-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathmain.py
170 lines (144 loc) · 10.5 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
import torch
import argparse
import ruamel_yaml as yaml
import numpy as np
from generation_api.metrics import compute_scores
from generation_api.optimizers import build_optimizer_blip, build_lr_scheduler
from generation_api.trainer_blip import Trainer
from generation_api.loss import compute_loss
from transformers import BertTokenizer
from generation_api.tokenizers_blip import Tokenizer
from models.blip import blip_decoder
from blip_original import create_loader, create_dataset
import os
from transformers import AutoTokenizer, AutoModel
def main(args, config):
torch.manual_seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
np.random.seed(args.seed)
# create tokenizer
if args.bert == 'base':
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
elif args.bert == 'sci':
tokenizer = AutoTokenizer.from_pretrained('allenai/scibert_scivocab_uncased')
elif args.bert == 'cli':
tokenizer = AutoTokenizer.from_pretrained('emilyalsentzer/Bio_ClinicalBERT')
tokenizer.add_special_tokens({'bos_token': '[DEC]'})
tokenizer.add_special_tokens({'additional_special_tokens': ['[ENC]']})
tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0]
# tokenizer = BertTokenizer.from_pretrained(args.text_encoder)
train_dataset, val_dataset, test_dataset = create_dataset('generation_%s'%args.dataset_name, args, config)
samplers = [None, None, None]
train_dataloader, val_dataloader, test_dataloader = create_loader([train_dataset, val_dataset, test_dataset], samplers,
batch_size=[args.batch_size] * 3,
num_workers=[4, 4, 4],
is_trains=[True, False, False],
collate_fns=[None, None, None])
model = blip_decoder(pretrained=args.pretrained, image_size=config['image_size'], vit=config['vit'],
vit_grad_ckpt=config['vit_grad_ckpt'], vit_ckpt_layer=config['vit_ckpt_layer'],
prompt=config['prompt'], tokenizer=tokenizer, args=args)
# get function handles of loss and metrics
criterion = compute_loss
metrics = compute_scores
# build optimizer, learning rate scheduler
optimizer = build_optimizer_blip(args, model)
lr_scheduler = build_lr_scheduler(args, optimizer)
# build trainer and start to train
trainer = Trainer(model, criterion, metrics, optimizer, args, lr_scheduler, train_dataloader, val_dataloader, test_dataloader, tokenizer)
trainer.train()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config', default='./configs/Generation.yaml')
parser.add_argument('--checkpoint', default='')
parser.add_argument('--pretrained', default='')
parser.add_argument('--output_dir', default='output/generation')
parser.add_argument('--evaluate', action='store_true')
parser.add_argument('--text_encoder', default='bert-base-uncased')
parser.add_argument('--text_decoder', default='bert-base-uncased')
parser.add_argument('--device', default='cuda')
parser.add_argument('--seed', default=42, type=int)
parser.add_argument('--world_size', default=1, type=int, help='number of distributed processes')
parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training')
parser.add_argument('--distributed', default=True, type=bool)
parser.add_argument('--image_dir', type=str,
default='./dataset/iu_xray/images&./dataset/MIMIC-CXR/mimic_cxr/images',
help='the path to the directory containing the data.')
parser.add_argument('--ann_path', type=str,
default='./annotations/iu-annotation.json&./annotations/mimic_annotation.json',
help='the path to the directory containing the data.')
parser.add_argument('--knowledge_path', type=str,
default='./annotations/iu_train_kg_AO.json&./annotations/mimic_train_kg_AO.json',
help='the path to the directory containing the data.')
# Data loader settings
parser.add_argument('--dataset_name', type=str, default='iu_xray', choices=['iu_xray', 'mimic_cxr'],
help='the dataset to be used.')
parser.add_argument('--max_seq_length', type=int, default=90, help='the maximum sequence length of the reports.')
parser.add_argument('--threshold', type=int, default=3, help='the cut off frequency for the words.')
parser.add_argument('--num_workers', type=int, default=2, help='the number of workers for dataloader.')
parser.add_argument('--batch_size', type=int, default=2, help='the number of samples for a batch')
# Model settings (for visual extractor)
parser.add_argument('--visual_extractor', type=str, default='resnet101', help='the visual extractor to be used.')
parser.add_argument('--visual_extractor_pretrained', type=bool, default=True,
help='whether to load the pretrained visual extractor')
# Model settings (for Transformer)
parser.add_argument('--d_model', type=int, default=512, help='the dimension of Transformer.')
