-
Notifications
You must be signed in to change notification settings - Fork 19.5k
/
training.py
2146 lines (1919 loc) · 90.5 KB
/
training.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
# -*- coding: utf-8 -*-
from __future__ import print_function
from __future__ import absolute_import
import warnings
import copy
import time
import numpy as np
import multiprocessing
import threading
import six
try:
import queue
except ImportError:
import Queue as queue
from .topology import Container
from .. import backend as K
from .. import optimizers
from .. import losses
from .. import metrics as metrics_module
from ..utils.generic_utils import Progbar
from .. import callbacks as cbks
from ..legacy import interfaces
def _standardize_input_data(data, names, shapes=None,
check_batch_axis=True,
exception_prefix=''):
"""Normalizes inputs and targets provided by users.
Users may pass data as a list of arrays, dictionary of arrays,
or as a single array. We normalize this to an ordered list of
arrays (same order as `names`), while checking that the provided
arrays have shapes that match the network's expectations.
# Arguments
data: User-provided input data (polymorphic).
names: List of expected array names.
shapes: Optional list of expected array shapes.
check_batch_axis: Boolean; whether to check that
the batch axis of the arrays matches the expected
value found in `shapes`.
exception_prefix: String prefix used for exception formatting.
# Returns
List of standardized input arrays (one array per model input).
# Raises
ValueError: in case of improperly formatted user-provided data.
"""
if not names:
return []
if data is None:
return [None for _ in range(len(names))]
if isinstance(data, dict):
arrays = []
for name in names:
if name not in data:
raise ValueError('No data provided for "' +
name + '". Need data for each key in: ' +
str(names))
arrays.append(data[name])
elif isinstance(data, list):
if len(data) != len(names):
if data and hasattr(data[0], 'shape'):
raise ValueError('Error when checking model ' +
exception_prefix +
': the list of Numpy arrays '
'that you are passing to your model '
'is not the size the model expected. '
'Expected to see ' + str(len(names)) +
' arrays but instead got '
'the following list of ' + str(len(data)) +
' arrays: ' + str(data)[:200] +
'...')
else:
if len(names) == 1:
data = [np.asarray(data)]
else:
raise ValueError(
'Error when checking model ' +
exception_prefix +
': you are passing a list as '
'input to your model, '
'but the model expects '
'a list of ' + str(len(names)) +
' Numpy arrays instead. '
'The list you passed was: ' +
str(data)[:200])
arrays = data
else:
if not hasattr(data, 'shape'):
raise TypeError('Error when checking model ' +
exception_prefix +
': data should be a Numpy array, '
'or list/dict of Numpy arrays. '
'Found: ' + str(data)[:200] + '...')
if len(names) > 1:
# Case: model expects multiple inputs but only received
# a single Numpy array.
raise ValueError('The model expects ' + str(len(names)) +
exception_prefix +
' arrays, but only received one array. '
'Found: array with shape ' + str(data.shape))
arrays = [data]
# Make arrays at least 2D.
for i in range(len(names)):
array = arrays[i]
if len(array.shape) == 1:
array = np.expand_dims(array, 1)
arrays[i] = array
# Check shapes compatibility.
if shapes:
for i in range(len(names)):
if shapes[i] is None:
continue
array = arrays[i]
if len(array.shape) != len(shapes[i]):
raise ValueError('Error when checking ' + exception_prefix +
': expected ' + names[i] +
' to have ' + str(len(shapes[i])) +
' dimensions, but got array with shape ' +
str(array.shape))
for j, (dim, ref_dim) in enumerate(zip(array.shape, shapes[i])):
if not j and not check_batch_axis:
# skip the first axis
continue
if ref_dim:
if ref_dim != dim:
raise ValueError(
'Error when checking ' + exception_prefix +
': expected ' + names[i] +
' to have shape ' + str(shapes[i]) +
' but got array with shape ' +
str(array.shape))
return arrays
def _standardize_sample_or_class_weights(x_weight, output_names, weight_type):
"""Maps `sample_weight` or `class_weight` to model outputs.
