COSFormer/train.py at main · caodoanh2001/COSFormer · 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
import argparse
from pathlib import Path
import numpy as np
import glob

from datasets import DataInterface
from models import ModelInterface
from utils.utils import *

# pytorch_lightning
import pytorch_lightning as pl
from pytorch_lightning import Trainer
import json
import os

def load_logits(past_logit_path, buffer):
    past_logit_dict = torch.load(past_logit_path)
    new_dict = dict()
    for slide_id in past_logit_dict.keys():
        if slide_id in buffer.all_slide_ids():
            new_dict[slide_id] = past_logit_dict[slide_id]

    return new_dict

def count_parameters(model):
    return sum(p.numel() for p in model.parameters())

#--->Setting parameters
def make_parse():
    parser = argparse.ArgumentParser()
    parser.add_argument('--stage', default='train', type=str)
    parser.add_argument('--config', default='Camelyon/TransMIL.yaml',type=str)
    parser.add_argument('--gpus', default = [2])
    parser.add_argument('--fold', default = 0)
    parser.add_argument('--model_path', default = None, type=str)
    args = parser.parse_args()
    return args

#---->main
def main(cfg):

    #---->Initialize seed
    pl.seed_everything(cfg.General.seed)

    #---->load loggers
    cfg.load_loggers = load_loggers(cfg)

    #---->load callbacks
    cfg.callbacks = load_callbacks(cfg)

    #---->Define Data
    DataInterface_dict = {'train_batch_size': cfg.Data.train_dataloader.batch_size,
                'train_num_workers': cfg.Data.train_dataloader.num_workers,
                'test_batch_size': cfg.Data.test_dataloader.batch_size,
                'test_num_workers': cfg.Data.test_dataloader.num_workers,
                'dataset_name': cfg.Data.dataset_name,
                'dataset_cfg': cfg.Data,}
    dm = DataInterface(**DataInterface_dict)
    dm.setup(); vocab_size = dm.vocab_size
    print("Vocab size:", vocab_size)

    #---->Define Model
    ModelInterface_dict = {'model': cfg.Model,
                            'loss': cfg.Loss,
                            'optimizer': cfg.Optimizer,
                            'data': cfg.Data,
                            'log': cfg.log_path,
                            'vocab_size': vocab_size,
                            'vocab_path': cfg.Data.vocab_path,
                            'max_seq_len': cfg.Data.max_seq_len,
                            'vocab': json.load(open(cfg.Data.vocab_path, 'r')),
                            'bos_tag': cfg.Data.bos_tag,
                            'eos_tag': cfg.Data.eos_tag,
                            'padding_idx': cfg.Data.padding_idx,
                            'buffer': dm.buffer,
                            }

    model = ModelInterface(**ModelInterface_dict)
    #---->Instantiate Trainer
    trainer = Trainer(
        num_sanity_val_steps=0,
        logger=cfg.load_loggers,
        callbacks=cfg.callbacks,
        max_epochs= cfg.General.epochs,
        gpus=cfg.General.gpus,
        amp_level=cfg.General.amp_level,
        precision=cfg.General.precision,
        accumulate_grad_batches=cfg.General.grad_acc,
        deterministic=True,
        check_val_every_n_epoch=1,
    )

    tau = cfg.Hyperparameter.tau
    beta = cfg.Hyperparameter.beta

    path_past_logit = "./buffer/current_buffer_" + str(cfg.Data.fold) + ".pth"
    #---->train or test
    if cfg.General.server == 'train':
        model.buffer.construct_slides_per_task()

        if os.path.exists(path_past_logit):
            past_logit_dict = load_logits(path_past_logit, model.buffer)
            model.buffer.past_logits = past_logit_dict

        model.model.update_fc_layers(tau=tau, beta=beta)
        if args.model_path is not None:
            ckpt = torch.load(args.model_path)['state_dict']
            ckpt = {'.'.join(k.split('.')[1:]):ckpt[k] for k in ckpt.keys() if k != 'model.word_emb.weight'}
            err = model.model.load_state_dict(ckpt, strict=False)
            print(err)
            print("Load weights from previous task successfully")
            model.model.update_learnable_weights_for_new_tasks()
        trainer.fit(model = model, datamodule = dm)

        print("Saving current buffer...")
        torch.save(model.buffer.past_logits, "./buffer/current_buffer_" + str(cfg.Data.fold) + ".pth")

    else:
        print("Number of parameters:", count_parameters(model.model))

        model.model.update_fc_layers(tau=tau, beta=beta)

        if not cfg.Data.first_task:
            model.model.update_learnable_weights_for_new_tasks()

        ckpt = torch.load(args.model_path)['state_dict']
        ckpt = {'.'.join(k.split('.')[1:]):ckpt[k] for k in ckpt.keys() if k != 'model.word_emb.weight'}
        err = model.model.load_state_dict(ckpt, strict=False)
        print(err)

        trainer.test(model=model, datamodule=dm)

if __name__ == '__main__':

    args = make_parse()
    cfg = read_yaml(args.config)

    #---->update
    cfg.config = args.config
    cfg.General.gpus = args.gpus
    cfg.General.server = args.stage
    cfg.Data.fold = args.fold

    #---->main
    main(cfg)