forked from neuroailab/Neural-Alignment
-
Notifications
You must be signed in to change notification settings - Fork 0
/
params.py
637 lines (585 loc) · 33.2 KB
/
params.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
'''
Builds a dictionary structured as tfutils expects it
'''
import copy
import tensorflow as tf
import dill
from tfutils import optimizer
# Data providers
from tfutils.imagenet_data import ImageNet
from tfutils.tests import mnist_data
from tfutils.db_interface import TFUTILS_HOME
import learning_rate as lr
import rate_scheduler
from Alignments import alignment
from Metrics import functions as metrics
from Metrics import losses
from Models import basic, resnet_model_google
from custom_optimizers import SWATSOptimizer, RAdamOptimizer
from custom_optimizers import build_noisy_optimizer
class Params:
def __init__(self):
self.params = {}
self.alignment_kwargs = {}
self.rate_scheduler_kwargs = {}
self.lr_scheduler_kwargs = {}
self.alignment_lr_scheduler_kwargs = {}
def _build_default_params(self, flags):
FLAGS = flags
# Dataset constants
if FLAGS.dataset == 'mnist':
self._LABEL_CLASSES = 10
self._NUM_CHANNELS = 1
self._NUM_TRAIN_IMAGES = 60000
self._NUM_EVAL_IMAGES = 10000
self.train_data_params = {
'func': mnist_data.build_data,
'batch_size': FLAGS.train_batch_size,
'group': 'train',
'directory': TFUTILS_HOME}
self.val_data_params = {
'func': mnist_data.build_data,
'batch_size': FLAGS.eval_batch_size,
'group': 'test',
'directory': TFUTILS_HOME}
elif FLAGS.dataset == 'imagenet':
self._LABEL_CLASSES = 1000
self._NUM_CHANNELS = 3
self._NUM_TRAIN_IMAGES = 1281167
self._NUM_EVAL_IMAGES = 49920
if FLAGS.use_resnet_v2:
self._data_prep_type = 'inception'
else:
self._data_prep_type = 'resnet'
# This is the GPU version of the data providers
assert(FLAGS.data_dir is not None)
self.train_data_params = {
'func': ImageNet(image_dir=FLAGS.data_dir,
prep_type=self._data_prep_type,
resize=FLAGS.image_size).dataset_func,
'is_train': True,
'batch_size': FLAGS.train_batch_size}
self.val_data_params = {
'func': ImageNet(image_dir=FLAGS.data_dir,
prep_type=self._data_prep_type,
resize=FLAGS.image_size).dataset_func,
'is_train': False,
'batch_size': FLAGS.eval_batch_size,
'q_cap': FLAGS.eval_batch_size,
'file_pattern': 'validation-*'}
self.NUM_BATCHES_PER_EPOCH = self._NUM_TRAIN_IMAGES / FLAGS.train_batch_size
self._MOMENTUM = 0.9
self.rate_scheduler_kwargs = {'loss_rate': FLAGS.loss_rate,
'alignment_rate': FLAGS.alignment_rate,
'delay_epochs': FLAGS.delay_epochs,
'alternate_step_freq': FLAGS.alternate_step_freq,
'constant_rate': FLAGS.constant_rate,
'num_batches_per_epoch': self.NUM_BATCHES_PER_EPOCH}
self.lr_scheduler_kwargs = {'learning_rate': FLAGS.learning_rate,
'num_batches_per_epoch': self.NUM_BATCHES_PER_EPOCH,
'train_batch_size': FLAGS.train_batch_size,
'base_batch_size': FLAGS.base_batch_size,
'rescale_lr': FLAGS.rescale_lr,
'constant_lr': FLAGS.constant_lr,
'warmup_epochs': FLAGS.warmup_epochs,
'alternate_step_freq': FLAGS.alternate_step_freq,
'delay_epochs': FLAGS.delay_epochs,
'delay_epochs_offset': None}
if FLAGS.load_step is None:
load_query = None # loads most recent step
else:
load_query = {'step': FLAGS.load_step}
self.params = {
'save_params': {
'host': 'localhost',
'port': FLAGS.port,
'dbname': FLAGS.dbname,
'collname': None,
'exp_id': None,
'do_save': True,
'save_valid_freq': (int)(FLAGS.epochs_per_checkpoint*self.NUM_BATCHES_PER_EPOCH),
'save_filters_freq': (int)(FLAGS.epochs_per_checkpoint*self.NUM_BATCHES_PER_EPOCH),
