-
Notifications
You must be signed in to change notification settings - Fork 2
/
main.py
234 lines (205 loc) · 10.7 KB
/
main.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
import os.path
import warnings
warnings.filterwarnings('ignore')
from coco_eval import model_evaluate, coco_caption_eval
from torch import optim, nn
#import tmp.diffcap_eval as diffcap_eval
from diff_models.diffcap_model import Diffuser, Diffuser_with_LN
from my_utils.blip_util import load_checkpoint
from diff_models.diffusion import *
# from dataload.dataloader import train_loader, val_loader, val_set
from torch.utils.data import DataLoader
from dataload import create_dataset
from torch.nn.parallel import DistributedDataParallel
from tqdm.auto import tqdm
from my_utils.train_util import batch_loss
import time
from transformers import get_linear_schedule_with_warmup
from transformers import get_constant_schedule_with_warmup
from transformers import get_cosine_schedule_with_warmup
from my_utils.detr_object import get_detr_objects
wandb_configs = {
"epochs": EPOCH_NUM,
"batch_size": TRAIN_BATCH_SIZE,
'length': MAX_LENGTH,
}
accelerator.init_trackers('Diff-Cap', config=wandb_configs,
init_kwargs={"wandb": {"name": notes}}) # , "entity": "xxx"}})
if not USING_TIME_LN:
model = Diffuser(image_size=224)
else:
model = Diffuser_with_LN(image_size=224)
data_config = {'image_size':224, 'ann_root':'datasets/COCO/', 'image_root': 'datasets/COCO'}
train_set, val_set, test_set = create_dataset('caption_coco', data_config)
train_loader = DataLoader(train_set, shuffle=True, batch_size=TRAIN_BATCH_SIZE, drop_last=True, num_workers=32)
val_loader = DataLoader(val_set, shuffle=False, batch_size=VAL_BATCH_SIZE, drop_last=True, num_workers=2)
optimizer = optim.AdamW(filter(lambda p: p.requires_grad, model.parameters()), lr=LEARNING_RATE)
num_step = len(train_loader) * EPOCH_NUM
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_step * WARMUP_RATIO,
num_training_steps=num_step)
# scheduler = get_constant_schedule_with_warmup(optimizer, num_warmup_steps=num_step * WARMUP_RATIO)
# scheduler = get_cosine_schedule_with_warmup(optimizer, num_warmup_steps=num_step * WARMUP_RATIO,
# num_training_steps=num_step, num_cycles=2)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
special = tokenizer(['.'], return_tensors='pt')
special_emb = model.space_encoder(special['input_ids'])[0][0][1]
def train_func(model, trainer, x, scheduler, train=True):
x_0 = model.space_encoder(input_ids=x['input_ids'], attention_mask=x['attention_mask'])[0]
# torch.save(torch.mean(x_0, dim=(0,1)), 'datasets/mean_emb_split.pickle')
# torch.save(torch.sqrt(torch.var(x_0, dim=(0, 1))), 'datasets/std_emb_split.pickle')
x_0 = (x_0 - X_MEAN.to(accelerator.device)) / X_SIGMA.to(accelerator.device)
atten_mask = x['attention_mask']
atten_mask = torch.roll(atten_mask, -1, 1)
atten_mask[:, 0] = 0
atten_mask[:, -1] = 0
# change pad cls sep to special token
x_0[atten_mask == 0] = special_emb.to(accelerator.device)
if USE_OBJECT:
# objects_list, objects_ids, objects_mask = get_detr_objects(x['detr_input'])
# objects_ids, objects_mask = objects_ids.to(accelerator.device), objects_mask.to(accelerator.device)
objects_ids, objects_mask = x['objects_ids'].to(accelerator.device), x['objects_mask'].to(accelerator.device)
