-
Notifications
You must be signed in to change notification settings - Fork 0
/
maxout.py
447 lines (357 loc) · 13.6 KB
/
maxout.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
#!/usr/bin/env python3
'''\
Main script file. Depending on the parameters, we can train or test with
different Maxout networks.
'''
import os
import argparse
import shutil
import tensorflow as tf
import numpy as np
from graphs import CGraph, EmaVariables
from tools import RunContexts
def training(args):
'''\
Training function. Saves checkpoints in models/<args.dataset>/model.
Args:
args: a namespace object. Run `maxout.py -h' for a list of the available
options.
'''
# Prints
print('| Training')
print('| Dataset:', args.dataset, flush=True)
# Start or continue? Check directories and set iteration range
if not args.cont:
# Start from scratch
_clear_saved(args.dataset)
steps_range = range(1, args.steps + 1)
else:
# Continue
if os.path.exists('logs/train') or os.path.exists('logs/val'):
raise FileExistsError(
"Move 'logs/train' and 'logs/val' if you want to continue.")
with open('logs/last_step.txt') as step_f:
last_step = int(step_f.read())
steps_range = range(last_step+1, last_step+args.steps+1)
# Instantiate the graph
graph = CGraph(args.dataset, args.batch, args.seed, args.regularization,
args.renormalization)
# Use it
with graph.graph.as_default():
# Constant seed for debugging
if args.seed:
tf.set_random_seed(args.seed)
# Optimizer
optimizer, rate, step_var = _select_optimizer(args)
minimize = optimizer.minimize(graph.loss, step_var)
# Tensorboard summaries
scalar_summaries = (
('1_accuracy', graph.accuracy),
('2_loss', graph.loss),
('3_learning_rate', rate))
for scalar in scalar_summaries:
tf.summary.scalar(*scalar)
#images = tf.get_collection('VISUALIZATIONS')
#if images: # Not saving images to save space
# tf.summary.image('tensor_images', images[0], max_outputs=10)
train_summaries_op = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('logs/train', graph=graph.graph)
val_writer = tf.summary.FileWriter('logs/val')
# Predict with running averages
variables = EmaVariables(args.ema)
# Variables initializer and saver
init = tf.global_variables_initializer()
saver = tf.train.Saver(max_to_keep=3)
# Run Options. Used if necessary
rConf = tf.ConfigProto(log_device_placement=True,
allow_soft_placement=True)
rOpts = tf.RunOptions(report_tensor_allocations_upon_oom=True)
# Run
with tf.Session() as sess:
# Graph is complete
tf.get_default_graph().finalize()
var_sizes = [np.prod(var.shape.as_list()) for var in variables.vars]
print('| Number of parameters:', np.sum(var_sizes), flush=True)
# Initialize variables
if not args.cont: # First time
sess.run(init)
else: # Continue
checkpoint = tf.train.latest_checkpoint(
checkpoint_dir=os.path.join('models',args.dataset))
saver.restore(sess, checkpoint)
print('| Variables restored.', flush=True)
# Initialize average
sess.run(variables.initialize_backups)
# Create contexts
contexts = RunContexts(sess,
train=(graph.use_train_data, variables.use_training_variables),
val=(graph.use_val_data, variables.use_ema_variables))
# Main loop
for step in steps_range:
# Train
with contexts.train:
sess.run( minimize,
feed_dict={
graph.dropouts[0]: args.dropout[0],
graph.dropouts[1]: args.dropout[1],
})
# Renormalization
if args.renormalization:
sess.run( graph.normalization_ops )
# Update average
sess.run( variables.update_op )
# Every log_every steps or at the end
if step % args.log_every == 0 or step == steps_range.stop-1:
# Predict on train set and validation set
with contexts.train:
train_loss, train_regular_loss, train_summaries = sess.run(
(graph.loss, graph.regular_loss, train_summaries_op))
with contexts.val:
val_summaries = predict_all_batches(sess, scalar_summaries)
# Log
print('| Step: ' + str(step) + ', train loss: ' + str(train_loss) +
' ( reg_loss: ' + str(train_regular_loss) + ' )', flush=True)
train_writer.add_summary(train_summaries, step)
val_writer.add_summary(val_summaries, step)
train_writer.flush()
val_writer.flush()
# Save parameters
model_name = 'model-step{}'.format(step)
with contexts.train:
saver.save(sess, os.path.join('models',args.dataset,model_name))
# Save step number
with open('logs/last_step.txt', 'wt') as step_f:
step_f.write(str(step))
def testing(args):
'''\
Test the performances of the net on the test set. Creates the tf Graph for
testing, loads the weights and evaluates the predictions. Parameters are
loaded from last checkpoint: models/<args.dataset>/model.
