ViT5P/train_continue.py at main · ADAning/ViT5P · GitHub

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
import argparse

import pytorch_lightning as pl
from pytorch_lightning.callbacks import ModelCheckpoint
from torch.utils.data import DataLoader
from transformers import ViTModel, T5ForConditionalGeneration

from data.dataset import ImageCaptionDataset
from trainer import TaskModel
from model.vl_model import ViT5P
from tools.utils import T5PegasusTokenizer

if __name__ == '__main__':

    parser = argparse.ArgumentParser()
    parser.add_argument('--min_length', type=int, default=0, help='captions min_length')
    parser.add_argument('--max_length', type=int, default=80, help='captions max_length')
    parser.add_argument('--total_batch_size', type=int, default=48, help='captions max_length')
    parser.add_argument('--epochs', type=int, default=80, help='captions max_length')
    parser.add_argument('--num_beams', default=4, type=int, help='num_beams')
    parser.add_argument('--out_file', default="predictions.jsonl", type=str,
                        help='outfile path and name')
    # pretrained model
    parser.add_argument('--ckpt_path', type=str,
                        help='outfile path and name')
    parser.add_argument('--visual_path', type=str,
                        help='vit model path')
    parser.add_argument('--language_path', type=str,
                        help='t5 pegasus model path')
    # datasets
    parser.add_argument('--image_path', type=str,
                        help='outfile path and name')
    parser.add_argument('--caption_path', type=str,
                        help='caption dataset path')
    args = parser.parse_args()
    out_file = args.out_file
    min_length = args.min_length
    max_length = args.max_length
    num_beams = args.num_beams
    total_batch_size = args.total_batch_size
    epochs = args.epochs
    # pretrained model
    ckpt_path = args.ckpt_path
    visual_path = args.visual_path
    language_path = args.language_path
    # datasets
    image_path = args.image_path
    caption_path = args.caption_path

    visual_model = ViTModel.from_pretrained(visual_path)
    language_model = T5ForConditionalGeneration.from_pretrained(language_path)
    tokenizer = T5PegasusTokenizer.from_pretrained(language_path)
    model = ViT5P(
        visual_model=visual_model,
        language_model=language_model,
        pseudo_label_num=0
    )
    train_dataset = ImageCaptionDataset(
        image_path.format("train"),
        caption_path.format("train"),
        visual_path,
        language_path,
        max_length,
    )

    valid_dataset = ImageCaptionDataset(
        image_path.format("valid"),
        caption_path.format("valid"),
        visual_path,
        language_path,
        max_length,
    )

    test_dataset = ImageCaptionDataset(
        image_path.format("test"),
        caption_path.format("test"),
        visual_path,
        language_path,
        max_length,
    )

    num_devices = 1
    batch_size = total_batch_size // num_devices
    num_wokers = 8
    train_loader = DataLoader(
        dataset=train_dataset,
        batch_size=batch_size,
        shuffle=True,
        num_workers=num_wokers,
        collate_fn=train_dataset.train_collate_finetune,
        pin_memory=True,
        persistent_workers=True,
    )

    valid_loader = DataLoader(
        dataset=valid_dataset,
        batch_size=batch_size,
        shuffle=False,
        num_workers=num_wokers,
        collate_fn=valid_dataset.test_collate_finetune,
        pin_memory=True,
        persistent_workers=True,
    )

    test_loader = DataLoader(
        dataset=test_dataset,
        batch_size=batch_size,
        shuffle=False,
        num_workers=num_wokers,
        collate_fn=test_dataset.test_collate_finetune,
        pin_memory=True,
        persistent_workers=True,
    )

    checkpoint = ModelCheckpoint(
        save_weights_only=True,
        save_on_train_epoch_end=False,
        monitor="valid CIDEr-D",
        mode="max",
        verbose=True,
        save_top_k=6,
    )

    visual_lr = 1e-5
    language_lr = 3e-4
    gradient_clip = False

    task_model = TaskModel.load_from_checkpoint(
        ckpt_path,
        model=model,
        tokenizer=tokenizer,
        visual_lr=visual_lr,
        language_lr=language_lr,
        max_length=max_length,
        training_mode="xe",
        num_beams=num_beams,
    )

    if num_devices > 1:
        print("only support 1 gpu! please set num_devices=1!")

    else:
        trainer = pl.Trainer(
            max_epochs=epochs,
            callbacks=[checkpoint],
            num_sanity_val_steps=0,
            accelerator="gpu",
            devices=1,
            precision=32,
            gradient_clip_val=1.0 if gradient_clip else None,
        )
        trainer.fit(task_model, train_loader, valid_loader)
        trainer.test(task_model, test_loader, ckpt_path="best")