'cache_filters_freq': (int)(FLAGS.epochs_per_checkpoint*self.NUM_BATCHES_PER_EPOCH),
'cache_dir': None, # model directory for google cloud bucket
},
'load_params': {
'do_restore': FLAGS.do_restore,
'query': load_query
},
'model_params': {
'func': None,
'model_prefix': 'model_0',
'num_classes': self._LABEL_CLASSES, # dataset dependent
'layers_list': [],
'num_gpus': len(FLAGS.gpu.split(',')),
'devices': ['/gpu:%i' % idx for idx in range(len(FLAGS.gpu.split(',')))],
# The following params only needed if tpu, passed as kwargs to model_fn
# Will only train on TPU is tpu_name is not None
'tpu_name': FLAGS.tpu_name,
'gcp_project': FLAGS.gcp_project,
'tpu_zone': FLAGS.tpu_zone,
'num_shards': FLAGS.num_shards,
'iterations_per_loop': FLAGS.iterations_per_loop,
},
'train_params': {
'targets': {'func': losses.loss_metric,
'target': 'labels',
'rate': rate_scheduler.build_schedule(**self.rate_scheduler_kwargs)},
'data_params': self.train_data_params,
'num_steps': (int)(FLAGS.train_epochs*self.NUM_BATCHES_PER_EPOCH), # number of steps to train
'thres_loss': float('Inf'),
'validate_first': FLAGS.validate_first, # You may want to turn this off at debugging
'include_global_step': False,
},
'loss_params': {
'targets': 'labels',
'agg_func': losses.mean_loss_with_reg,
'agg_func_kwargs': {'rate': rate_scheduler.build_schedule(**self.rate_scheduler_kwargs)},
'loss_per_case_func': losses.category_loss,
},
'learning_rate_params': {
'func': lr.build_lr_schedule(**self.lr_scheduler_kwargs)
},
'optimizer_params': {
'optimizer': optimizer.ClipOptimizer,
'optimizer_class': tf.train.GradientDescentOptimizer,
'clip': FLAGS.grad_clip,
'clipping_value': FLAGS.grad_clipping_value,
'clipping_method': FLAGS.grad_clipping_method
},
'validation_params': {
'topn_val': {
'data_params': self.val_data_params,
'targets': {
'func': metrics.metric_fn,
'num_classes': self._LABEL_CLASSES
},
'num_steps': self._NUM_EVAL_IMAGES // FLAGS.eval_batch_size,
'agg_func': metrics.concat_agg_func,
'online_agg_func': metrics.online_agg_append,
}
},
'skip_check': FLAGS.skip_check,
}
if FLAGS.minibatch_size is not None:
self.params['train_params']['minibatch_size'] = FLAGS.minibatch_size
if FLAGS.load_db:
self.params['load_params'] = {'host': 'localhost',
'port': FLAGS.load_port,
'dbname': FLAGS.load_dbname,
'collname': FLAGS.load_collname,
'exp_id': FLAGS.load_exp_id,
'do_restore': True,
'query': load_query}
def _set_optimizer(self, optimizer_class):
if optimizer_class == 'momentum':
print("Using Momentum Optimizer")
optimizer_params = {
'optimizer_class': tf.train.MomentumOptimizer,
'optimizer_kwargs': {'momentum': self._MOMENTUM, 'use_nesterov': True}
}
elif optimizer_class == 'adagrad':
print("Using Adagrad Optimizer")
optimizer_params = {
'optimizer_class': tf.train.AdagradOptimizer
}
elif optimizer_class == 'rmsprop':
print("Using RMSProp Optimizer")
# Note: default momentum value for RMSProp is 0.0
optimizer_params = {
'optimizer_class': tf.train.RMSPropOptimizer
}
elif optimizer_class == 'adam':
print("Using ADAM Optimizer")
optimizer_params = {
'optimizer_class': tf.train.AdamOptimizer
}
elif optimizer_class == 'swats':
print("Using SWATS Optimizer")
optimizer_params = {
'optimizer_class': SWATSOptimizer,
'optimizer_kwargs': {'rectified_adam': False,
'include_global_step': True}
}
elif optimizer_class == 'swrats':
print("Using SWATS Optimizer with RAdam")
optimizer_params = {
'optimizer_class': SWATSOptimizer,
'optimizer_kwargs': {'rectified_adam': True,
