-
Notifications
You must be signed in to change notification settings - Fork 21
/
train.py
280 lines (243 loc) · 12.9 KB
/
train.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
# Use tensorboard
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.optim as optim
import time
from six.moves import cPickle
import opts
import models
from dataloader import *
import eval_utils
import misc.utils as utils
try:
import tensorflow as tf
except ImportError:
print("Tensorflow not installed; No tensorboard logging.")
tf = None
def add_summary_value(writer, key, value, iteration):
summary = tf.Summary(value=[tf.Summary.Value(tag=key, simple_value=value)])
writer.add_summary(summary, iteration)
# def unwrap_self(arg, **kwarg):
# return DataLoader.get_batch_one(*arg, **kwarg)
def train(opt):
np.random.seed(42)
warnings.filterwarnings('ignore')
opt.use_att = utils.if_use_att(opt.caption_model)
loader = DataLoader(opt)
opt.vocab_size = loader.vocab_size
opt.seq_length = loader.seq_length
# for debug purposes
# a=get_batch_one(opt, [loader.split_ix, loader.shuffle, loader.iterators, loader.label_start_ix, loader.label_end_ix])
# loader.get_batch('train')
if not os.path.exists(opt.checkpoint_path+'tensorboard/'):
os.makedirs(opt.checkpoint_path+'tensorboard/')
else:
for path in os.listdir(opt.checkpoint_path+'tensorboard/'):
os.remove(opt.checkpoint_path+'tensorboard/'+path)
tf_summary_writer = tf and tf.summary.FileWriter(opt.checkpoint_path+'tensorboard/')
np.random.seed(42)
infos = {}
histories = {}
if opt.start_from is not None:
# open old infos and check if models are compatible
with open(os.path.join(opt.start_from, 'infos_'+opt.id+'.pkl')) as f:
infos = cPickle.load(f)
saved_model_opt = infos['opt']
need_be_same=["caption_model", "rnn_type", "rnn_size", "num_layers"]
for checkme in need_be_same:
assert vars(saved_model_opt)[checkme] == vars(opt)[checkme], "Command line argument and saved model disagree on '%s' " % checkme
if os.path.isfile(os.path.join(opt.start_from, 'histories_'+opt.id+'.pkl')):
with open(os.path.join(opt.start_from, 'histories_'+opt.id+'.pkl')) as f:
histories = cPickle.load(f)
iteration = infos.get('iter', 0)
# iteration = 26540
epoch = infos.get('epoch', 0)
val_result_history = histories.get('val_result_history', {})
loss_history = histories.get('loss_history', {})
lr_history = histories.get('lr_history', {})
ss_prob_history = histories.get('ss_prob_history', {})
loader.iterators = infos.get('iterators', loader.iterators)
if opt.load_best_score == 1:
best_val_score = infos.get('best_val_score', None)
cnn_model = utils.build_cnn(opt)
cnn_model.cuda()
model = models.setup(opt)
model.cuda()
update_lr_flag = True
# Assure in training mode
model.train()
crit = utils.LanguageModelCriterion()
optimizer = optim.Adam(model.parameters(), lr=opt.learning_rate)
if opt.finetune_cnn_after != -1:
# only finetune the layer2 to layer4
cnn_optimizer = optim.Adam([\
{'params': module.parameters()} for module in cnn_model._modules.values()[5:]\
], lr=opt.cnn_learning_rate, weight_decay=opt.cnn_weight_decay)
# Load the optimizer
if vars(opt).get('start_from', None) is not None:
if os.path.isfile(os.path.join(opt.start_from, 'optimizer.pth')):
optimizer.load_state_dict(torch.load(os.path.join(opt.start_from, 'optimizer.pth')))
if opt.finetune_cnn_after != -1 and epoch >= opt.finetune_cnn_after:
if os.path.isfile(os.path.join(opt.start_from, 'optimizer-cnn.pth')):
cnn_optimizer.load_state_dict(torch.load(os.path.join(opt.start_from, 'optimizer-cnn.pth')))
