-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_rl.py
304 lines (249 loc) · 11.2 KB
/
train_rl.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
import argparse
import logging
import os
import time
import json
import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
from torch.nn.parallel import DistributedDataParallel
from torch.nn.utils import clip_grad_norm_
from transformers.modeling_bert import BertConfig
from transformers.optimization import AdamW, WarmupCosineSchedule
from config import _C as config
from dataset import COCOCaptionDataset, collate_fn_train
from modeling import Generator, LabelSmoothingLoss
from utils import get_rank, mkdir, synchronize
from utils.checkpointer import Checkpointer
from utils.dataloader import make_data_loader
from utils.logger import setup_logger
from utils.tokenizer import EOS, MASK, PAD, num_tokens, LENGTH, tokenizer
from pycocoevalcap.cider.cider import Cider
from pycocoevalcap.meteor.meteor import Meteor
from pycocoevalcap.tokenizer.ptbtokenizer import PTBTokenizer
class SelfCriticalLoss(torch.nn.Module):
def __init__(self, scorers, ptb_tokenizer, gt_caption, entropy_weight):
super(SelfCriticalLoss, self).__init__()
self.scorers = scorers
self.tokenizer = ptb_tokenizer
self.gt_captions = gt_caption
self.entropy_weight = entropy_weight
def forward(self, new_caption, probs, image_id):
"""
Args:
old_caption: (N, L), long
new_caption: (N, L), long
probs: (N,), float
image_id: (N,), long
mask: (N, L), float
"""
# new_caption = new_caption.cpu().numpy()
# image_id = image_id.cpu().numpy()
ref = dict()
new_hypo = dict()
for new, id in zip(new_caption, image_id):
id = str(id.cpu().numpy())
if id in ref.keys():
id_ = id + '_'
else:
id_ = id
ref[id_] = self.gt_captions[id]
new = tokenizer.decode(new.cpu().numpy(), end_flags=[EOS])
new_hypo[id_] = [{'caption': new}]
new_hypo = self.tokenizer.tokenize(new_hypo)
rewards = 0.0
for (scorer, weight) in self.scorers:
_, new_scores = scorer.compute_score(ref, new_hypo)
rewards += np.asarray(new_scores) * weight
rewards = torch.from_numpy(rewards).to(probs.device).float().unsqueeze(1)
logprobs = probs.log()
entropy = (-logprobs * probs).mean()
ascent_objective = (logprobs * rewards).mean()
ascent_objective += self.entropy_weight * entropy
return -ascent_objective, rewards.mean()
def train(generator, optimizer, data_loader, scheduler, checkpointer,
device, log_time, checkpoint_time, arguments):
logger = logging.getLogger("train")
logger.info("Start training")
max_iter = len(data_loader)
start_iter = arguments['iteration']
generator.train()
if config.loss.balance_weight != 1.0:
balance_weight = torch.ones(
num_tokens, dtype=torch.float32, device=device)
balance_weight[EOS] = config.loss.balance_weight
else:
balance_weight = None
criterion = LabelSmoothingLoss(
num_tokens, balance_weight, config.loss.label_smoothing)
crossEntropyLoss = nn.CrossEntropyLoss()
scorers = [(Cider(), 0.1), (Meteor(), 1.0)]
ptb_tokenizer = PTBTokenizer()
with open(os.path.join(config.data_dir, 'id2captions_train.json')) as f:
gt_captions = json.load(f)
gt_captions = ptb_tokenizer.tokenize(gt_captions)
rl_criterion = SelfCriticalLoss(
scorers, ptb_tokenizer, gt_captions, config.loss.entropy_weight)
end = time.time()
for iteration, batch in enumerate(data_loader, start_iter):
iteration = iteration + 1
arguments['iteration'] = iteration
token_type_ids = batch[0].to(device) # (N, L), long
input_token_ids = batch[1].to(device) # (N, L), long
masked_token_ids = batch[2].to(device) # (N, L), long
region_features = batch[3].to(device) # (N, 100, 2048), float
region_class = batch[4].to(device) # (N, 100, 1601), float
region_spatial = batch[5].to(device) # (N, 100, 6), float
gt_maxlength = batch[6].to(device) # (N), float
image_id = batch[7].to(device)
num_img_tokens = region_spatial.size(1)
seq_length = input_token_ids.size(1)
batch_size = input_token_ids.size(0)
pred_levelp_list = list()
pred_ids_list = list()
region_spatial[:, :, [0, 2]] /= region_spatial[:, :, [2]] + 1e-5
region_spatial[:, :, [1, 3]] /= region_spatial[:, :, [3]] + 1e-5
rel_area = (region_spatial[:, :, [3]] - region_spatial[:, :, [1]]) * \
(region_spatial[:, :, [2]] - region_spatial[:, :, [0]])
region_spatial = torch.cat((region_spatial[:, :, :4],
rel_area.clamp_(0), region_spatial[:, :, 5:]), dim=-1)
position_features = torch.cat((F.layer_norm(region_spatial, [6]),
F.layer_norm(region_class, [1601])), dim=-1)
position_ids = torch.arange(seq_length, dtype=torch.long, device=device)
position_ids = position_ids.unsqueeze(0).expand_as(input_token_ids)
region_type = position_ids.new_full(
region_features.shape[:2], len(config.boundaries) + 1)
'''
token_type_ids = torch.cat((region_type, token_type_ids), dim=1)
attention_mask = (masked_token_ids != PAD).float()
_attention_mask = attention_mask.new_ones((batch_size, num_img_tokens))
attention_mask = torch.cat((_attention_mask, attention_mask), dim=1)
mask_position = (masked_token_ids == MASK).to(torch.long).view(-1)
mask_position = mask_position.nonzero().squeeze()
'''
for l, (low, high) in enumerate(config.boundaries, 1):
token_type_ids = region_class.new_full((batch_size, high), l, dtype=torch.long)
masked_token_ids = token_type_ids.new_full((batch_size, high), MASK)
attention_mask = rel_area.new_ones((batch_size, high + num_img_tokens))
position_ids = torch.arange(high, dtype=torch.long, device=device)
position_ids = position_ids.unsqueeze(0).expand_as(masked_token_ids)
token_type_ids = torch.cat((region_type, token_type_ids), dim=1)
pred_scores, length_score = generator(
region_features, position_features,
masked_token_ids, token_type_ids,
position_ids, attention_mask)
pred_scores = pred_scores[:, num_img_tokens:, :]
#pred_scores = pred_scores.contiguous().view(-1, num_tokens)
#pred_scores = pred_scores[mask_position]
#gt_length = token_type_ids[:, 100]
#gt_token_ids = input_token_ids[:,1:]#.contiguous().view(-1)[mask_position]
_, pred_token_ids = F.softmax(pred_scores, dim=-1).max(dim=-1)
#pred_token_ids.
