-
Notifications
You must be signed in to change notification settings - Fork 1
/
image_caption_demo.py
173 lines (129 loc) · 7.54 KB
/
image_caption_demo.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
import os
from config import get_args
import torch
import torch.nn as nn
import numpy as np
import logging
from data import DPC2022, DPC2022_for_generate
from torch.utils.data import DataLoader
import pandas as pd
# from transformers import AdamW
from models.transformer import Transformer_visualgpt, VisualEncoder, ScaledDotProductAttention
from loss import BatchWeightedCE
import random
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
from train import train_for_epoch
from eval import eval_for_epoch
if __name__ == '__main__':
args = get_args()
print(args)
# logging 初始化要在distribute之前
logging.basicConfig(filename=os.path.join(args.log_dir, args.batch_weighted_ce, args.use_model+'_image_caption.txt'),
level=logging.INFO,
datefmt='%a, %d %b %Y %H:%M:%S'
)
logging.info(args)
dist.init_process_group("nccl")
rank = dist.get_rank()
device = torch.device(f'cuda:{rank}')
torch.cuda.set_device(rank)
# gpt2_tokenizer.add_special_tokens({'sep_token': "<|sepftext|>"})
# gpt2_tokenizer.add_special_tokens({'pad_token': "+="})
# Model and dataloaders
encoder =VisualEncoder(3, 0, attention_module=ScaledDotProductAttention)
model = Transformer_visualgpt(encoder)
optimizer = torch.optim.AdamW(model.parameters(),lr=args.lr, betas=(0.9, 0.999), eps=1e-8)
gen_ce_loss_fn = nn.NLLLoss(ignore_index=model.tokenizer.pad_token_id)
loss_fn = BatchWeightedCE(ignore_index = model.tokenizer.pad_token_id, args=args)
train_split_by_val_scored_comment_id_pair_df = pd.read_csv(os.path.join(args.comments_root, 'train_split_by_val_scored_comment_id_pair.csv'))
val_df = pd.read_csv(os.path.join(args.test_and_val_root, 'val.csv'))
train_dataset = DPC2022(train_split_by_val_scored_comment_id_pair_df, model.tokenizer, args)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
batch_trainDataLoader = DataLoader(train_dataset, batch_size=args.train_batch_size, shuffle=False, num_workers=4, collate_fn=train_dataset.collate_fn, drop_last=True, sampler=train_sampler)
val_dataset = DPC2022_for_generate(val_df, model.tokenizer , args)
val_sampler = torch.utils.data.distributed.DistributedSampler(val_dataset)
batch_valDataLoader = DataLoader(val_dataset, batch_size=args.val_batch_size, shuffle=False, num_workers=4, collate_fn=val_dataset.collate_fn, drop_last=False, sampler=val_sampler)
start_epoch = 0
eval_result_dict_list = []
if args.train_from_scratch == True:
print('training from scratch')
else:
print('training from last')
model_fname = '%s/%s/%s_image_caption_last_model.pth' % (args.models_dir, args.batch_weighted_ce ,args.use_model)
# model_fname = '%s/%s/%s_image_caption_epoch_14_model.pth' % (args.models_dir, args.batch_weighted_ce ,args.use_model)
if os.path.exists(model_fname):
# model_checkpoint = torch.load(model_fname, map_location='cuda:0')
model_checkpoint = torch.load(model_fname, map_location={'cuda:0':f'cuda:{rank}'})
torch.set_rng_state(model_checkpoint['torch_rng_state'].cpu())
torch.cuda.set_rng_state(model_checkpoint['cuda_rng_state'].cpu())
np.random.set_state(model_checkpoint['numpy_rng_state'])
random.setstate(model_checkpoint['random_rng_state'])
model.load_state_dict(model_checkpoint['state_dict'], strict=False)
optimizer.load_state_dict(model_checkpoint['optimizer'])
# 要把optimizer参数放到gpu, 不然报错
for state in optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.cuda()
start_epoch = model_checkpoint['epoch']
print(f"start_epoch:{start_epoch}, eval_result_dict_list:{model_checkpoint['eval_result_dict_list']}")
# print('generator loadingg %s' %(model_fname))
else:
print('no exits %s, training from scratch' % (model_fname))
# 等待加载完成
# dist.barrier()
# 多gpu
model.cuda()
# model=torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
model = DDP(model, device_ids=[rank], output_device=rank, find_unused_parameters=True)
training_loss = 0
world_size = dist.get_world_size()
for epoch in range(start_epoch, start_epoch+args.max_epoch):
eval_result_dict_list = []
training_loss = train_for_epoch(model, batch_trainDataLoader, loss_fn, optimizer, epoch, args)
print('epoch %d, model training_loss : %.4f'%(epoch, training_loss))
if epoch >= start_epoch+0:
beam_eval_result_dict = eval_for_epoch(model, batch_valDataLoader, epoch, 5, args)
eval_result_dict_list.append({'num_beams':5, 'eval_result_dict':beam_eval_result_dict})
dist.barrier()
# beam_eval_result_dict = eval_for_epoch(model, batch_valDataLoader, epoch, 128, args)
# eval_result_dict_list.append({'num_beams':128, 'eval_result_dict':beam_eval_result_dict})
# dist.barrier()
rank = dist.get_rank()
if rank == 0:
# All processes should see same parameters as they all start from same
# random parameters and gradients are synchronized in backward passes.
# Therefore, saving it in one process is sufficient.
logging.info('epoch %d --- caption evaluation ---'%(epoch))
logging.info('epoch %d, training_loss : %.4f'%(epoch, training_loss))
print('epoch %d, training_loss : %.4f'%(epoch, training_loss))
for item in eval_result_dict_list:
logging.info(f'epoch {epoch}, num_beams {item["num_beams"]}')
eval_result_dict = item['eval_result_dict']
for metric_dict in eval_result_dict.values():
for metric, score in metric_dict.items():
print(f'{metric}: {metric_dict[metric]:.4f}')
logging.info('epoch %d, %s : %.4f'%(epoch, metric, metric_dict[metric]))
torch.save({
'torch_rng_state': torch.get_rng_state(),
'cuda_rng_state': torch.cuda.get_rng_state(),
'numpy_rng_state': np.random.get_state(),
'random_rng_state': random.getstate(),
'state_dict': model.module.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch,
'eval_result_dict_list': eval_result_dict_list,
}, '%s/%s/%s_image_caption_last_model.pth' % (args.models_dir, args.batch_weighted_ce ,args.use_model))
torch.save({
'torch_rng_state': torch.get_rng_state(),
'cuda_rng_state': torch.cuda.get_rng_state(),
'numpy_rng_state': np.random.get_state(),
'random_rng_state': random.getstate(),
'state_dict': model.module.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch,
'eval_result_dict_list': eval_result_dict_list,
}, '%s/%s/%s_image_caption_epoch_%d_model.pth' % (args.models_dir, args.batch_weighted_ce ,args.use_model, epoch))
# 等待模型保存完毕
dist.barrier()