Args:
args: a namespace object. Run `maxout.py -h' for a list of the available
options.
'''
# Prints
print('| Testing')
print('| Dataset:', args.dataset, flush=True)
# Instantiate the graph
graph = CGraph(args.dataset, seed=args.seed)
# Use it
with graph.graph.as_default():
# Constant seed for debugging
if args.seed:
tf.set_random_seed(args.seed)
# Predict with running averages
variables = EmaVariables(args.ema)
# Create a Saver
saver = tf.train.Saver()
# Run
with tf.Session() as sess:
# Graph is complete
tf.get_default_graph().finalize()
# Restore parameters
checkpoint = tf.train.latest_checkpoint(
checkpoint_dir=os.path.join('models',args.dataset))
saver.restore(sess, checkpoint)
# Create context
contexts = RunContexts(sess,
test_set=(graph.use_test_data, variables.use_ema_variables))
# Predict
with contexts.test_set:
summary = predict_all_batches(sess, (
('accuracy', graph.accuracy),
('loss', graph.loss)
))
# Out
print('| Accuracy:', summary.value[0].simple_value)
print('| Loss:', summary.value[1].simple_value)
def debug(args):
'''\
Debugging
'''
# Prints
print('| Debug')
import pdb
pdb.set_trace()
def predict_all_batches(sess, summaries, feed_dict=None):
'''\
Sometimes the test/validation set can't be predicted as a single batch due
to memory constraints of large nets. This function iterates on all batches
averaging the given metrics until the dataset ends (it should end, since
it's the validation/test set). The dataset is not explicitly passed (we use
initializable iterators).
Args:
sess: an open tf Session
summaries: list of (name, tensor) for each summary to compute
feed_dict: feed_dict argument to pass
Returns:
A Summary
'''
tensors = [tensor for (_, tensor) in summaries]
names = [name for (name, _) in summaries]
values = []
# Compute for each batch
while True:
try:
values.append(sess.run(tensors, feed_dict=feed_dict))
except tf.errors.OutOfRangeError:
break
# Average and create summary
values = np.mean(values, axis=0)
summary_values = [tf.Summary.Value(tag=name, simple_value=val)
for name, val in zip(names, values)]
summary = tf.Summary(value=summary_values)
return summary
def _clear_saved(dataset):
'''\
Removes all files from 'models/<dataset>/', 'logs/train' and 'logs/val'.
Ask for confirmation at terminal.
Args:
dataset: The name of the dataset to use.
'''
# Confirm
print('| Clearing previous savings. Press enter to confirm.')
input()
# To remove
join = os.path.join
model = join('models', dataset)
logs = [join('logs',x) for x in ['train','val','debug']]
paths = [model] + logs
# Rm
for p in paths:
if os.path.exists(p):
shutil.rmtree(p)
def _select_optimizer(args):
'''\
Returns a tf optimizer, initialized with options. See the tf api of these
optimizers to see what options are available for each one.
Args:
args: namespace of options with these fields:
rate: learning rate
optimizer: identifier of the optimizer to use. Choices:
gd: GradientDescentOptimizer
rms: RMSPropOptimizer
adagrad: AdagradOptimizer
adadelta: AdadeltaOptimizer. Use rate=1 and other options.
adam: AdamOptimizer
parameters: a list of `opt=val' options to pass to the constructor.
val is numeric.
decay_after: (STEPS, FACTOR). If not None, after this number of STEPS,
the learning rate decreases by FACTOR.