object_emb = model.space_encoder(input_ids=objects_ids.to(accelerator.device), attention_mask=objects_mask.to(accelerator.device))[0]
object_emb = (object_emb - X_MEAN.to(accelerator.device)) / X_SIGMA.to(accelerator.device)
object_atten_mask = objects_mask
object_atten_mask = torch.roll(object_atten_mask, -1, 1)
object_atten_mask[:, 0] = 0
object_atten_mask[:, -1] = 0
object_emb[object_atten_mask == 0] = special_emb.to(accelerator.device)
# randomly mask some objects
object_rand_mask = torch.rand(object_emb.shape[0]) < OBJECT_MASK_RATIO
object_emb[object_rand_mask==1] = repeat(special_emb, 'd -> seq d', seq=object_emb.shape[1]).to(accelerator.device)
t = torch.randint(0, STEP_TOT, (x_0.shape[0],), device=accelerator.device) # 随机采样batchsize个时间
if X_0_PREDICTION or EPSILON_PRED:
x_t = diffuse_t(x_0, t) # bsz, seqlen, dmodel
x_tgt = None
# else:
# x_t, x_tgt = generate_diffuse_pair(x_0, t, torch.max(t - X_T_STEP_INTERVAL,
# torch.zeros(t.shape, device=accelerator.device,
# dtype=torch.int64)))
x_1 = diffuse_t(x_0, torch.ones(1, dtype=torch.int64, device=accelerator.device))
image, mask = x['image'].to(accelerator.device), x['attention_mask'].to(accelerator.device)
if CLASSIFIER_FREE_PROB > 0:
classifier_mask = (torch.rand(TRAIN_BATCH_SIZE) > CLASSIFIER_FREE_PROB).type(torch.float32).to(
accelerator.device) # generate mask
image = image * (repeat(classifier_mask,'b -> b c h w', c = 3, h = image.shape[2], w=image.shape[3]))
x_pred = torch.zeros_like(x_t)
if USE_OBJECT:
object_emb_selfcond = torch.concat([object_emb, object_emb], dim=-1)
# add self attentioning
if SELF_COND and random.random() > SELF_COND_PROB:
concat_x_t = torch.cat([x_t, x_pred], dim=-1)
concat_x_t = torch.cat([concat_x_t, object_emb_selfcond], dim=-2)
x_pred = model(image, concat_x_t, torch.concat([mask, objects_mask], dim=-1), t)
x_pred = x_pred.detach()
# x_t restore loss
x_pred = model(image, torch.concat([torch.cat([x_t, x_pred], dim=-1), object_emb_selfcond],dim=-2), torch.concat([mask, objects_mask], dim=-1), t)
else:
if SELF_COND and random.random() > SELF_COND_PROB:
concat_x_t = torch.cat([x_t, x_pred], dim=-1)
# concat_x_t = torch.cat([concat_x_t, object_emb_selfcond], dim=-2)
x_pred = model(image, concat_x_t, mask, t)
x_pred = x_pred.detach()
# x_t restore loss
x_pred = model(image, torch.cat([x_t, x_pred], dim=-1), mask, t)
x_t_loss, x_1_loss, prob_loss, valid_token_loss, pad_loss = batch_loss(
model, x_pred, x_t, x_tgt, x_0,
x["attention_mask"],
x["input_ids"],
LOSS_FUNC
)
l = x_t_loss + x_1_loss + prob_loss
if train:
trainer.zero_grad()
accelerator.backward(l)
# accelerator.clip_grad_norm_(model.parameters(), 1.0)
trainer.step()
scheduler.step()
return l, x_t_loss, x_1_loss, prob_loss, valid_token_loss, pad_loss
def validate(model):
val_acc_x_t = 0
val_acc_x_1 = 0
val_acc_prob = 0
val_loss = 0
model.eval()
with torch.no_grad():
for batch_num, x in tqdm(enumerate(val_loader), disable=not accelerator.is_local_main_process):
l, x_t_loss, x_1_loss, prob_loss, valid_token_loss, pad_loss = train_func(model, optimizer, x, scheduler, train=False)
val_acc_x_t += x_t_loss
val_acc_x_1 += x_1_loss
val_acc_prob += prob_loss
