This repository has been archived by the owner on Oct 31, 2023. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 211
/
train_model.py
490 lines (425 loc) · 18.4 KB
/
train_model.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
#!/usr/bin/env python3
# Copyright 2017-present, Facebook, Inc.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import sys
import os
import argparse
import json
import random
import shutil
import torch
torch.backends.cudnn.enabled = True
from torch.autograd import Variable
import torch.nn.functional as F
import numpy as np
import h5py
import iep.utils as utils
import iep.preprocess
from iep.data import ClevrDataset, ClevrDataLoader
from iep.models import ModuleNet, Seq2Seq, LstmModel, CnnLstmModel, CnnLstmSaModel
parser = argparse.ArgumentParser()
# Input data
parser.add_argument('--train_question_h5', default='data/train_questions.h5')
parser.add_argument('--train_features_h5', default='data/train_features.h5')
parser.add_argument('--val_question_h5', default='data/val_questions.h5')
parser.add_argument('--val_features_h5', default='data/val_features.h5')
parser.add_argument('--feature_dim', default='1024,14,14')
parser.add_argument('--vocab_json', default='data/vocab.json')
parser.add_argument('--loader_num_workers', type=int, default=1)
parser.add_argument('--use_local_copies', default=0, type=int)
parser.add_argument('--cleanup_local_copies', default=1, type=int)
parser.add_argument('--family_split_file', default=None)
parser.add_argument('--num_train_samples', default=None, type=int)
parser.add_argument('--num_val_samples', default=10000, type=int)
parser.add_argument('--shuffle_train_data', default=1, type=int)
# What type of model to use and which parts to train
parser.add_argument('--model_type', default='PG',
choices=['PG', 'EE', 'PG+EE', 'LSTM', 'CNN+LSTM', 'CNN+LSTM+SA'])
parser.add_argument('--train_program_generator', default=1, type=int)
parser.add_argument('--train_execution_engine', default=1, type=int)
parser.add_argument('--baseline_train_only_rnn', default=0, type=int)
# Start from an existing checkpoint
parser.add_argument('--program_generator_start_from', default=None)
parser.add_argument('--execution_engine_start_from', default=None)
parser.add_argument('--baseline_start_from', default=None)
# RNN options
parser.add_argument('--rnn_wordvec_dim', default=300, type=int)
parser.add_argument('--rnn_hidden_dim', default=256, type=int)
parser.add_argument('--rnn_num_layers', default=2, type=int)
parser.add_argument('--rnn_dropout', default=0, type=float)
# Module net options
parser.add_argument('--module_stem_num_layers', default=2, type=int)
parser.add_argument('--module_stem_batchnorm', default=0, type=int)
parser.add_argument('--module_dim', default=128, type=int)
parser.add_argument('--module_residual', default=1, type=int)
parser.add_argument('--module_batchnorm', default=0, type=int)
# CNN options (for baselines)
parser.add_argument('--cnn_res_block_dim', default=128, type=int)
parser.add_argument('--cnn_num_res_blocks', default=0, type=int)
parser.add_argument('--cnn_proj_dim', default=512, type=int)
parser.add_argument('--cnn_pooling', default='maxpool2',
choices=['none', 'maxpool2'])
# Stacked-Attention options
parser.add_argument('--stacked_attn_dim', default=512, type=int)
parser.add_argument('--num_stacked_attn', default=2, type=int)
# Classifier options
parser.add_argument('--classifier_proj_dim', default=512, type=int)
parser.add_argument('--classifier_downsample', default='maxpool2',
choices=['maxpool2', 'maxpool4', 'none'])
parser.add_argument('--classifier_fc_dims', default='1024')
parser.add_argument('--classifier_batchnorm', default=0, type=int)
parser.add_argument('--classifier_dropout', default=0, type=float)
# Optimization options
parser.add_argument('--batch_size', default=64, type=int)
parser.add_argument('--num_iterations', default=100000, type=int)
parser.add_argument('--learning_rate', default=5e-4, type=float)
parser.add_argument('--reward_decay', default=0.9, type=float)
# Output options
parser.add_argument('--checkpoint_path', default='data/checkpoint.pt')
parser.add_argument('--randomize_checkpoint_path', type=int, default=0)
parser.add_argument('--record_loss_every', type=int, default=1)
parser.add_argument('--checkpoint_every', default=10000, type=int)
def main(args):
if args.randomize_checkpoint_path == 1:
name, ext = os.path.splitext(args.checkpoint_path)
num = random.randint(1, 1000000)
args.checkpoint_path = '%s_%06d%s' % (name, num, ext)
vocab = utils.load_vocab(args.vocab_json)
if args.use_local_copies == 1:
shutil.copy(args.train_question_h5, '/tmp/train_questions.h5')
shutil.copy(args.train_features_h5, '/tmp/train_features.h5')
shutil.copy(args.val_question_h5, '/tmp/val_questions.h5')
shutil.copy(args.val_features_h5, '/tmp/val_features.h5')
args.train_question_h5 = '/tmp/train_questions.h5'
args.train_features_h5 = '/tmp/train_features.h5'
args.val_question_h5 = '/tmp/val_questions.h5'
args.val_features_h5 = '/tmp/val_features.h5'