'include_global_step': True}
}
elif optimizer_class == 'radam':
print("Using RADAM Optimizer")
optimizer_params = {
'optimizer_class': RAdamOptimizer,
'optimizer_kwargs': {'include_global_step': True}
}
else:
print("Using default Gradient Descent Optimizer")
optimizer_params = {
'optimizer_class': tf.train.GradientDescentOptimizer
}
return optimizer_params
def customize(self, flags):
FLAGS = flags
self._build_default_params(FLAGS)
# Choosing the model
if FLAGS.model == 'fc':
self.params['model_params']['func'] = basic.fc
self.params['model_params']['activation'] = FLAGS.activation
self.params['save_params']['collname'] = 'fc'
if FLAGS.layers_list is not None:
layers_list = [int(l) for l in FLAGS.layers_list.split(",")]
assert type(layers_list) == list
for l in layers_list:
assert type(l) == int
self.params['model_params']['layers_list'] = layers_list
self.params['save_params']['collname'] = 'fc_' + str(FLAGS.layers_list.replace(',','-'))
elif 'resnet' in FLAGS.model:
self.params['save_params']['collname'] = FLAGS.model
if FLAGS.use_resnet_v2:
self.params['save_params']['collname'] += 'v2'
self.params['model_params']['func'] = resnet_model_google.google_resnet_func
self.params['model_params']['tf_layers'] = FLAGS.tf_layers
self.params['model_params']['resnet_size'] = (int)(FLAGS.model.split('resnet')[-1])
self.params['model_params']['use_v2'] = FLAGS.use_resnet_v2
self.params['model_params']['bn_trainable'] = FLAGS.bn_trainable
print("regularize_weights_via_model", FLAGS.regularize_weights_via_model)
print("regularize_weights_via_model (type)", type(FLAGS.regularize_weights_via_model))
self.params['model_params']['regularize_weights'] = FLAGS.regularize_weights_via_model
self.params['save_params']['exp_id'] = str(FLAGS.alignment)
self.params['save_params']['exp_id'] += FLAGS.exp_id_suffix
# Common alignment kwargs used by the Alignment parent class
self.alignment_kwargs = {
'update_forward': FLAGS.update_forward,
'input_distribution': FLAGS.input_distribution,
'input_stddev': FLAGS.input_stddev,
'use_bias_forward': FLAGS.use_bias_forward,
'use_bias_backward': FLAGS.use_bias_backward,
'activation_fn_override': FLAGS.activation_fn_override,
'activation_forward': FLAGS.activation_forward,
'activation_backward': FLAGS.activation_backward,
'batch_center_backward_input': FLAGS.batch_center_backward_input,
'center_input': FLAGS.center_input,
'normalize_input': FLAGS.normalize_input,
'batch_center_forward_output': FLAGS.batch_center_forward_output,
'center_forward_output': FLAGS.center_forward_output,
'normalize_forward_output': FLAGS.normalize_forward_output,
'center_backward_output': FLAGS.center_backward_output,
'normalize_backward_output': FLAGS.normalize_backward_output}
# Alignment coefficient kwargs used to build rate schedulers
self.alignment_coefficient_kwargs = {
'alpha': {'start': FLAGS.alpha_start,
'stop': FLAGS.alpha_stop,
'cycle': FLAGS.alpha_cycle,
'schedule_rate': FLAGS.alpha_schedule_rate,
'schedule_type': FLAGS.alpha_schedule_type,
'num_batches_per_epoch': self.NUM_BATCHES_PER_EPOCH,
'train_epochs': FLAGS.train_epochs},
'beta': {'start': FLAGS.beta_start,
'stop': FLAGS.beta_stop,
'cycle': FLAGS.beta_cycle,
'schedule_rate': FLAGS.beta_schedule_rate,
'schedule_type': FLAGS.beta_schedule_type,
'num_batches_per_epoch': self.NUM_BATCHES_PER_EPOCH,
'train_epochs': FLAGS.train_epochs},
'gamma': {'start': FLAGS.gamma_start,
'stop': FLAGS.gamma_stop,
'cycle': FLAGS.gamma_cycle,
'schedule_rate': FLAGS.gamma_schedule_rate,