pred_zeros = torch.zeros(pred_token_ids.shape[0], 25).to(pred_token_ids.device)
pred_zeros[:, :pred_token_ids.shape[1]] = pred_token_ids
pred_levelp_list.append(length_score)
pred_ids_list.append(pred_zeros)
pred_levelp_list = torch.stack(pred_levelp_list, 1)
levelp_max_fin, levelp_max = torch.max(pred_levelp_list,1) #b*1
pred_ids_fin = [pred_ids_list[levelp_max[i,0]][i] for i in range(len(levelp_max))]
#pred_ids_fin = [pred_ids_list[levelp_max[0, 0]][0] for i in range(len(levelp_max))]
masker_loss, masker_reward = rl_criterion(pred_ids_fin, levelp_max_fin, image_id)
#loss_length = crossEntropyLoss(length_score, gt_maxlength-1)
#loss = loss_words + masker_loss
loss = masker_loss
optimizer.zero_grad()
loss.backward()
clip_grad_norm_(generator.parameters(), config.solver.grad_clip)
optimizer.step()
scheduler.step()
batch_time = time.time() - end
end = time.time()
if iteration % log_time == 0 or iteration == max_iter:
logger.info(
' '.join([
"iter: {iter}", "time: {time:.4f}", "mem: {mem:.2f}",
"lr: {lr:.8f}","loss: {loss:.4f}"
]).format(
iter=iteration, time=batch_time, loss=loss,
#loss_words = loss_words, #loss_length = loss_length,
lr=optimizer.param_groups[0]["lr"],
mem=torch.cuda.max_memory_allocated() / 1024.0 ** 3,
))
if iteration % checkpoint_time == 0 or iteration == max_iter:
checkpointer.save("model_{:07d}".format(iteration), **arguments)
if __name__ == "__main__":
os.environ["PATH"] += ':/home/dingning/anaconda3/envs/vlp/bin'
parser = argparse.ArgumentParser(description="train")
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument("opts", default=None, nargs=argparse.REMAINDER)
args = parser.parse_args()
if config.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group("nccl", init_method="env://")
synchronize()
config.merge_from_list(args.opts)
config.freeze()
save_dir = os.path.join(config.save_dir, f'train')
mkdir(save_dir)
logger = setup_logger("train", save_dir, get_rank())
logger.info("Running with config:\n{}".format(config))
arguments = {'iteration': 0}
device = torch.device(config.device)
bert_config = BertConfig(type_vocab_size=len(config.boundaries) + 2)
generator = Generator(bert_config)
generator = generator.to(device)
optimizer = AdamW(
params=generator.parameters(),
lr=config.solver.lr,
weight_decay=config.solver.weight_decay,
betas=config.solver.betas
)
scheduler = WarmupCosineSchedule(
optimizer=optimizer,
warmup_steps=config.scheduler.warmup_steps,
t_total=config.scheduler.max_steps
)
checkpointer = Checkpointer(
model=generator,
optimizer=optimizer,
scheduler=scheduler,
save_dir=save_dir,
save_to_disk=get_rank() == 0,
logger=logger
)
if config.model_path == '':
generator.load_weights(config.pretrained_bert)
else:
extra_checkpoint_data = checkpointer.load(config.model_path)
arguments.update(extra_checkpoint_data)
dataset = COCOCaptionDataset(
root=config.data_dir,
split='trainrestval',
boundaries=config.boundaries,
)
data_loader = make_data_loader(
dataset=dataset,
collate_fn=collate_fn_train,
batch_size=config.samples_per_gpu,
num_workers=config.num_workers,
max_iter=config.scheduler.max_steps,
split='trainrestval',
is_distributed=config.distributed,
start_iter=arguments['iteration'],
)
if config.distributed:
generator = DistributedDataParallel(
module=generator,
device_ids=[args.local_rank],
output_device=args.local_rank,
)
train(generator=generator,
optimizer=optimizer,
data_loader=data_loader,
scheduler=scheduler,
checkpointer=checkpointer,
device=device,
log_time=config.log_time,
checkpoint_time=config.checkpoint_time,
arguments=arguments)