# Arguments
x_weight: User-provided `sample_weight` or `class_weight` argument.
output_names: List of output names (strings) in the model.
weight_type: A string used purely for exception printing.
# Returns
A list of `sample_weight` or `class_weight` where there are exactly
one element per model output.
# Raises
ValueError: In case of invalid user-provided argument.
"""
if x_weight is None or len(x_weight) == 0:
return [None for _ in output_names]
if len(output_names) == 1:
if isinstance(x_weight, list) and len(x_weight) == 1:
return x_weight
if isinstance(x_weight, dict) and output_names[0] in x_weight:
return [x_weight[output_names[0]]]
else:
return [x_weight]
if isinstance(x_weight, list):
if len(x_weight) != len(output_names):
raise ValueError('Provided `' + weight_type + '` was a list of ' +
str(len(x_weight)) +
' elements, but the model has ' +
str(len(output_names)) + ' outputs. '
'You should provide one `' + weight_type + '`'
'array per model output.')
return x_weight
if isinstance(x_weight, dict):
x_weights = []
for name in output_names:
x_weights.append(x_weight.get(name))
return x_weights
else:
raise TypeError('The model has multiple outputs, so `' +
weight_type + '` '
'should be either a list of a dict. '
'Provided `' + weight_type +
'` type not understood: ' +
str(x_weight))
def _standardize_class_weights(class_weight, output_names):
return _standardize_sample_or_class_weights(class_weight,
output_names,
'class_weight')
def _standardize_sample_weights(sample_weight, output_names):
return _standardize_sample_or_class_weights(sample_weight,
output_names,
'sample_weight')
def _check_array_lengths(inputs, targets, weights):
"""Does user input validation for numpy arrays.
# Arguments
inputs: list of Numpy arrays of inputs.
targets: list of Numpy arrays of targets.
weights: list of Numpy arrays of sample weights.
# Raises
ValueError: in case of incorrectly formatted data.
"""
x_lengths = [x.shape[0] for x in inputs]
y_lengths = [y.shape[0] for y in targets]
w_lengths = [w.shape[0] for w in weights]
set_x = set(x_lengths)
if len(set_x) > 1:
raise ValueError('All input arrays (x) should have '
'the same number of samples. Got array shapes: ' +
str([x.shape for x in inputs]))
set_y = set(y_lengths)
if len(set_y) > 1:
raise ValueError('All target arrays (y) should have '
'the same number of samples. Got array shapes: ' +
str([y.shape for y in targets]))
set_w = set(w_lengths)
if len(set_w) > 1:
raise ValueError('All sample_weight arrays should have '
'the same number of samples. Got array shapes: ' +
str([w.shape for w in weights]))
if set_x and set_y and list(set_x)[0] != list(set_y)[0]:
raise ValueError('Input arrays should have '
'the same number of samples as target arrays. '
'Found ' + str(list(set_x)[0]) + ' input samples '
'and ' + str(list(set_y)[0]) + ' target samples.')
if set_y and set_w and list(set_y)[0] != list(set_w)[0]:
raise ValueError('Sample_weight arrays should have '
'the same number of samples as target arrays. Got ' +
str(list(set_y)[0]) + ' input samples and ' +
str(list(set_w)[0]) + ' target samples.')
def _check_loss_and_target_compatibility(targets, loss_fns, output_shapes):
"""Does validation on the compatibility of targets and loss functions.
This helps prevent users from using loss functions incorrectly.
# Arguments
targets: list of Numpy arrays of targets.
loss_fns: list of loss functions.
output_shapes: list of shapes of model outputs.
# Raises
ValueError: if a loss function or target array
is incompatible with an output.