while True:
if update_lr_flag:
# Assign the learning rate
if epoch > opt.learning_rate_decay_start and opt.learning_rate_decay_start >= 0:
frac = (epoch - opt.learning_rate_decay_start) // opt.learning_rate_decay_every
decay_factor = opt.learning_rate_decay_rate ** frac
opt.current_lr = opt.learning_rate * decay_factor
utils.set_lr(optimizer, opt.current_lr) # set the decayed rate
else:
opt.current_lr = opt.learning_rate
# Assign the scheduled sampling prob
if epoch > opt.scheduled_sampling_start and opt.scheduled_sampling_start >= 0:
frac = (epoch - opt.scheduled_sampling_start) // opt.scheduled_sampling_increase_every
opt.ss_prob = min(opt.scheduled_sampling_increase_prob * frac, opt.scheduled_sampling_max_prob)
model.ss_prob = opt.ss_prob
# Update the training stage of cnn
if opt.finetune_cnn_after == -1 or epoch < opt.finetune_cnn_after:
for p in cnn_model.parameters():
p.requires_grad = False
cnn_model.eval()
else:
for p in cnn_model.parameters():
p.requires_grad = True
# Fix the first few layers:
for module in cnn_model._modules.values()[:5]:
for p in module.parameters():
p.requires_grad = False
cnn_model.train()
update_lr_flag = False
# torch.cuda.synchronize()
start = time.time()
# Load data from train split (0)
# for validation training change the split to 'val'
# data = loader.get_batch('val')
data = loader.get_batch('train')
data['images'] = utils.prepro_images(data['images'], True)
# torch.cuda.synchronize()
print('Read data:', time.time() - start)
# torch.cuda.synchronize()
start = time.time()
tmp = [data['images'], data['labels'], data['masks']]
tmp = [Variable(torch.from_numpy(_), requires_grad=False).cuda() for _ in tmp]
images, labels, masks = tmp
att_feats = cnn_model(images).permute(0, 2, 3, 1)
fc_feats = att_feats.mean(2).mean(1)
if not opt.use_att:
att_feats = Variable(torch.FloatTensor(1, 1,1,1).cuda())
att_feats = att_feats.unsqueeze(1).expand(*((att_feats.size(0), opt.seq_per_img,) +
att_feats.size()[1:])).contiguous().view(*((att_feats.size(0) * opt.seq_per_img,)
+ att_feats.size()[1:]))
fc_feats = fc_feats.unsqueeze(1).expand(*((fc_feats.size(0), opt.seq_per_img,) +
fc_feats.size()[1:])).contiguous().view(*((fc_feats.size(0) * opt.seq_per_img,) +
fc_feats.size()[1:]))
model.zero_grad()
optimizer.zero_grad()
if opt.finetune_cnn_after != -1 and epoch >= opt.finetune_cnn_after:
cnn_optimizer.zero_grad()
if opt.sentence_embed:
sen_embed = Variable(torch.from_numpy(np.array(data['sen_embed'])).cuda())
out = model(fc_feats, att_feats, labels, sen_embed)
loss = crit(out, labels[:, 1:], masks[:, 1:])
# loss += cov
else:
loss = crit(model(fc_feats, att_feats, labels), labels[:,1:], masks[:,1:])
# - 0.001 * crit(model(torch.zeros(fc_feats.size()).cuda(), torch.zeros(att_feats.size()).cuda(), labels), labels[:,1:], masks[:,1:])
loss.backward()
# utils.clip_gradient(optimizer, opt.grad_clip)
optimizer.step()
if opt.finetune_cnn_after != -1 and epoch >= opt.finetune_cnn_after:
utils.clip_gradient(cnn_optimizer, opt.grad_clip)
cnn_optimizer.step()
# train_loss = loss.data[0]
train_loss = loss.item()
# torch.cuda.synchronize()
end = time.time()
print("Step [{}/{}], Epoch [{}/{}], train_loss = {:.3f}, time/batch = {:.3f}" \
.format((iteration+1)%int(len(loader)/vars(opt)['batch_size']), int(len(loader)/vars(opt)['batch_size']),
epoch, vars(opt)['max_epochs'], train_loss, end - start))
# Update the iteration and epoch
iteration += 1
if data['bounds']['wrapped']:
epoch += 1
update_lr_flag = True
