(batch_seq_len), dropout)
# Compute the CTC loss using either TensorFlow's `ctc_loss` or Baidu's `warp_ctc_loss`.
if FLAGS.use_warpctc:
total_loss = tf.contrib.warpctc.warp_ctc_loss(labels=batch_y, inputs=logits, sequence_length=batch_seq_len)
else:
total_loss = tf.nn.ctc_loss(labels=batch_y, inputs=logits, sequence_length=batch_seq_len)
# Calculate the average loss across the batch
avg_loss = tf.reduce_mean(total_loss)
# Beam search decode the batch
decoded, _ = decode_with_lm(logits, batch_seq_len, merge_repeated=False, beam_width=FLAGS.beam_width)
# Compute the edit (Levenshtein) distance
distance = tf.edit_distance(tf.cast(decoded[0], tf.int32), batch_y)
# Compute the mean edit distance
mean_edit_distance = tf.reduce_mean(distance)
# Finally we return the
# - calculated total and
# - average losses,
# - the Levenshtein distance,
# - the recognition mean edit distance,
# - the decoded batch and
# - the original batch_y (which contains the verified transcriptions).
return total_loss, avg_loss, distance, mean_edit_distance, decoded, batch_y
# Adam Optimization
# =================
# In constrast to 'Deep Speech: Scaling up end-to-end speech recognition'
# (http://arxiv.org/abs/1412.5567),
# in which 'Nesterov's Accelerated Gradient Descent'
# (www.cs.toronto.edu/~fritz/absps/momentum.pdf) was used,
# we will use the Adam method for optimization (http://arxiv.org/abs/1412.6980),
# because, generally, it requires less fine-tuning.
def create_optimizer():
optimizer = tf.train.AdamOptimizer(learning_rate=FLAGS.learning_rate,
beta1=FLAGS.beta1,
beta2=FLAGS.beta2,
epsilon=FLAGS.epsilon)
return optimizer
# Towers
# ======
# In order to properly make use of multiple GPU's, one must introduce new abstractions,
# not present when using a single GPU, that facilitate the multi-GPU use case.
# In particular, one must introduce a means to isolate the inference and gradient
# calculations on the various GPU's.
# The abstraction we intoduce for this purpose is called a 'tower'.
# A tower is specified by two properties:
# * **Scope** - A scope, as provided by `tf.name_scope()`,
# is a means to isolate the operations within a tower.
# For example, all operations within 'tower 0' could have their name prefixed with `tower_0/`.
# * **Device** - A hardware device, as provided by `tf.device()`,
# on which all operations within the tower execute.
# For example, all operations of 'tower 0' could execute on the first GPU `tf.device('/gpu:0')`.
def get_tower_results(model_feeder, optimizer):
r'''
With this preliminary step out of the way, we can for each GPU introduce a
tower for which's batch we calculate
* The CTC decodings ``decoded``,
* The (total) loss against the outcome (Y) ``total_loss``,
* The loss averaged over the whole batch ``avg_loss``,
* The optimization gradient (computed based on the averaged loss),
* The Levenshtein distances between the decodings and their transcriptions ``distance``,
* The mean edit distance of the outcome averaged over the whole batch ``mean_edit_distance``
and retain the original ``labels`` (Y).
``decoded``, ``labels``, the optimization gradient, ``distance``, ``mean_edit_distance``,
``total_loss`` and ``avg_loss`` are collected into the corresponding arrays
``tower_decodings``, ``tower_labels``, ``tower_gradients``, ``tower_distances``,
``tower_mean_edit_distances``, ``tower_total_losses``, ``tower_avg_losses`` (dimension 0 being the tower).
Finally this new method ``get_tower_results()`` will return those tower arrays.
In case of ``tower_mean_edit_distances`` and ``tower_avg_losses``, it will return the
averaged values instead of the arrays.