"""
key_losses = {'mean_square_error',
'binary_crossentropy',
'categorical_crossentropy'}
for y, loss, shape in zip(targets, loss_fns, output_shapes):
if loss is None:
continue
if loss.__name__ == 'categorical_crossentropy':
if y.shape[-1] == 1:
raise ValueError(
'You are passing a target array of shape ' + str(y.shape) +
' while using as loss `categorical_crossentropy`. '
'`categorical_crossentropy` expects '
'targets to be binary matrices (1s and 0s) '
'of shape (samples, classes). '
'If your targets are integer classes, '
'you can convert them to the expected format via:\n'
'```\n'
'from keras.utils.np_utils import to_categorical\n'
'y_binary = to_categorical(y_int)\n'
'```\n'
'\n'
'Alternatively, you can use the loss function '
'`sparse_categorical_crossentropy` instead, '
'which does expect integer targets.')
if loss.__name__ in key_losses:
for target_dim, out_dim in zip(y.shape[1:], shape[1:]):
if out_dim is not None and target_dim != out_dim:
raise ValueError(
'A target array with shape ' + str(y.shape) +
' was passed for an output of shape ' + str(shape) +
' while using as loss `' + loss.__name__ + '`. '
'This loss expects '
'targets to have the same shape '
'as the output.')
def _collect_metrics(metrics, output_names):
"""Maps metric functions to model outputs.
# Arguments
metrics: a list or dict of metric functions.
output_names: a list of the names (strings) of model outputs.
# Returns
A list (one entry per model output) of lists of metric functions.
For instance, if the model has 2 outputs, and for the first output
we want to compute "binary_accuracy" and "binary_crossentropy",
and just "binary_accuracy" for the second output,
the list would look like:
`[[binary_accuracy, binary_crossentropy], [binary_accuracy]]`
# Raises
TypeError: if an incorrect type is passed for the `metrics` argument.
"""
if not metrics:
return [[] for _ in output_names]
if isinstance(metrics, list):
# we then apply all metrics to all outputs.
return [copy.copy(metrics) for _ in output_names]
elif isinstance(metrics, dict):
nested_metrics = []
for name in output_names:
output_metrics = metrics.get(name, [])
if not isinstance(output_metrics, list):
output_metrics = [output_metrics]
nested_metrics.append(output_metrics)
return nested_metrics
else:
raise TypeError('Type of `metrics` argument not understood. '
'Expected a list or dictionary, found: ' +
str(metrics))
def _batch_shuffle(index_array, batch_size):
"""Shuffles an array in a batch-wise fashion.
Useful for shuffling HDF5 arrays
(where one cannot access arbitrary indices).
# Arguments
index_array: array of indices to be shuffled.
batch_size: integer.
# Returns
The `index_array` array, shuffled in a batch-wise fashion.
"""
batch_count = int(len(index_array) / batch_size)
# to reshape we need to be cleanly divisible by batch size
# we stash extra items and reappend them after shuffling
last_batch = index_array[batch_count * batch_size:]
index_array = index_array[:batch_count * batch_size]
index_array = index_array.reshape((batch_count, batch_size))
np.random.shuffle(index_array)
index_array = index_array.flatten()
return np.append(index_array, last_batch)
def _make_batches(size, batch_size):
"""Returns a list of batch indices (tuples of indices).
# Arguments
size: Integer, total size of the data to slice into batches.
batch_size: Integer, batch size.
# Returns
A list of tuples of array indices.
"""
num_batches = int(np.ceil(size / float(batch_size)))
return [(i * batch_size, min(size, (i + 1) * batch_size))
for i in range(0, num_batches)]
def _slice_arrays(arrays, start=None, stop=None):
"""Slice an array or list of arrays.
This takes an array-like, or a list of
array-likes, and outputs:
- arrays[start:stop] if `arrays` is an array-like
- [x[start:stop] for x in arrays] if `arrays` is a list
Can also work on list/array of indices: `_slice_arrays(x, indices)`
# Arguments
arrays: Single array or list of arrays.
start: can be an integer index (start index)
or a list/array of indices
stop: integer (stop index); should be None if
`start` was a list.