'schedule_type': FLAGS.gamma_schedule_type,
'num_batches_per_epoch': self.NUM_BATCHES_PER_EPOCH,
'train_epochs': FLAGS.train_epochs}
}
# Set the alignment
if FLAGS.alignment == 'feedback':
print("Using Feedback Alignment")
self.params['model_params']['alignment'] = alignment.Feedback()
elif FLAGS.alignment == 'symmetric':
print("Using Symmetric Alignment")
# alpha scheduler
self.alignment_coefficient_kwargs['alpha']['value'] = FLAGS.alpha if FLAGS.alpha is not None else 1.0e-3
self.alignment_kwargs['alpha'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['alpha'])
# beta scheduler
self.alignment_coefficient_kwargs['beta']['value'] = FLAGS.beta if FLAGS.beta is not None else 2.0e-3
self.alignment_kwargs['beta'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['beta'])
# build alignment class
self.params['model_params']['alignment'] = alignment.Symmetric(**self.alignment_kwargs)
elif FLAGS.alignment == 'activation':
print("Using Activation Alignment")
# alpha scheduler
self.alignment_coefficient_kwargs['alpha']['value'] = FLAGS.alpha if FLAGS.alpha is not None else 1.0e-3
self.alignment_kwargs['alpha'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['alpha'])
# beta scheduler
self.alignment_coefficient_kwargs['beta']['value'] = FLAGS.beta if FLAGS.beta is not None else 2.0e-3
self.alignment_kwargs['beta'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['beta'])
# build alignment class
self.params['model_params']['alignment'] = alignment.Activation(**self.alignment_kwargs)
elif FLAGS.alignment == 'mirror':
print("Using Weight Mirror")
# alpha scheduler
self.alignment_coefficient_kwargs['alpha']['value'] = FLAGS.alpha if FLAGS.alpha is not None else 1.0e-3
self.alignment_kwargs['alpha'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['alpha'])
# beta scheduler
self.alignment_coefficient_kwargs['beta']['value'] = FLAGS.beta if FLAGS.beta is not None else 1.0e-3
self.alignment_kwargs['beta'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['beta'])
# build alignment class
self.params['model_params']['alignment'] = alignment.Mirror(**self.alignment_kwargs)
elif FLAGS.alignment == 'information':
print("Using Information Alignment")
# alpha scheduler
self.alignment_coefficient_kwargs['alpha']['value'] = FLAGS.alpha if FLAGS.alpha is not None else 2.0e-3
self.alignment_kwargs['alpha'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['alpha'])
# beta scheduler
self.alignment_coefficient_kwargs['beta']['value'] = FLAGS.beta if FLAGS.beta is not None else 1.0e-3
self.alignment_kwargs['beta'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['beta'])
# gamma scheduler
self.alignment_coefficient_kwargs['gamma']['value'] = FLAGS.gamma if FLAGS.gamma is not None else 1.0e-3
self.alignment_kwargs['gamma'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['gamma'])
# boolean hyperparameters
self.alignment_kwargs['reconstruction_reversal'] = FLAGS.reconstruction_reversal
self.alignment_kwargs['reconstruction_amp'] = FLAGS.reconstruction_amp
self.alignment_kwargs['use_sparse'] = FLAGS.use_sparse
# build alignment class
self.params['model_params']['alignment'] = alignment.Information(**self.alignment_kwargs)
elif FLAGS.alignment == 'kolen_pollack':
print("Using Kolen Pollack")
# alpha scheduler
self.alignment_coefficient_kwargs['alpha']['value'] = FLAGS.alpha if FLAGS.alpha is not None else 1.0e-3
self.alignment_kwargs['alpha'] = rate_scheduler.build_alignment_coefficient_schedule(**self.alignment_coefficient_kwargs['alpha'])
