SOLD/train_encoder_decoder.py at main · SAIS-LifeScience/SOLD · 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
import warnings
warnings.filterwarnings("ignore")

from datetime import datetime
from torch.utils.tensorboard import SummaryWriter
from utils import *
from dataset.graph_dataset import GraphDataset
from torch.utils.data import DataLoader
from models import *
import torch.nn as nn
from tqdm import tqdm

def train_encoder_decoder(config):
    print("Training Encoder-Decoder model...")
    data_config = config['data_config']
    train_config = config['train_config']
    model_config = config['model_config']
    llm_config = model_config['llm_config']
    llm_config['device'] = config['device']
    data_config['llm_config'] = llm_config

    encoder_config = model_config['encoder']
    decoder_config = model_config['decoder']

    seed = config['seed']
    seeding(seed)

    llm_name = llm_config['name']
    encoder_name = encoder_config['name']
    decoder_name = decoder_config['name']
    compress_dim = decoder_config['config']['encoder_hidden_dim']

    result_folder = os.path.join(config['train_config']['output_dir'], f'{llm_name}_encoder_{encoder_name}_decoder_{decoder_name}_compress_dim_{compress_dim}')
    if not os.path.exists(result_folder):
        os.makedirs(result_folder)
    save_config(os.path.join(result_folder, 'config.yaml'), config)

    # Initialize TensorBoard
    current_time = datetime.now().strftime("%Y%m%d-%H%M%S")
    log_dir = os.path.join(result_folder, "runs", current_time)
    writer = SummaryWriter(log_dir=log_dir)
    print(f"TensorBoard logs will be saved to: {log_dir}")

    train_dataset = GraphDataset(data_path=data_config['train_data_path'],
                                 data_config=data_config)

    valid_dataset = GraphDataset(data_path=data_config['valid_data_path'],
                                 data_config=data_config)

    test_dataset = GraphDataset(data_path=data_config['test_data_path'],
                                data_config=data_config)

    train_dataloader = DataLoader(train_dataset,
                                  batch_size=train_config['batch_size'],
                                  shuffle=True,
                                  collate_fn=train_dataset.collate_fn)
    valid_dataloader = DataLoader(valid_dataset,
                                  batch_size=train_config['batch_size'],
                                  shuffle=False,
                                  collate_fn=valid_dataset.collate_fn)
    test_dataloader = DataLoader(test_dataset,
                                  batch_size=train_config['batch_size'],
                                  shuffle=False,
                                  collate_fn=test_dataset.collate_fn)

    if model_config['decoder']['name'] == 'mlp':
        decoder = MLPDecoder(**model_config['decoder']['config'])
        if model_config.get("encoder"):
            if model_config["encoder"]["name"] == "mlp":
                encoder = MLPEncoder(**model_config["encoder"]["config"])
            else:
                raise NotImplementedError(model_config["encoder"]["name"])
            encoder_decoder = EncoderMlpDecoder(decoder=decoder, encoder=encoder).to(config['device'])
        else:
            encoder_decoder = EncoderMlpDecoder(decoder=decoder).to(config['device'])
    else:
        raise NotImplementedError(model_config['decoder']['name'])

    optimizer = torch.optim.AdamW(decoder.parameters(), lr=train_config['lr'])
    loss_fn = nn.CrossEntropyLoss()

    print("---------- Start Training Encoder-Decoder ----------")
    gradient_step = 0
    accum_step = 0
    best_valid_loss = np.inf
    patient = 0
    for epoch in range(train_config['max_train_epochs']):
        print(f"---------- Epoch {epoch} ----------")

        encoder_decoder.eval()
        val_loss, val_recovery_seq, val_recovery_aa = validate_encoder_decoder(encoder_decoder, config['device'], valid_dataloader, loss_fn, 'validate', epoch)
        print(f"epoch: {epoch}, val seq recovery: {val_recovery_seq}, val recovery aa: {val_recovery_aa}")
        test_loss, test_recovery_seq, test_recovery_aa = validate_encoder_decoder(encoder_decoder, config['device'], test_dataloader, loss_fn, 'test', epoch)
        print(f"epoch: {epoch}, test seq recovery: {test_recovery_seq}, test recovery aa: {test_recovery_aa}")
        writer.add_scalar(f"Validation/loss", val_loss, epoch)
        writer.add_scalar(f"Validation/recovery_seq", val_recovery_seq, epoch)
        writer.add_scalar(f"Validation/recovery_aa", val_recovery_aa, epoch)
        writer.add_scalar(f"Test/loss", test_loss, epoch)
        writer.add_scalar(f"Test/recovery_seq", test_recovery_seq, epoch)
        writer.add_scalar(f"Test/recovery_aa", test_recovery_aa, epoch)
        if val_loss < best_valid_loss:
            best_valid_loss = val_loss
            patient = 0
            state_dict = {
                "encoder": encoder_decoder.encoder.state_dict(),
                "decoder": encoder_decoder.decoder.state_dict(),
                "encoder_decoder": encoder_decoder.state_dict(),
            }
            print("now saving model on epoch {} ......".format(epoch))
            torch.save(state_dict, os.path.join(result_folder, 'model_separate_epoch_{}.pt'.format(epoch)))
        else:
            patient += 1

        if patient > train_config['patience']:
            print("Early Stopping .......")
            break

        encoder_decoder.train()
        total_loss = 0
        accum_loss = 0
        epoch_iterator = tqdm(train_dataloader)
        for batch in epoch_iterator:
            logits = encoder_decoder(batch)
            gt = torch.argmax(batch['seq'], dim=-1).to(config['device'])
            loss = loss_fn(logits, gt)
            total_loss += loss.item()
            accum_loss += loss.item()
            accum_step += 1
            loss.backward()
            epoch_iterator.set_postfix(loss=loss.item())

            if accum_step % train_config['gradient_accumulate_every'] == 0:
                writer.add_scalar("Train-Step/loss", accum_loss / train_config['gradient_accumulate_every'], gradient_step)
                gradient_step += 1
                accum_loss = 0
                optimizer.step()
                optimizer.zero_grad()

        train_epoch_loss = total_loss / len(train_dataloader)
        writer.add_scalar(f'Train-Epoch/loss', train_epoch_loss, epoch)

    writer.close()
    print("training complete ......")