# Returns
A slice of the array(s).
"""
if isinstance(arrays, list):
if hasattr(start, '__len__'):
# hdf5 datasets only support list objects as indices
if hasattr(start, 'shape'):
start = start.tolist()
return [x[start] for x in arrays]
else:
return [x[start:stop] for x in arrays]
else:
if hasattr(start, '__len__'):
if hasattr(start, 'shape'):
start = start.tolist()
return arrays[start]
else:
return arrays[start:stop]
def _weighted_masked_objective(fn):
"""Adds support for masking and sample-weighting to an objective function.
It transforms an objective function `fn(y_true, y_pred)`
into a sample-weighted, cost-masked objective function
`fn(y_true, y_pred, weights, mask)`.
# Arguments
fn: The objective function to wrap,
with signature `fn(y_true, y_pred)`.
# Returns
A function with signature `fn(y_true, y_pred, weights, mask)`.
"""
if fn is None:
return None
def weighted(y_true, y_pred, weights, mask=None):
"""Wrapper function.
# Arguments
y_true: `y_true` argument of `fn`.
y_pred: `y_pred` argument of `fn`.
weights: Weights tensor.
mask: Mask tensor.
# Returns
Scalar tensor.
"""
# score_array has ndim >= 2
score_array = fn(y_true, y_pred)
if mask is not None:
# Cast the mask to floatX to avoid float64 upcasting in theano
mask = K.cast(mask, K.floatx())
# mask should have the same shape as score_array
score_array *= mask
# the loss per batch should be proportional
# to the number of unmasked samples.
score_array /= K.mean(mask)
# reduce score_array to same ndim as weight array
ndim = K.ndim(score_array)
weight_ndim = K.ndim(weights)
score_array = K.mean(score_array, axis=list(range(weight_ndim, ndim)))
# apply sample weighting
if weights is not None:
score_array *= weights
score_array /= K.mean(K.cast(K.not_equal(weights, 0), K.floatx()))
return K.mean(score_array)
return weighted
def _masked_objective(fn):
"""Adds support for masking to an objective function.
It transforms an objective function `fn(y_true, y_pred)`
into a cost-masked objective function
`fn(y_true, y_pred, mask)`.
# Arguments
fn: The objective function to wrap,
with signature `fn(y_true, y_pred)`.
# Returns
A function with signature `fn(y_true, y_pred, mask)`.
"""
def masked(y_true, y_pred, mask=None):
"""Wrapper function.
# Arguments
y_true: `y_true` argument of `fn`.
y_pred: `y_pred` argument of `fn`.
mask: Mask tensor.
# Returns
Scalar tensor.
"""
# score_array has ndim >= 2
score_array = fn(y_true, y_pred)
if mask is not None:
# Cast the mask to floatX to avoid float64 upcasting in theano
mask = K.cast(mask, K.floatx())
# mask should have the same shape as score_array
score_array *= mask
# the loss per batch should be proportional
# to the number of unmasked samples.
score_array /= K.mean(mask)
return K.mean(score_array)
return masked
def _standardize_weights(y, sample_weight=None, class_weight=None,
sample_weight_mode=None):
"""Performs sample weight validation and standardization.
Everything gets normalized to a single sample-wise (or timestep-wise)
weight array.
# Arguments
y: Numpy array of model targets to be weighted.
sample_weight: User-provided `sample_weight` argument.
class_weight: User-provided `class_weight` argument.
sample_weight_mode: One of `None` or `"temporal"`.
`"temporal"` indicated that we expect 2D weight data
that will be applied to the last 2 dimensions of
the targets (i.e. we are weighting timesteps, not samples).
# Returns
A numpy array of target weights, one entry per sample to weight.
# Raises
ValueError: In case of invalid user-provided arguments.