question_families = None
if args.family_split_file is not None:
with open(args.family_split_file, 'r') as f:
question_families = json.load(f)
train_loader_kwargs = {
'question_h5': args.train_question_h5,
'feature_h5': args.train_features_h5,
'vocab': vocab,
'batch_size': args.batch_size,
'shuffle': args.shuffle_train_data == 1,
'question_families': question_families,
'max_samples': args.num_train_samples,
'num_workers': args.loader_num_workers,
}
val_loader_kwargs = {
'question_h5': args.val_question_h5,
'feature_h5': args.val_features_h5,
'vocab': vocab,
'batch_size': args.batch_size,
'question_families': question_families,
'max_samples': args.num_val_samples,
'num_workers': args.loader_num_workers,
}
with ClevrDataLoader(**train_loader_kwargs) as train_loader, \
ClevrDataLoader(**val_loader_kwargs) as val_loader:
train_loop(args, train_loader, val_loader)
if args.use_local_copies == 1 and args.cleanup_local_copies == 1:
os.remove('/tmp/train_questions.h5')
os.remove('/tmp/train_features.h5')
os.remove('/tmp/val_questions.h5')
os.remove('/tmp/val_features.h5')
def train_loop(args, train_loader, val_loader):
vocab = utils.load_vocab(args.vocab_json)
program_generator, pg_kwargs, pg_optimizer = None, None, None
execution_engine, ee_kwargs, ee_optimizer = None, None, None
baseline_model, baseline_kwargs, baseline_optimizer = None, None, None
baseline_type = None
pg_best_state, ee_best_state, baseline_best_state = None, None, None
# Set up model
if args.model_type == 'PG' or args.model_type == 'PG+EE':
program_generator, pg_kwargs = get_program_generator(args)
pg_optimizer = torch.optim.Adam(program_generator.parameters(),
lr=args.learning_rate)
print('Here is the program generator:')
print(program_generator)
if args.model_type == 'EE' or args.model_type == 'PG+EE':
execution_engine, ee_kwargs = get_execution_engine(args)
ee_optimizer = torch.optim.Adam(execution_engine.parameters(),
lr=args.learning_rate)
print('Here is the execution engine:')
print(execution_engine)
if args.model_type in ['LSTM', 'CNN+LSTM', 'CNN+LSTM+SA']:
baseline_model, baseline_kwargs = get_baseline_model(args)
params = baseline_model.parameters()
if args.baseline_train_only_rnn == 1:
params = baseline_model.rnn.parameters()
baseline_optimizer = torch.optim.Adam(params, lr=args.learning_rate)
print('Here is the baseline model')
print(baseline_model)
baseline_type = args.model_type
loss_fn = torch.nn.CrossEntropyLoss().cuda()
stats = {
'train_losses': [], 'train_rewards': [], 'train_losses_ts': [],
'train_accs': [], 'val_accs': [], 'val_accs_ts': [],
'best_val_acc': -1, 'model_t': 0,
}
t, epoch, reward_moving_average = 0, 0, 0
set_mode('train', [program_generator, execution_engine, baseline_model])
print('train_loader has %d samples' % len(train_loader.dataset))
print('val_loader has %d samples' % len(val_loader.dataset))
while t < args.num_iterations:
epoch += 1
print('Starting epoch %d' % epoch)
for batch in train_loader:
t += 1
questions, _, feats, answers, programs, _ = batch
questions_var = Variable(questions.cuda())
feats_var = Variable(feats.cuda())
answers_var = Variable(answers.cuda())
if programs[0] is not None:
programs_var = Variable(programs.cuda())
reward = None
if args.model_type == 'PG':
# Train program generator with ground-truth programs
pg_optimizer.zero_grad()
loss = program_generator(questions_var, programs_var)
loss.backward()
pg_optimizer.step()
elif args.model_type == 'EE':
# Train execution engine with ground-truth programs
ee_optimizer.zero_grad()
scores = execution_engine(feats_var, programs_var)
loss = loss_fn(scores, answers_var)
loss.backward()
ee_optimizer.step()
elif args.model_type in ['LSTM', 'CNN+LSTM', 'CNN+LSTM+SA']:
baseline_optimizer.zero_grad()
baseline_model.zero_grad()
scores = baseline_model(questions_var, feats_var)
loss = loss_fn(scores, answers_var)
loss.backward()
baseline_optimizer.step()
elif args.model_type == 'PG+EE':
programs_pred = program_generator.reinforce_sample(questions_var)
scores = execution_engine(feats_var, programs_pred)
loss = loss_fn(scores, answers_var)
_, preds = scores.data.cpu().max(1)
raw_reward = (preds == answers).float()
reward_moving_average *= args.reward_decay
reward_moving_average += (1.0 - args.reward_decay) * raw_reward.mean()
centered_reward = raw_reward - reward_moving_average
if args.train_execution_engine == 1:
ee_optimizer.zero_grad()
loss.backward()
ee_optimizer.step()
if args.train_program_generator == 1:
pg_optimizer.zero_grad()
program_generator.reinforce_backward(centered_reward.cuda())
pg_optimizer.step()
if t % args.record_loss_every == 0:
print(t, loss.data[0])
stats['train_losses'].append(loss.data[0])
stats['train_losses_ts'].append(t)
if reward is not None:
stats['train_rewards'].append(reward)
if t % args.checkpoint_every == 0:
print('Checking training accuracy ... ')
train_acc = check_accuracy(args, program_generator, execution_engine,
baseline_model, train_loader)
print('train accuracy is', train_acc)
print('Checking validation accuracy ...')