val_loss += l
model.train()
return val_loss / len(val_loader), val_acc_x_t / len(val_loader), val_acc_x_1 / len(val_loader), val_acc_prob / len(
val_loader),
model, optimizer, train_loader, scheduler, X_MEAN, X_SIGMA = accelerator.prepare(
model, optimizer, train_loader, scheduler, X_MEAN, X_SIGMA
)
if START_EPOCH > 0:
accelerator.load_state(f'{LOG_DIR}/{MODEL_NAME}/acc_epoch_{START_EPOCH}/')
if isinstance(model, DistributedDataParallel):
model = model.module
# early_stopped = False
######################################################################################
#################### begin training #################################################
######################################################################################
if not os.path.exists(f'{LOG_DIR}/{MODEL_NAME}'):
os.makedirs(f'{LOG_DIR}/{MODEL_NAME}', exist_ok=True)
accelerator.print("start training")
start_time = time.time()
start_epoch = START_EPOCH
model.train()
for epoch in range(start_epoch, EPOCH_NUM):
accelerator.print(f'current epoch{epoch}')
acc_x_t = 0
acc_x_1 = 0
acc_prob = 0
acc_l = 0
accelerator.print("the number of batchs is", len(train_loader))
accelerator.print('before training', (time.time() - start_time) / 60, 'min')
for batch_num, x in tqdm(enumerate(train_loader), disable=not accelerator.is_local_main_process):
l, x_t_loss, x_1_loss, prob_loss, valid_token_loss, pad_loss = train_func(model, optimizer, x, scheduler)
if batch_num % 50 == 0:
accelerator.log({'loss': l,
'x_t_loss': x_t_loss,
'x_1_loss': x_1_loss,
'prob_loss': prob_loss,
'valid_token_loss': valid_token_loss,
'pad_loss': pad_loss}
)
acc_x_t += x_t_loss
acc_x_1 += x_1_loss
acc_prob += prob_loss
acc_l += l
if DYNAMIC_ROUNDING_WEIGHT > 0:
ROUNDING_WEIGHT = ((acc_x_t + acc_x_1) / acc_prob).detach() * DYNAMIC_ROUNDING_WEIGHT
if DEBUG:
break
accelerator.print('after a epoch training', (time.time() - start_time) / 60, 'min')
accelerator.wait_for_everyone()
accelerator.print('after sync', (time.time() - start_time) / 60, 'min')
# l, x_t_loss, x_1_loss, prob_loss, valid_token_loss, pad_loss = validate(model)
# accelerator.log({'val_loss': l,
# 'val_x_t_loss': x_t_loss,
# 'val_x_1_loss': x_1_loss,
# 'val_prob_loss': prob_loss,
# 'val_valid_token_loss': valid_token_loss,
# 'val_pad_loss': pad_loss}
# )
# unwrapped_model = accelerator.unwrap_model(model)
# accelerator.save(unwrapped_model.state_dict(), f"./checkpoint/{MODEL_NAME}/epoch_{epoch}.pickle")
# model = model.to(accelerator.device)
accelerator.save_state(f"{LOG_DIR}/{MODEL_NAME}/acc_epoch_{epoch}/")
accelerator.print('after saving', (time.time() - start_time) / 60, 'min')
accelerator.wait_for_everyone()
# unwrapped_model = accelerator.unwrap_model(model)
# accelerator.save(unwrapped_model.state_dict(), f"./checkpoint/{MODEL_NAME}.pickle")
# model = model.to(accelerator.device)
accelerator.print('Done!')
if accelerator.is_local_main_process:
# bleu = diffcap_eval.evaluate(model, val_set, val_loader)
# accelerator.log({'bleu': bleu})
model_evaluate(model, val_set, val_loader)
if not os.path.exists('result'):
os.makedirs('result', exist_ok=True)
coco_caption_eval('result/', f'result/{RESULT_FILE}.json', split='val')
accelerator.end_training()