"""
if sample_weight_mode is not None:
if sample_weight_mode != 'temporal':
raise ValueError('"sample_weight_mode '
'should be None or "temporal". '
'Found: ' + str(sample_weight_mode))
if len(y.shape) < 3:
raise ValueError('Found a sample_weight array for '
'an input with shape ' +
str(y.shape) + '. '
'Timestep-wise sample weighting (use of '
'sample_weight_mode="temporal") is restricted to '
'outputs that are at least 3D, i.e. that have '
'a time dimension.')
if sample_weight is not None and len(sample_weight.shape) != 2:
raise ValueError('Found a sample_weight array with shape ' +
str(sample_weight.shape) + '. '
'In order to use timestep-wise sample weighting, '
'you should pass a 2D sample_weight array.')
else:
if sample_weight is not None and len(sample_weight.shape) != 1:
raise ValueError('Found a sample_weight array with shape ' +
str(sample_weight.shape) + '. '
'In order to use timestep-wise sample weights, '
'you should specify '
'sample_weight_mode="temporal" '
'in compile(). If you just mean to use '
'sample-wise weights, make sure your '
'sample_weight array is 1D.')
if sample_weight is not None:
if len(sample_weight.shape) > len(y.shape):
raise ValueError('Found a sample_weight with shape' +
str(sample_weight.shape) + '.'
'Expected sample_weight with rank '
'less than or equal to ' + str(len(y.shape)))
if y.shape[:sample_weight.ndim] != sample_weight.shape:
raise ValueError('Found a sample_weight array with shape ' +
str(sample_weight.shape) + ' for an input with shape ' +
str(y.shape) + '. '
'sample_weight cannot be broadcast.')
return sample_weight
elif isinstance(class_weight, dict):
if len(y.shape) > 2:
raise ValueError('class_weight not supported for '
'3+ dimensional targets.')
if y.shape[1] > 1:
y_classes = y.argmax(axis=1)
elif y.shape[1] == 1:
y_classes = np.reshape(y, y.shape[0])
else:
y_classes = y
weights = np.asarray([class_weight[cls] for cls in y_classes])
return weights
else:
if sample_weight_mode is None:
return np.ones((y.shape[0],), dtype=K.floatx())
else:
return np.ones((y.shape[0], y.shape[1]), dtype=K.floatx())
class GeneratorEnqueuer(object):
"""Builds a queue out of a data generator.
Used in `fit_generator`, `evaluate_generator`, `predict_generator`.
# Arguments
generator: a generator function which endlessly yields data
pickle_safe: use multiprocessing if True, otherwise threading
"""
def __init__(self, generator, pickle_safe=False):
self._generator = generator
self._pickle_safe = pickle_safe
self._threads = []
self._stop_event = None
self.queue = None
def start(self, workers=1, max_q_size=10, wait_time=0.05):
"""Kicks off threads which add data from the generator into the queue.
# Arguments
workers: number of worker threads
max_q_size: queue size (when full, threads could block on put())
wait_time: time to sleep in-between calls to put()
"""
def data_generator_task():
while not self._stop_event.is_set():
try:
if self._pickle_safe or self.queue.qsize() < max_q_size:
generator_output = next(self._generator)
self.queue.put(generator_output)
else:
time.sleep(wait_time)
except Exception:
self._stop_event.set()
raise
try:
if self._pickle_safe:
self.queue = multiprocessing.Queue(maxsize=max_q_size)
self._stop_event = multiprocessing.Event()
else:
self.queue = queue.Queue()
self._stop_event = threading.Event()
for _ in range(workers):
if self._pickle_safe:
# Reset random seed else all children processes
# share the same seed
np.random.seed()
thread = multiprocessing.Process(target=data_generator_task)
thread.daemon = True
else:
thread = threading.Thread(target=data_generator_task)
self._threads.append(thread)
thread.start()
except:
self.stop()
raise
def is_running(self):
return self._stop_event is not None and not self._stop_event.is_set()
def stop(self, timeout=None):
"""Stop running threads and wait for them to exit, if necessary.