val_acc = check_accuracy(args, program_generator, execution_engine,
baseline_model, val_loader)
print('val accuracy is ', val_acc)
stats['train_accs'].append(train_acc)
stats['val_accs'].append(val_acc)
stats['val_accs_ts'].append(t)
if val_acc > stats['best_val_acc']:
stats['best_val_acc'] = val_acc
stats['model_t'] = t
best_pg_state = get_state(program_generator)
best_ee_state = get_state(execution_engine)
best_baseline_state = get_state(baseline_model)
checkpoint = {
'args': args.__dict__,
'program_generator_kwargs': pg_kwargs,
'program_generator_state': best_pg_state,
'execution_engine_kwargs': ee_kwargs,
'execution_engine_state': best_ee_state,
'baseline_kwargs': baseline_kwargs,
'baseline_state': best_baseline_state,
'baseline_type': baseline_type,
'vocab': vocab
}
for k, v in stats.items():
checkpoint[k] = v
print('Saving checkpoint to %s' % args.checkpoint_path)
torch.save(checkpoint, args.checkpoint_path)
del checkpoint['program_generator_state']
del checkpoint['execution_engine_state']
del checkpoint['baseline_state']
with open(args.checkpoint_path + '.json', 'w') as f:
json.dump(checkpoint, f)
if t == args.num_iterations:
break
def parse_int_list(s):
return tuple(int(n) for n in s.split(','))
def get_state(m):
if m is None:
return None
state = {}
for k, v in m.state_dict().items():
state[k] = v.clone()
return state
def get_program_generator(args):
vocab = utils.load_vocab(args.vocab_json)
if args.program_generator_start_from is not None:
pg, kwargs = utils.load_program_generator(args.program_generator_start_from)
cur_vocab_size = pg.encoder_embed.weight.size(0)
if cur_vocab_size != len(vocab['question_token_to_idx']):
print('Expanding vocabulary of program generator')
pg.expand_encoder_vocab(vocab['question_token_to_idx'])
kwargs['encoder_vocab_size'] = len(vocab['question_token_to_idx'])
else:
kwargs = {
'encoder_vocab_size': len(vocab['question_token_to_idx']),
'decoder_vocab_size': len(vocab['program_token_to_idx']),
'wordvec_dim': args.rnn_wordvec_dim,
'hidden_dim': args.rnn_hidden_dim,
'rnn_num_layers': args.rnn_num_layers,
'rnn_dropout': args.rnn_dropout,
}
pg = Seq2Seq(**kwargs)
pg.cuda()
pg.train()
return pg, kwargs
def get_execution_engine(args):
vocab = utils.load_vocab(args.vocab_json)
if args.execution_engine_start_from is not None:
ee, kwargs = utils.load_execution_engine(args.execution_engine_start_from)