# build alignment class
self.params['model_params']['alignment'] = alignment.Kolen_Pollack(**self.alignment_kwargs)
elif FLAGS.alignment is None:
print("No alignment specified, defaulting to backprop")
else:
raise ValueError
# This is to pass the alignment to the metric_fn, so we can plot the
# alpha, beta and gamma schedules
if FLAGS.save_alignment_coefficients:
print("Saving alignment coefficients :) ")
self.params['validation_params']['topn_val']['targets'].update({'alignment':self.params['model_params']['alignment']})
# Caches
if FLAGS.cache_dir is not None:
self.params['save_params']['cache_dir'] = '{}/localhost:{}/{}/{}/{}'.format(FLAGS.cache_dir,
self.params['save_params']['port'],
self.params['save_params']['dbname'],
self.params['save_params']['collname'],
self.params['save_params']['exp_id'])
# LR drops
if FLAGS.manual_lr:
if FLAGS.rescale_lr:
scaled_lr = FLAGS.learning_rate * (FLAGS.train_batch_size / (float)(FLAGS.base_batch_size))
else:
scaled_lr = FLAGS.learning_rate
self.lr_scheduler_kwargs = {'scaled_lr': scaled_lr,
'drop': FLAGS.drop,
'boundary_step': FLAGS.boundary_step,
'num_batches_per_epoch': self.NUM_BATCHES_PER_EPOCH,
'warmup_epochs': FLAGS.warmup_epochs}
self.params['learning_rate_params'] = lr.manual_lr(**self.lr_scheduler_kwargs)
# Optimizer
optimizer_params = self._set_optimizer(optimizer_class=FLAGS.optimizer)
if FLAGS.use_noisy_global_opt:
print("Using NoisyOptimizer on the global optimizer")
apply_filter = 'backward' if FLAGS.alignment == 'kolen_pollack' else ''
if FLAGS.noisy_global_opt_distribution is not None:
noisy_global_opt = build_noisy_optimizer(optimizer_params['optimizer_class'],
FLAGS.noisy_global_opt_distribution,
FLAGS.noisy_global_opt_variance,
apply_filter=apply_filter)
else:
noisy_global_opt = build_noisy_optimizer(optimizer_params['optimizer_class'],
FLAGS.noisy_opt_distribution,
FLAGS.noisy_opt_variance,
apply_filter=apply_filter)
optimizer_params.update({'optimizer_class': noisy_global_opt})
self.params['optimizer_params'].update(optimizer_params)
opt_req_global_step = ['swats', 'radam', 'swrats']
if FLAGS.alignment_optimizer in opt_req_global_step or \
FLAGS.optimizer in opt_req_global_step:
self.params['train_params'].update({'include_global_step': True})
if FLAGS.alignment_optimizer is not None:
# have loss returned be a list ([model + reg_loss, alignment losses])
self.params['loss_params']['agg_func_kwargs']['return_list'] = True
# pass in optimizer params per optimizer
alignment_optimizer_params = self._set_optimizer(optimizer_class=FLAGS.alignment_optimizer)
if FLAGS.use_noisy_alignment_opt:
print("Using NoisyOptimizer on the alignment optimizer")
apply_filter = 'backward' if FLAGS.alignment == 'kolen_pollack' else ''
if FLAGS.noisy_alignment_opt_distribution is not None:
noisy_alignment_opt = build_noisy_optimizer(alignment_optimizer_params['optimizer_class'],
FLAGS.noisy_alignment_opt_distribution,
FLAGS.noisy_alignment_opt_variance,
apply_filter=apply_filter)
else:
noisy_alignment_opt = build_noisy_optimizer(alignment_optimizer_params['optimizer_class'],
FLAGS.noisy_opt_distribution,
FLAGS.noisy_opt_variance,
apply_filter=apply_filter)
alignment_optimizer_params.update({'optimizer_class': noisy_alignment_opt})
for k in ['optimizer_class', 'optimizer_kwargs']:
if k == 'optimizer_class':
self.params['optimizer_params'][k] = [self.params['optimizer_params'][k], alignment_optimizer_params[k]]
elif k == 'optimizer_kwargs':