Should be called by the same thread which called start().
# Arguments
timeout: maximum time to wait on thread.join()
"""
if self.is_running():
self._stop_event.set()
for thread in self._threads:
if thread.is_alive():
if self._pickle_safe:
thread.terminate()
else:
thread.join(timeout)
if self._pickle_safe:
if self.queue is not None:
self.queue.close()
self._threads = []
self._stop_event = None
self.queue = None
class Model(Container):
"""The `Model` class adds training & evaluation routines to a `Container`.
"""
def compile(self, optimizer, loss, metrics=None, loss_weights=None,
sample_weight_mode=None, **kwargs):
"""Configures the model for training.
# Arguments
optimizer: str (name of optimizer) or optimizer object.
See [optimizers](/optimizers).
loss: str (name of objective function) or objective function.
See [losses](/losses).
If the model has multiple outputs, you can use a different loss
on each output by passing a dictionary or a list of losses.
The loss value that will be minimized by the model
will then be the sum of all individual losses.
metrics: list of metrics to be evaluated by the model
during training and testing.
Typically you will use `metrics=['accuracy']`.
To specify different metrics for different outputs of a
multi-output model, you could also pass a dictionary,
such as `metrics={'output_a': 'accuracy'}`.
loss_weights: Optional list or dictionary specifying scalar
coefficients (Python floats) to weight the loss contributions
of different model outputs.
The loss value that will be minimized by the model
will then be the *weighted sum* of all individual losses,
weighted by the `loss_weights` coefficients.
If a list, it is expected to have a 1:1 mapping
to the model's outputs. If a tensor, it is expected to map
output names (strings) to scalar coefficients.
sample_weight_mode: if you need to do timestep-wise
sample weighting (2D weights), set this to `"temporal"`.
`None` defaults to sample-wise weights (1D).
If the model has multiple outputs, you can use a different
`sample_weight_mode` on each output by passing a
dictionary or a list of modes.
**kwargs: when using the Theano backend, these arguments
are passed into K.function. When using the Tensorflow backend,
these arguments are passed into `tf.Session.run`.
# Raises
ValueError: In case of invalid arguments for
`optimizer`, `loss`, `metrics` or `sample_weight_mode`.
"""
loss = loss or {}
self.optimizer = optimizers.get(optimizer)
self.sample_weight_mode = sample_weight_mode
self.loss = loss
self.loss_weights = loss_weights
# Prepare loss functions.
if isinstance(loss, dict):
for name in loss:
if name not in self.output_names:
raise ValueError('Unknown entry in loss '
'dictionary: "' + name + '". '
'Only expected the following keys: ' +
str(self.output_names))
loss_functions = []
for name in self.output_names:
if name not in loss:
warnings.warn('Output "' + name +
'" missing from loss dictionary. '
'We assume this was done on purpose, '
'and we will not be expecting '
'any data to be passed to "' + name +
'" during training.', stacklevel=2)
loss_functions.append(losses.get(loss.get(name)))
elif isinstance(loss, list):
if len(loss) != len(self.outputs):
raise ValueError('When passing a list as loss, '
'it should have one entry per model outputs. '
'The model has ' + str(len(self.outputs)) +
' outputs, but you passed loss=' +
str(loss))
loss_functions = [losses.get(l) for l in loss]
else:
loss_function = losses.get(loss)
loss_functions = [loss_function for _ in range(len(self.outputs))]
self.loss_functions = loss_functions
weighted_losses = [_weighted_masked_objective(fn) for fn in loss_functions]
skip_indices = []
self._feed_outputs = []
self._feed_output_names = []
self._feed_output_shapes = []
self._feed_loss_fns = []
for i in range(len(weighted_losses)):
if weighted_losses[i] is None:
skip_indices.append(i)
else:
self._feed_outputs.append(self.outputs[i])
self._feed_output_names.append(self.output_names[i])
self._feed_output_shapes.append(self.internal_output_shapes[i])
self._feed_loss_fns.append(self.loss_functions[i])