'''
# Tower labels to return
tower_labels = []
# Tower decodings to return
tower_decodings = []
# Tower distances to return
tower_distances = []
# Tower total batch losses to return
tower_total_losses = []
# Tower gradients to return
tower_gradients = []
# To calculate the mean of the mean edit distances
tower_mean_edit_distances = []
# To calculate the mean of the losses
tower_avg_losses = []
with tf.variable_scope(tf.get_variable_scope()):
# Loop over available_devices
for i in range(len(available_devices)):
# Execute operations of tower i on device i
if len(FLAGS.ps_hosts) == 0:
device = available_devices[i]
else:
device = tf.train.replica_device_setter(worker_device=available_devices[i], cluster=cluster)
with tf.device(device):
# Create a scope for all operations of tower i
with tf.name_scope('tower_%d' % i) as scope:
# Calculate the avg_loss and mean_edit_distance and retrieve the decoded
# batch along with the original batch's labels (Y) of this tower
total_loss, avg_loss, distance, mean_edit_distance, decoded, labels = \
calculate_mean_edit_distance_and_loss(model_feeder, i, no_dropout if optimizer is None else dropout_rates)
# Allow for variables to be re-used by the next tower
tf.get_variable_scope().reuse_variables()
# Retain tower's labels (Y)
tower_labels.append(labels)
# Retain tower's decoded batch
tower_decodings.append(decoded)
# Retain tower's distances
tower_distances.append(distance)
# Retain tower's total losses
tower_total_losses.append(total_loss)
# Compute gradients for model parameters using tower's mini-batch
gradients = optimizer.compute_gradients(avg_loss)
# Retain tower's gradients
tower_gradients.append(gradients)
# Retain tower's mean edit distance
tower_mean_edit_distances.append(mean_edit_distance)
# Retain tower's avg losses
tower_avg_losses.append(avg_loss)
# Return the results tuple, the gradients, and the means of mean edit distances and losses
return (tower_labels, tower_decodings, tower_distances, tower_total_losses), \
tower_gradients, \
tf.reduce_mean(tower_mean_edit_distances, 0), \
tf.reduce_mean(tower_avg_losses, 0)
def average_gradients(tower_gradients):
r'''
A routine for computing each variable's average of the gradients obtained from the GPUs.
Note also that this code acts as a syncronization point as it requires all
GPUs to be finished with their mini-batch before it can run to completion.
'''
# List of average gradients to return to the caller
average_grads = []
# Run this on cpu_device to conserve GPU memory
with tf.device(cpu_device):
# Loop over gradient/variable pairs from all towers
for grad_and_vars in zip(*tower_gradients):
# Introduce grads to store the gradients for the current variable
grads = []
# Loop over the gradients for the current variable
for g, _ in grad_and_vars:
# Add 0 dimension to the gradients to represent the tower.
expanded_g = tf.expand_dims(g, 0)
# Append on a 'tower' dimension which we will average over below.
grads.append(expanded_g)
# Average over the 'tower' dimension
grad = tf.concat(grads, 0)
grad = tf.reduce_mean(grad, 0)
# Create a gradient/variable tuple for the current variable with its average gradient
grad_and_var = (grad, grad_and_vars[0][1])
# Add the current tuple to average_grads
average_grads.append(grad_and_var)
# Return result to caller
return average_grads
# Logging
# =======
def log_variable(variable, gradient=None):
r'''
We introduce a function for logging a tensor variable's current state.
It logs scalar values for the mean, standard deviation, minimum and maximum.
Furthermore it logs a histogram of its state and (if given) of an optimization gradient.
'''
name = variable.name
mean = tf.reduce_mean(variable)
tf.summary.scalar(name='%s/mean' % name, tensor=mean)
tf.summary.scalar(name='%s/sttdev' % name, tensor=tf.sqrt(tf.reduce_mean(tf.square(variable - mean))))
tf.summary.scalar(name='%s/max' % name, tensor=tf.reduce_max(variable))
tf.summary.scalar(name='%s/min' % name, tensor=tf.reduce_min(variable))
tf.summary.histogram(name=name, values=variable)
if gradient is not None:
if isinstance(gradient, tf.IndexedSlices):
grad_values = gradient.values
else:
grad_values = gradient
if grad_values is not None:
tf.summary.histogram(name='%s/gradients' % name, values=grad_values)
def log_grads_and_vars(grads_and_vars):
r'''
Let's also introduce a helper function for logging collections of gradient/variable tuples.