model_optimizer_kwargs = self.params['optimizer_params'][k] if k in self.params['optimizer_params'].keys() else {}
alignment_optimizer_kwargs = alignment_optimizer_params[k] if k in alignment_optimizer_params.keys() else {}
self.params['optimizer_params'][k] = [model_optimizer_kwargs, alignment_optimizer_kwargs]
else:
raise ValueError
# set learning rate schedule
if FLAGS.alignment_manual_lr:
if FLAGS.alignment_rescale_lr:
alignment_scaled_lr = FLAGS.alignment_learning_rate * (FLAGS.train_batch_size / (float)(FLAGS.base_batch_size))
else:
alignment_scaled_lr = FLAGS.alignment_learning_rate
self.alignment_lr_scheduler_kwargs = {
'scaled_lr': alignment_scaled_lr,
'drop': FLAGS.drop if FLAGS.alignment_drop is None else FLAGS.alignment_drop,
'boundary_step': FLAGS.boundary_step if FLAGS.alignment_boundary_step is None else FLAGS.alignment_boundary_step,
'num_batches_per_epoch': self.NUM_BATCHES_PER_EPOCH,
'warmup_epochs': FLAGS.alignment_warmup_epochs
}
alignment_lr_params = lr.manual_lr(**self.alignment_lr_scheduler_kwargs)
else:
self.alignment_lr_scheduler_kwargs = {
'learning_rate': FLAGS.alignment_learning_rate,
'num_batches_per_epoch': self.NUM_BATCHES_PER_EPOCH,
'train_batch_size': FLAGS.train_batch_size,
'base_batch_size': FLAGS.base_batch_size,
'rescale_lr': FLAGS.alignment_rescale_lr,
'constant_lr': FLAGS.alignment_constant_lr,
'warmup_epochs': FLAGS.alignment_warmup_epochs,
'alternate_step_freq': FLAGS.alternate_step_freq,
# delay epochs in rate_scheduler only sets the class loss
# rate to 0 so should only apply to categorization to set
# its lr to 0 until then delay, but not alignment
'delay_epochs': None,
# to sync the alignment lr drops with class loss, we
# add this to the LR schedule since alignment will be
# running for delay epochs longer with nonzero lr
'delay_epochs_offset': FLAGS.delay_epochs
}
alignment_lr_params = {'func': lr.build_lr_schedule(**self.alignment_lr_scheduler_kwargs)}
self.params['learning_rate_params'] = {
'func': lr.combined_lr,
'lr_params':[self.params['learning_rate_params'], alignment_lr_params]
}
# TPU compatibility
if FLAGS.tpu_name is not None:
self.make_tpu_compatible(FLAGS)
# Debugging statements
print("Saving cache at: {}".format(self.params['save_params']['cache_dir']))
def make_tpu_compatible(self, flags):
FLAGS = flags
if FLAGS.cache_dir is not None:
tpu_cache_dir = FLAGS.cache_dir
else:
tpu_cache_dir = 'neur-al'
# Update the data params
self.params['train_params']['data_params'] = {
'func': ImageNet(image_dir=FLAGS.data_dir,
prep_type=self._data_prep_type,
is_train=True,
resize=FLAGS.image_size).dataset_func_tpu,
'batch_size': FLAGS.train_batch_size}
self.params['validation_params']['topn_val']['data_params'] = {
'func': ImageNet(image_dir=FLAGS.data_dir,
prep_type=self._data_prep_type,
is_train=False,
resize=FLAGS.image_size).dataset_func_tpu,
'batch_size': FLAGS.eval_batch_size}
# Send the model the gpu_mode flag in False
self.params['model_params']['gpu_mode'] = False
# Use the tpu version of the loss agg func
self.params['loss_params']['agg_func'] = losses.mean_loss_with_reg_tpu
# Change the metric function
self.params['validation_params']['topn_val']['targets']['func'] = metrics.metric_fn_tpu
self.params['validation_params']['topn_val'].pop('agg_func')
self.params['validation_params']['topn_val'].pop('online_agg_func')
# max checkpoints to keep in gcloud cache
self.params['save_params']['checkpoint_max'] = FLAGS.checkpoint_max
# Save caches in glcoud
self.params['save_params']['cache_dir'] = 'gs://{}/{}/{}/{}/'.format(tpu_cache_dir,