Returns:
A tf.train.Optimizer, the learning_rate, and the global_step
'''
# Extract parameters
params = dict()
if args.parameters:
for p in args.parameters:
key, val = p.split('=',maxsplit=1)
params[key.strip()] = float(val)
# Learning rate decay
rate = tf.convert_to_tensor(args.rate, name='learning_rate')
step_var = tf.train.create_global_step()
if args.decay_after:
rate = tf.train.exponential_decay(rate,
step_var, args.decay_after[0], args.decay_after[1], staircase=True)
# Select optimizer
if args.optimizer == 'gd':
opt = tf.train.GradientDescentOptimizer(rate, **params)
elif args.optimizer == 'rms':
opt = tf.train.RMSPropOptimizer(rate, **params)
elif args.optimizer == 'adagrad':
opt = tf.train.AdagradOptimizer(rate, **params)
elif args.optimizer == 'adadelta':
opt = tf.train.AdadeltaOptimizer(rate, **params)
elif args.optimizer == 'adam':
opt = tf.train.AdamOptimizer(rate, **params)
else:
raise ValueError(args.optimizer+
' is not an optimizer. See help(maxout._select_optimizer)')
print('| Using', opt.get_name())
return opt, rate, step_var
def main():
'''\
Main function. Called when this file is executed as script.
'''
# Defaults
learning_rate = 0.05
n_steps = 200
log_every = 20
optimizer = 'adam'
seed = 4134631
dataset = 'cifar10'
ema = 0.0 # Off
## Parsing arguments
parser = argparse.ArgumentParser(description='Training and testing with\
the Maxout network')
parser.add_argument('op', choices=['train','test','debug'],
help='What to do with the net. Most options only affect training.')
parser.add_argument('-d', '--dataset', default=dataset,
choices=['example','mnist','cifar10'], help='Which dataset to load')
parser.add_argument('-r', '--rate', type=float, default=learning_rate,
help='Learning rate / step size. Depends on the optimizer.')
parser.add_argument('-s', '--steps', type=int, default=n_steps,
help='Number of steps of the optimization')
parser.add_argument('-l', '--log_every', type=int, default=log_every,
help='Interval of number of steps between logs/saved models')
parser.add_argument('-o', '--optimizer', default=optimizer,
choices=['gd', 'rms', 'adagrad', 'adadelta', 'adam'],
help='Name of the optimizer to use.\
See `help(maxout._select_optimizer)\' to know more.')
parser.add_argument('-p', '--parameters',
nargs='+', metavar='PARAMETER',
help='If the optimizer needs other arguments than just --rate,\
use this option. One or more `opt=val\' for any opt argument of the\
optimizer selected (see tf doc). val is assumed numeric.')
parser.add_argument('-c', '--continue', action='store_true', dest='cont',
help='Loads most recent saved model and resumes training from there.\
Continue with the same optimizer.')
parser.add_argument('--dropout', type=float, nargs=2, metavar=('INPUT_RATE',
'HIDDEN_RATE'),
help='Dropout probability: drop probability for input and hidden units.')
parser.add_argument('-b', '--batch', type=int,
help='Batch size. Without this parameter, the whole dataset is used.')
parser.add_argument('--pseudorand', action='store_const', const=seed,
dest='seed', help='Always use the same seed for reproducible results')
parser.add_argument('--regularization', type=float,
help='Regularization scale. 0 means no regularization')
parser.add_argument('--renormalization', type=float,
help='If set, this is the maximum norm for all vectors in weigts\
matrices')
parser.add_argument('--ema', type=float, default=ema,
help='Running average decay rate (something like 0.99).')
parser.add_argument('--decay_after', nargs=2, metavar=('STEPS', 'FACTOR'),
type=float, help='Number of STEPS after which the learnin rate decays\
by FACTOR')
args = parser.parse_args()
# Some checks
if not args.dropout:
args.dropout = (0,0) # Keep everything
else:
if not (0 <= args.dropout[0] <= 1 and 0 <= args.dropout[1] <= 1):
raise ValueError(
'--dropout arguments is not a probability. It must be in [0, 1].')
if args.op == 'test':
print('Warning: dropout argument is not used in testing.')
if args.renormalization != None and args.renormalization <= 0:
raise ValueError(
'--renormalization must be a positive length.')
if args.ema and not (0 <= args.ema < 1):
raise ValueError(
'--ema must be a decay rate in [0, 1).')
if args.decay_after:
if not args.decay_after[0].is_integer() or args.decay_after[0] <= 0:
raise ValueError(
'--decay_after STEPS must be a positive number of steps (int)')
if not 0 <= args.decay_after[1] <= 1:
raise ValueError(
'--decay_after FACTOR must be in [0, 1].')
# Go
if args.op == 'train':
training(args)
elif args.op == 'test':
testing(args)
elif args.op == 'debug':
debug(args)
if __name__ == '__main__':
main()