'''
for gradient, variable in grads_and_vars:
log_variable(variable, gradient=gradient)
def get_git_revision_hash():
return subprocess.check_output(['git', 'rev-parse', 'HEAD']).strip()
def get_git_branch():
return subprocess.check_output(['git', 'rev-parse', '--abbrev-ref', 'HEAD']).strip()
# Helpers
# =======
def calculate_report(results_tuple):
r'''
This routine will calculate a WER report.
It'll compute the `mean` WER and create ``Sample`` objects of the ``report_count`` top lowest
loss items from the provided WER results tuple (only items with WER!=0 and ordered by their WER).
'''
samples = []
items = list(zip(*results_tuple))
mean_wer = 0.0
for label, decoding, distance, loss in items:
sample_wer = wer(label, decoding)
sample = Sample(label, decoding, loss, distance, sample_wer)
samples.append(sample)
mean_wer += sample_wer
# Getting the mean WER from the accumulated one
mean_wer = mean_wer / len(items)
# Filter out all items with WER=0
samples = [s for s in samples if s.wer > 0]
# Order the remaining items by their loss (lowest loss on top)
samples.sort(key=lambda s: s.loss)
# Take only the first report_count items
samples = samples[:FLAGS.report_count]
# Order this top FLAGS.report_count items by their WER (lowest WER on top)
samples.sort(key=lambda s: s.wer)
return mean_wer, samples
def collect_results(results_tuple, returns):
r'''
This routine will help collecting partial results for the WER reports.
The ``results_tuple`` is composed of an array of the original labels,
an array of the corresponding decodings, an array of the corrsponding
distances and an array of the corresponding losses. ``returns`` is built up
in a similar way, containing just the unprocessed results of one
``session.run`` call (effectively of one batch).
Labels and decodings are converted to text before splicing them into their
corresponding results_tuple lists. In the case of decodings,
for now we just pick the first available path.
'''
# Each of the arrays within results_tuple will get extended by a batch of each available device
for i in range(len(available_devices)):
# Collect the labels
results_tuple[0].extend(sparse_tensor_value_to_texts(returns[0][i], alphabet))
# Collect the decodings - at the moment we default to the first one
results_tuple[1].extend(sparse_tensor_value_to_texts(returns[1][i][0], alphabet))
# Collect the distances
results_tuple[2].extend(returns[2][i])
# Collect the losses
results_tuple[3].extend(returns[3][i])
# For reporting we also need a standard way to do time measurements.
def stopwatch(start_duration=0):
r'''
This function will toggle a stopwatch.
The first call starts it, second call stops it, third call continues it etc.
So if you want to measure the accumulated time spent in a certain area of the code,
you can surround that code by stopwatch-calls like this:
.. code:: python
fun_time = 0 # initializes a stopwatch
[...]
for i in range(10):
[...]
# Starts/continues the stopwatch - fun_time is now a point in time (again)
fun_time = stopwatch(fun_time)
fun()
# Pauses the stopwatch - fun_time is now a duration
fun_time = stopwatch(fun_time)
[...]
# The following line only makes sense after an even call of :code:`fun_time = stopwatch(fun_time)`.
print 'Time spent in fun():', format_duration(fun_time)
'''
if start_duration == 0:
return datetime.datetime.utcnow()
else:
return datetime.datetime.utcnow() - start_duration
def format_duration(duration):
'''Formats the result of an even stopwatch call as hours:minutes:seconds'''
duration = duration if isinstance(duration, int) else duration.seconds
m, s = divmod(duration, 60)
h, m = divmod(m, 60)
return '%d:%02d:%02d' % (h, m, s)
# Execution
# =========
# String constants for different services of the web handler
PREFIX_NEXT_INDEX = '/next_index_'
PREFIX_GET_JOB = '/get_job_'
# Global ID counter for all objects requiring an ID
id_counter = 0
def new_id():
'''Returns a new ID that is unique on process level. Not thread-safe.