parser.add_argument('--d_ff', type=int, default=512, help='the dimension of FFN.')
parser.add_argument('--d_vf', type=int, default=2048, help='the dimension of the patch features.')
parser.add_argument('--num_heads', type=int, default=8, help='the number of heads in Transformer.')
parser.add_argument('--num_layers', type=int, default=3, help='the number of layers of Transformer.')
parser.add_argument('--dropout', type=float, default=0.1, help='the dropout rate of Transformer.')
parser.add_argument('--logit_layers', type=int, default=1, help='the number of the logit layer.')
parser.add_argument('--bos_idx', type=int, default=0, help='the index of <bos>.')
parser.add_argument('--eos_idx', type=int, default=0, help='the index of <eos>.')
parser.add_argument('--pad_idx', type=int, default=0, help='the index of <pad>.')
parser.add_argument('--use_bn', type=int, default=0, help='whether to use batch normalization.')
parser.add_argument('--drop_prob_lm', type=float, default=0.5, help='the dropout rate of the output layer.')
# Sample related
parser.add_argument('--sample_method', type=str, default='beam_search',
help='the sample methods to sample a report.')
parser.add_argument('--beam_size', type=int, default=3, help='the beam size when beam searching.')
parser.add_argument('--temperature', type=float, default=1.0, help='the temperature when sampling.')
parser.add_argument('--sample_n', type=int, default=1, help='the sample number per image.')
parser.add_argument('--group_size', type=int, default=1, help='the group size.')
parser.add_argument('--output_logsoftmax', type=int, default=1, help='whether to output the probabilities.')
parser.add_argument('--decoding_constraint', type=int, default=0, help='whether decoding constraint.')
parser.add_argument('--block_trigrams', type=int, default=1, help='whether to use block trigrams.')
# Trainer settings
parser.add_argument('--n_gpu', type=int, default=1, help='the number of gpus to be used.')
parser.add_argument('--epochs', type=int, default=30, help='the number of training epochs.')
parser.add_argument('--save_dir', type=str, default='results/fair', help='the patch to save the models.')
parser.add_argument('--record_dir', type=str, default='records/generation/',
help='the patch to save the results of experiments')
parser.add_argument('--save_period', type=int, default=1, help='the saving period.')
parser.add_argument('--monitor_mode', type=str, default='max', choices=['min', 'max'],
help='whether to max or min the metric.')
parser.add_argument('--monitor_metric', type=str, default='BLEU_4', help='the metric to be monitored.')
parser.add_argument('--early_stop', type=int, default=50, help='the patience of training.')
# Optimization
parser.add_argument('--optim', type=str, default='Adam', help='the type of the optimizer.')
parser.add_argument('--lr_ve', type=float, default=1e-5, help='the learning rate for the visual extractor.')
parser.add_argument('--lr_ed', type=float, default=1e-4, help='the learning rate for the remaining parameters.')
parser.add_argument('--weight_decay', type=float, default=5e-5, help='the weight decay.')
parser.add_argument('--amsgrad', type=bool, default=True, help='.')
# Learning Rate Scheduler
parser.add_argument('--lr_scheduler', type=str, default='StepLR', help='the type of the learning rate scheduler.')
parser.add_argument('--step_size', type=int, default=50, help='the step size of the learning rate scheduler.')
parser.add_argument('--gamma', type=float, default=0.1, help='the gamma of the learning rate scheduler.')
# Others
# parser.add_argument('--seed', type=int, default=9233, help='.')
parser.add_argument('--resume', type=str, help='whether to resume the training from existing checkpoints.')
parser.add_argument('--test_visual', type=bool, default=False, help='whether to test the visual encoder.')
parser.add_argument('--test_text', type=bool, default=False, help='whether to test the text encoder.')
parser.add_argument('--test_text_cross', type=bool, default=False, help='whether to test the text and cross encoder.')
parser.add_argument('--test_visual_text', type=bool, default=False, help='whether to test the text and visual encoder.')
parser.add_argument('--test_best', type=bool, default=False, help='whether to test the best model.')
parser.add_argument('--add_memory', type=bool, default=False, help='whether to test the best model.')
parser.add_argument('--tokenizer', type=str, default='blip', choices=['r2gen', 'blip'],
help='the dataset to be used.')
parser.add_argument('--bert', type=str, default='base', choices=['base', 'sci', 'cli'],
help='the dataset to be used.')
parser.add_argument('--concat', default=False, type=bool)
args = parser.parse_args()
config = yaml.load(open(args.config, 'r'), Loader=yaml.Loader)
main(args, config)