# Prepare output masks.
masks = self.compute_mask(self.inputs, mask=None)
if masks is None:
masks = [None for _ in self.outputs]
if not isinstance(masks, list):
masks = [masks]
# Prepare loss weights.
if loss_weights is None:
loss_weights_list = [1. for _ in range(len(self.outputs))]
elif isinstance(loss_weights, dict):
for name in loss_weights:
if name not in self.output_names:
raise ValueError('Unknown entry in loss_weights '
'dictionary: "' + name + '". '
'Only expected the following keys: ' +
str(self.output_names))
loss_weights_list = []
for name in self.output_names:
loss_weights_list.append(loss_weights.get(name, 1.))
elif isinstance(loss_weights, list):
if len(loss_weights) != len(self.outputs):
raise ValueError('When passing a list as loss_weights, '
'it should have one entry per model outputs. '
'The model has ' + str(len(self.outputs)) +
' outputs, but you passed loss_weights=' +
str(loss_weights))
loss_weights_list = loss_weights
else:
raise TypeError('Could not interpret loss_weights argument: ' +
str(loss_weights) +
' - expected a list of dicts.')
# Prepare sample weights.
sample_weights = []
sample_weight_modes = []
if isinstance(sample_weight_mode, dict):
for name in sample_weight_mode:
if name not in self.output_names:
raise ValueError('Unknown entry in '
'sample_weight_mode dictionary: "' +
name + '". '
'Only expected the following keys: ' +
str(self.output_names))
for i, name in enumerate(self.output_names):
if i in skip_indices:
weight = None
sample_weight_modes.append(None)
else:
if name not in sample_weight_mode:
raise ValueError('Output "' + name +
'" missing from sample_weight_modes '
'dictionary')
if sample_weight_mode.get(name) == 'temporal':
weight = K.placeholder(ndim=2,
name=name + '_sample_weights')
sample_weight_modes.append('temporal')
else:
weight = K.placeholder(ndim=1,
name=name + '_sample_weights')
sample_weight_modes.append(None)
sample_weights.append(weight)
elif isinstance(sample_weight_mode, list):
if len(sample_weight_mode) != len(self.outputs):
raise ValueError('When passing a list as sample_weight_mode, '
'it should have one entry per model outputs. '
'The model has ' + str(len(self.outputs)) +
' outputs, but you passed '
'sample_weight_mode=' +
str(sample_weight_mode))
for i in range(len(self.output_names)):
if i in skip_indices:
weight = None
sample_weight_modes.append(None)
else:
mode = sample_weight_mode[i]
name = self.output_names[i]
if mode == 'temporal':
weight = K.placeholder(ndim=2,
name=name + '_sample_weights')
sample_weight_modes.append('temporal')
else:
weight = K.placeholder(ndim=1,
name=name + '_sample_weights')
sample_weight_modes.append(None)
sample_weights.append(weight)
else:
for i, name in enumerate(self.output_names):
if i in skip_indices:
sample_weight_modes.append(None)
sample_weights.append(None)
else:
if sample_weight_mode == 'temporal':
sample_weights.append(
K.placeholder(ndim=2,
name=name + '_sample_weights'))
sample_weight_modes.append('temporal')
else:
sample_weights.append(
K.placeholder(ndim=1,
name=name + '_sample_weights'))
sample_weight_modes.append(None)
self.sample_weight_modes = sample_weight_modes
self._feed_sample_weight_modes = []
for i in range(len(self.outputs)):
if i not in skip_indices:
self._feed_sample_weight_modes.append(self.sample_weight_modes[i])