Returns:
int. The new ID
'''
global id_counter
id_counter += 1
return id_counter
class Sample(object):
def __init__(self, src, res, loss, mean_edit_distance, sample_wer):
'''Represents one item of a WER report.
Args:
src (str): source text
res (str): resulting text
loss (float): computed loss of this item
mean_edit_distance (float): computed mean edit distance of this item
'''
self.src = src
self.res = res
self.loss = loss
self.mean_edit_distance = mean_edit_distance
self.wer = sample_wer
def __str__(self):
return 'WER: %f, loss: %f, mean edit distance: %f\n - src: "%s"\n - res: "%s"' % (self.wer, self.loss, self.mean_edit_distance, self.src, self.res)
class WorkerJob(object):
def __init__(self, epoch_id, index, set_name, steps, report):
'''Represents a job that should be executed by a worker.
Args:
epoch_id (int): the ID of the 'parent' epoch
index (int): the epoch index of the 'parent' epoch
set_name (str): the name of the data-set - one of 'train', 'dev', 'test'
steps (int): the number of `session.run` calls
report (bool): if this job should produce a WER report
'''
self.id = new_id()
self.epoch_id = epoch_id
self.index = index
self.worker = -1
self.set_name = set_name
self.steps = steps
self.report = report
self.loss = -1
self.mean_edit_distance = -1
self.wer = -1
self.samples = []
def __str__(self):
return 'Job (ID: %d, worker: %d, epoch: %d, set_name: %s)' % (self.id, self.worker, self.index, self.set_name)
class Epoch(object):
'''Represents an epoch that should be executed by the Training Coordinator.
Creates `num_jobs` `WorkerJob` instances in state 'open'.
Args:
index (int): the epoch index of the 'parent' epoch
num_jobs (int): the number of jobs in this epoch
Kwargs:
set_name (str): the name of the data-set - one of 'train', 'dev', 'test'
report (bool): if this job should produce a WER report
'''
def __init__(self, index, num_jobs, set_name='train', report=False):
self.id = new_id()
self.index = index
self.num_jobs = num_jobs
self.set_name = set_name
self.report = report
self.wer = -1
self.loss = -1
self.mean_edit_distance = -1
self.jobs_open = []
self.jobs_running = []
self.jobs_done = []
self.samples = []
for i in range(self.num_jobs):
self.jobs_open.append(WorkerJob(self.id, self.index, self.set_name, FLAGS.iters_per_worker, self.report))
def name(self):
'''Gets a printable name for this epoch.
Returns:
str. printable name for this epoch
'''
if self.index >= 0:
ename = ' of Epoch %d' % self.index
else:
ename = ''
if self.set_name == 'train':
return 'Training%s' % ename
elif self.set_name == 'dev':
return 'Validation%s' % ename
else:
return 'Test%s' % ename
def get_job(self, worker):
'''Gets the next open job from this epoch. The job will be marked as 'running'.
Args:
worker (int): index of the worker that takes the job
Returns:
WorkerJob. job that has been marked as running for this worker
'''
if len(self.jobs_open) > 0:
job = self.jobs_open.pop(0)
self.jobs_running.append(job)
job.worker = worker
return job
else:
return None
def finish_job(self, job):
'''Finishes a running job. Removes it from the running jobs list and adds it to the done jobs list.
Args:
job (WorkerJob): the job to put into state 'done'
'''
index = next((i for i in range(len(self.jobs_running)) if self.jobs_running[i].id == job.id), -1)
if index >= 0:
self.jobs_running.pop(index)
self.jobs_done.append(job)
log_traffic('%s - Moved %s from running to done.' % (self.name(), job))
else:
log_warn('%s - There is no job with ID %d registered as running.' % (self.name(), job.id))
def done(self):
'''Checks, if all jobs of the epoch are in state 'done'.
It also lazy-prepares a WER report from the result data of all jobs.
Returns:
bool. if all jobs of the epoch are 'done'
'''
if len(self.jobs_open) == 0 and len(self.jobs_running) == 0:
num_jobs = len(self.jobs_done)