# Write the training loss summary
if (iteration % opt.losses_log_every == 0):
if tf is not None:
add_summary_value(tf_summary_writer, 'train_loss', train_loss, iteration)
add_summary_value(tf_summary_writer, 'learning_rate', opt.current_lr, iteration)
add_summary_value(tf_summary_writer, 'scheduled_sampling_prob', model.ss_prob, iteration)
tf_summary_writer.flush()
loss_history[iteration] = train_loss
lr_history[iteration] = opt.current_lr
ss_prob_history[iteration] = model.ss_prob
# make evaluation on validation set, and save model
if (iteration % opt.save_checkpoint_every == 0):
# eval model
eval_kwargs = {'split': 'val',
'dataset': opt.input_json}
eval_kwargs.update(vars(opt))
val_loss, predictions, lang_stats = eval_utils.eval_split(cnn_model, model, crit, loader, eval_kwargs)
# Write validation result into summary
if tf is not None:
add_summary_value(tf_summary_writer, 'validation loss', val_loss, iteration)
for k,v in lang_stats.items():
add_summary_value(tf_summary_writer, k, v, iteration)
tf_summary_writer.flush()
val_result_history[iteration] = {'loss': val_loss, 'lang_stats': lang_stats, 'predictions': predictions}
# Save model if is improving on validation result
if opt.language_eval == 1:
current_score = lang_stats['CIDEr']
else:
current_score = - val_loss
best_flag = False
if True: # if true
if best_val_score is None or current_score > best_val_score:
best_val_score = current_score
best_flag = True
checkpoint_path = os.path.join(opt.checkpoint_path + opt.caption_model, 'model.pth')
cnn_checkpoint_path = os.path.join(opt.checkpoint_path + opt.caption_model, 'model-cnn.pth')
torch.save(model.state_dict(), checkpoint_path)
torch.save(cnn_model.state_dict(), cnn_checkpoint_path)
print("model saved to {}".format(checkpoint_path))
print("cnn model saved to {}".format(cnn_checkpoint_path))
optimizer_path = os.path.join(opt.checkpoint_path + opt.caption_model, 'optimizer.pth')
torch.save(optimizer.state_dict(), optimizer_path)
if opt.finetune_cnn_after != -1 and epoch >= opt.finetune_cnn_after:
cnn_optimizer_path = os.path.join(opt.checkpoint_path + opt.caption_model, 'optimizer-cnn.pth')
torch.save(cnn_optimizer.state_dict(), cnn_optimizer_path)
# Dump miscalleous informations
infos['iter'] = iteration
infos['epoch'] = epoch
infos['iterators'] = loader.iterators
infos['best_val_score'] = best_val_score
infos['opt'] = opt
infos['vocab'] = loader.get_vocab()
histories['val_result_history'] = val_result_history
histories['loss_history'] = loss_history
histories['lr_history'] = lr_history
histories['ss_prob_history'] = ss_prob_history
with open(os.path.join(opt.checkpoint_path + opt.caption_model, 'infos_'+opt.id+'.pkl'), 'wb') as f:
cPickle.dump(infos, f)
with open(os.path.join(opt.checkpoint_path + opt.caption_model, 'histories_'+opt.id+'.pkl'), 'wb') as f:
cPickle.dump(histories, f)
if best_flag:
checkpoint_path = os.path.join(opt.checkpoint_path+ opt.caption_model, 'model-best.pth')
cnn_checkpoint_path = os.path.join(opt.checkpoint_path+ opt.caption_model, 'model-cnn-best.pth')
torch.save(model.state_dict(), checkpoint_path)
torch.save(cnn_model.state_dict(), cnn_checkpoint_path)
print("model saved to {}".format(checkpoint_path))
print("cnn model saved to {}".format(cnn_checkpoint_path))
with open(os.path.join(opt.checkpoint_path+ opt.caption_model, 'infos_'+opt.id+'-best.pkl'), 'wb') as f:
cPickle.dump(infos, f)
# Stop if reaching max epochs
if epoch >= opt.max_epochs and opt.max_epochs != -1:
break
opt = opts.parse_opt()
train(opt)