# TODO: Adjust vocab?
else:
kwargs = {
'vocab': vocab,
'feature_dim': parse_int_list(args.feature_dim),
'stem_batchnorm': args.module_stem_batchnorm == 1,
'stem_num_layers': args.module_stem_num_layers,
'module_dim': args.module_dim,
'module_residual': args.module_residual == 1,
'module_batchnorm': args.module_batchnorm == 1,
'classifier_proj_dim': args.classifier_proj_dim,
'classifier_downsample': args.classifier_downsample,
'classifier_fc_layers': parse_int_list(args.classifier_fc_dims),
'classifier_batchnorm': args.classifier_batchnorm == 1,
'classifier_dropout': args.classifier_dropout,
}
ee = ModuleNet(**kwargs)
ee.cuda()
ee.train()
return ee, kwargs
def get_baseline_model(args):
vocab = utils.load_vocab(args.vocab_json)
if args.baseline_start_from is not None:
model, kwargs = utils.load_baseline(args.baseline_start_from)
elif args.model_type == 'LSTM':
kwargs = {
'vocab': vocab,
'rnn_wordvec_dim': args.rnn_wordvec_dim,
'rnn_dim': args.rnn_hidden_dim,
'rnn_num_layers': args.rnn_num_layers,
'rnn_dropout': args.rnn_dropout,
'fc_dims': parse_int_list(args.classifier_fc_dims),
'fc_use_batchnorm': args.classifier_batchnorm == 1,
'fc_dropout': args.classifier_dropout,
}
model = LstmModel(**kwargs)
elif args.model_type == 'CNN+LSTM':
kwargs = {
'vocab': vocab,
'rnn_wordvec_dim': args.rnn_wordvec_dim,
'rnn_dim': args.rnn_hidden_dim,
'rnn_num_layers': args.rnn_num_layers,
'rnn_dropout': args.rnn_dropout,
'cnn_feat_dim': parse_int_list(args.feature_dim),
'cnn_num_res_blocks': args.cnn_num_res_blocks,
'cnn_res_block_dim': args.cnn_res_block_dim,
'cnn_proj_dim': args.cnn_proj_dim,
'cnn_pooling': args.cnn_pooling,
'fc_dims': parse_int_list(args.classifier_fc_dims),
'fc_use_batchnorm': args.classifier_batchnorm == 1,
'fc_dropout': args.classifier_dropout,
}
model = CnnLstmModel(**kwargs)
elif args.model_type == 'CNN+LSTM+SA':
kwargs = {
'vocab': vocab,
'rnn_wordvec_dim': args.rnn_wordvec_dim,
'rnn_dim': args.rnn_hidden_dim,
'rnn_num_layers': args.rnn_num_layers,
'rnn_dropout': args.rnn_dropout,
'cnn_feat_dim': parse_int_list(args.feature_dim),
'stacked_attn_dim': args.stacked_attn_dim,
'num_stacked_attn': args.num_stacked_attn,
'fc_dims': parse_int_list(args.classifier_fc_dims),
'fc_use_batchnorm': args.classifier_batchnorm == 1,
'fc_dropout': args.classifier_dropout,
}
model = CnnLstmSaModel(**kwargs)
if model.rnn.token_to_idx != vocab['question_token_to_idx']:
# Make sure new vocab is superset of old
for k, v in model.rnn.token_to_idx.items():
assert k in vocab['question_token_to_idx']
assert vocab['question_token_to_idx'][k] == v
for token, idx in vocab['question_token_to_idx'].items():
model.rnn.token_to_idx[token] = idx
kwargs['vocab'] = vocab
model.rnn.expand_vocab(vocab['question_token_to_idx'])
model.cuda()
model.train()
return model, kwargs
def set_mode(mode, models):
assert mode in ['train', 'eval']
for m in models:
if m is None: continue
if mode == 'train': m.train()
if mode == 'eval': m.eval()
def check_accuracy(args, program_generator, execution_engine, baseline_model, loader):
set_mode('eval', [program_generator, execution_engine, baseline_model])
num_correct, num_samples = 0, 0
for batch in loader:
questions, _, feats, answers, programs, _ = batch
questions_var = Variable(questions.cuda(), volatile=True)
feats_var = Variable(feats.cuda(), volatile=True)
answers_var = Variable(feats.cuda(), volatile=True)
if programs[0] is not None:
programs_var = Variable(programs.cuda(), volatile=True)
scores = None # Use this for everything but PG
if args.model_type == 'PG':
vocab = utils.load_vocab(args.vocab_json)
for i in range(questions.size(0)):
program_pred = program_generator.sample(Variable(questions[i:i+1].cuda(), volatile=True))
program_pred_str = iep.preprocess.decode(program_pred, vocab['program_idx_to_token'])
program_str = iep.preprocess.decode(programs[i], vocab['program_idx_to_token'])
if program_pred_str == program_str:
num_correct += 1
num_samples += 1
elif args.model_type == 'EE':
scores = execution_engine(feats_var, programs_var)
elif args.model_type == 'PG+EE':
programs_pred = program_generator.reinforce_sample(
questions_var, argmax=True)
scores = execution_engine(feats_var, programs_pred)
elif args.model_type in ['LSTM', 'CNN+LSTM', 'CNN+LSTM+SA']:
scores = baseline_model(questions_var, feats_var)
if scores is not None:
_, preds = scores.data.cpu().max(1)
num_correct += (preds == answers).sum()
num_samples += preds.size(0)
if num_samples >= args.num_val_samples:
break
set_mode('train', [program_generator, execution_engine, baseline_model])
acc = float(num_correct) / num_samples
return acc
if __name__ == '__main__':
args = parser.parse_args()
main(args)