# Prepare targets of model.
self.targets = []
self._feed_targets = []
for i in range(len(self.outputs)):
if i in skip_indices:
self.targets.append(None)
else:
shape = self.internal_output_shapes[i]
name = self.output_names[i]
target = K.placeholder(ndim=len(shape),
name=name + '_target',
sparse=K.is_sparse(self.outputs[i]),
dtype=K.dtype(self.outputs[i]))
self.targets.append(target)
self._feed_targets.append(target)
# Prepare metrics.
self.metrics = metrics
self.metrics_names = ['loss']
self.metrics_tensors = []
# Compute total loss.
total_loss = None
for i in range(len(self.outputs)):
if i in skip_indices:
continue
y_true = self.targets[i]
y_pred = self.outputs[i]
weighted_loss = weighted_losses[i]
sample_weight = sample_weights[i]
mask = masks[i]
loss_weight = loss_weights_list[i]
output_loss = weighted_loss(y_true, y_pred,
sample_weight, mask)
if len(self.outputs) > 1:
self.metrics_tensors.append(output_loss)
self.metrics_names.append(self.output_names[i] + '_loss')
if total_loss is None:
total_loss = loss_weight * output_loss
else:
total_loss += loss_weight * output_loss
if total_loss is None:
if not self.losses:
raise RuntimeError('The model cannot be compiled '
'because it has no loss to optimize.')
else:
total_loss = 0.
# Add regularization penalties
# and other layer-specific losses.
for loss_tensor in self.losses:
total_loss += loss_tensor
# List of same size as output_names.
# contains tuples (metrics for output, names of metrics).
nested_metrics = _collect_metrics(metrics, self.output_names)
def append_metric(layer_num, metric_name, metric_tensor):
"""Helper function used in loop below."""
if len(self.output_names) > 1:
metric_name = self.output_layers[layer_num].name + '_' + metric_name
self.metrics_names.append(metric_name)
self.metrics_tensors.append(metric_tensor)
for i in range(len(self.outputs)):
if i in skip_indices:
continue
y_true = self.targets[i]
y_pred = self.outputs[i]
output_metrics = nested_metrics[i]
for metric in output_metrics:
if metric == 'accuracy' or metric == 'acc':
# custom handling of accuracy
# (because of class mode duality)
output_shape = self.internal_output_shapes[i]
acc_fn = None
if (output_shape[-1] == 1 or
self.loss_functions[i] == losses.binary_crossentropy):
# case: binary accuracy
acc_fn = metrics_module.binary_accuracy
elif self.loss_functions[i] == losses.sparse_categorical_crossentropy:
# case: categorical accuracy with sparse targets
acc_fn = metrics_module.sparse_categorical_accuracy
else:
acc_fn = metrics_module.categorical_accuracy
masked_fn = _masked_objective(acc_fn)
append_metric(i, 'acc', masked_fn(y_true, y_pred, mask=masks[i]))
else:
metric_fn = metrics_module.get(metric)
masked_metric_fn = _masked_objective(metric_fn)
metric_result = masked_metric_fn(y_true, y_pred, mask=masks[i])
metric_result = {
metric_fn.__name__: metric_result
}
for name, tensor in six.iteritems(metric_result):
append_metric(i, name, tensor)
# Prepare gradient updates and state updates.
self.total_loss = total_loss
self.sample_weights = sample_weights
self._feed_sample_weights = []
for i in range(len(self.sample_weights)):
if i not in skip_indices:
self._feed_sample_weights.append(sample_weights[i])
# Functions for train, test and predict will
# be compiled lazily when required.
# This saves time when the user is not using all functions.
self._function_kwargs = kwargs
self.train_function = None
self.test_function = None
self.predict_function = None
# Collected trainable weights and sort them deterministically.
trainable_weights = self.trainable_weights
# Sort weights by name.
if trainable_weights:
if K.backend() == 'theano':
trainable_weights.sort(key=lambda x: x.name if x.name else x.auto_name)
else:
trainable_weights.sort(key=lambda x: x.name)