self.params['save_params']['dbname'],
self.params['save_params']['collname'],
self.params['save_params']['exp_id'])
def save(self, filename):
# TODO: check if we need to pop the alignment, rate_scheduler, coefficient kwargs
# and sve them separtately, and recreate the objects when loading
params_to_save = self.get_params_copy()
with open(filename, 'wb') as f:
dill.dump({'params': params_to_save}, f)
print("Params were saved at {}".format(filename)) #TODO: expand the full path here
def load(self, filename, flags=None):
# Parse the initial flags (will set some class attributes)
if flags is not None:
self._build_default_params(flags)
# Load the config
with open(filename, 'rb') as f:
saved_params = dill.load(f)
self.params = saved_params['params']
# Parse the relevant flags
if flags is not None:
self._customize_save(flags)
def _customize_save(self, flags):
FLAGS = flags
# Overwrite save_params
self.params['save_params'].update({
'host': 'localhost',
'port': FLAGS.port,
'dbname': FLAGS.dbname,
'cache_dir': FLAGS.cache_dir})
self.params['save_params']['exp_id'] += FLAGS.exp_id_suffix
# Allow customizing which device to run on
self.params['model_params'].update({
'num_gpus': len(FLAGS.gpu.split(',')),
'devices': ['/gpu:%i' % idx for idx in range(len(FLAGS.gpu.split(',')))],
# The following params only needed if tpu, passed as kwargs to model_fn
# Will only train on TPU is tpu_name is not None
'tpu_name': FLAGS.tpu_name,
'gcp_project': FLAGS.gcp_project,
'tpu_zone': FLAGS.tpu_zone,
'num_shards': FLAGS.num_shards,
'iterations_per_loop': FLAGS.iterations_per_loop})
# Overwrite data params for GPU
# NOTE: FLAGS.image_size is the only thing not directly saved
# in the params dictionary we loaded, so we rely on the user
# setting it appropriately
self.params['train_params']['data_params'] = {
'func': ImageNet(image_dir=FLAGS.data_dir,
prep_type=self._data_prep_type,
resize=FLAGS.image_size).dataset_func,
'is_train': True,
'batch_size': FLAGS.train_batch_size}
self.params['validation_params']['topn_val']['data_params'] = {
'func': ImageNet(image_dir=FLAGS.data_dir,
prep_type=self._data_prep_type,
resize=FLAGS.image_size).dataset_func,
'is_train': False,
'batch_size': FLAGS.eval_batch_size,
'q_cap': FLAGS.eval_batch_size,
'file_pattern': 'validation-*'}
# Allow for TPU training
if FLAGS.tpu_name is not None:
# Make sure the flags take these values from the loaded config file
FLAGS.train_batch_size = self.params['train_params']['data_params']['batch_size']
FLAGS.eval_batch_size = self.params['validation_params']['topn_val']['data_params']['batch_size']
FLAGS.checkpoint_max = self.params['save_params']['checkpoint_max']
self.make_tpu_compatible(FLAGS)
# Allow using different types of resnets
self.params['model_params']['resnet_size'] = (int)(FLAGS.model.split('resnet')[-1])
self.params['model_params']['use_v2'] = FLAGS.use_resnet_v2
if FLAGS.use_resnet_v2:
self.params['save_params']['collname'] += 'v2'
def get_params_copy(self):
return copy.deepcopy(self.params)
def get_alignment_kwargs_copy(self):
return copy.deepcopy(self.alignment_kwargs)
def get_rate_scheduler_kwargs_copy(self):
return copy.deepcopy(self.rate_scheduler_kwargs)
def get_lr_scheduler_kwargs_copy(self):
return copy.deepcopy(self.lr_scheduler_kwargs)
def get_alignment_lr_scheduler_kwargs_copy(self):
return copy.deepcopy(self.alignment_lr_scheduler_kwargs)
def get_alignment_coefficient_kwargs_copy(self):
return copy.deepcopy(self.alignment_coefficient_kwargs)