forked from jinhojsk515/SPMM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
d_classification_multilabel.py
191 lines (160 loc) · 8.07 KB
/
d_classification_multilabel.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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
import argparse
from tqdm import tqdm
import torch
import numpy as np
import time
import torch.backends.cudnn as cudnn
from transformers import BertTokenizer, WordpieceTokenizer
import datetime
from dataset import SMILESDataset_Clintox, SMILESDataset_SIDER
from torch.utils.data import DataLoader
import torch.optim as optim
from scheduler import create_scheduler
import random
from sklearn.metrics import roc_auc_score
import torch.nn as nn
from xbert import BertConfig, BertForMaskedLM
class AttrDict(dict):
def __init__(self, *args, **kwargs):
super(AttrDict, self).__init__(*args, **kwargs)
self.__dict__ = self
class SPMM_classifier(nn.Module):
def __init__(self, tokenizer=None, config=None, n_output=2):
super().__init__()
self.tokenizer = tokenizer
bert_config = BertConfig.from_json_file(config['bert_config_text'])
self.text_encoder = BertForMaskedLM(config=bert_config)
for i in range(bert_config.fusion_layer, bert_config.num_hidden_layers): self.text_encoder.bert.encoder.layer[i] = nn.Identity()
self.text_encoder.cls = nn.Identity()
text_width = self.text_encoder.config.hidden_size
self.reg_head = nn.Sequential(
nn.Linear(text_width * 1, text_width * 1),
nn.GELU(),
nn.Linear(text_width * 1, n_output)
)
def forward(self, text_input_ids, text_attention_mask, value, eval=False):
vl_embeddings = self.text_encoder.bert(text_input_ids, attention_mask=text_attention_mask, return_dict=True, mode='text').last_hidden_state[:, 0, :]
pred = self.reg_head(vl_embeddings)
if eval: return pred
loss = nn.BCELoss()(torch.sigmoid(pred), value)
return loss
def train(model, data_loader, optimizer, tokenizer, epoch, warmup_steps, device, scheduler):
# train
model.train()
header = 'Train Epoch: [{}]'.format(epoch)
print_freq = 50
step_size = 50
warmup_iterations = warmup_steps * step_size
tqdm_data_loader = tqdm(data_loader, miniters=print_freq, desc=header)
for i, (text, label) in enumerate(tqdm_data_loader):
optimizer.zero_grad()
label = label.to(device, non_blocking=True)
text_input = tokenizer(text, padding='longest', truncation=True, max_length=100, return_tensors="pt").to(device)
loss = model(text_input.input_ids[:, 1:], text_input.attention_mask[:, 1:], label)
loss.backward()
optimizer.step()
tqdm_data_loader.set_description(f'loss={loss.item():.4f}, lr={optimizer.param_groups[0]["lr"]:.6f}')
if epoch == 0 and i % step_size == 0 and i <= warmup_iterations:
scheduler.step(i // step_size)
@torch.no_grad()
def evaluate(model, data_loader, tokenizer, device):
# test
model.eval()
scores = []
answers = []
for text, label in data_loader:
label = label.to(device, non_blocking=True)
text_input = tokenizer(text, padding='longest', return_tensors="pt").to(device)
prediction = model(text_input.input_ids[:, 1:], text_input.attention_mask[:, 1:], label, eval=True)
score = torch.sigmoid_(prediction)
scores.append(score.cpu())
answers.append(label.cpu())
scores = torch.cat(scores, dim=0)
answers = torch.cat(answers, dim=0)
return roc_auc_score(answers, scores)
def main(args, config):
device = torch.device(args.device)
print('DATASET:', args.name)
# === Dataset === #
name = args.name
if name == 'clintox':
dataset_train = SMILESDataset_Clintox('data/4_MoleculeNet/Clintox_train.csv')
dataset_val = SMILESDataset_Clintox('data/4_MoleculeNet/Clintox_valid.csv')
dataset_test = SMILESDataset_Clintox('data/4_MoleculeNet/Clintox_test.csv')
elif name == 'sider':
dataset_train = SMILESDataset_SIDER('data/4_MoleculeNet/SIDER_train.csv')
dataset_val = SMILESDataset_SIDER('data/4_MoleculeNet/SIDER_valid.csv')
dataset_test = SMILESDataset_SIDER('data/4_MoleculeNet/SIDER_test.csv')
else:
print('Dataset not found', name)
print(len(dataset_train), len(dataset_val), len(dataset_test))
train_loader = DataLoader(dataset_train, batch_size=config['batch_size_train'], num_workers=8, pin_memory=True, drop_last=False)
val_loader = DataLoader(dataset_val, batch_size=config['batch_size_test'], num_workers=8, pin_memory=True, drop_last=False)
test_loader = DataLoader(dataset_test, batch_size=config['batch_size_test'], num_workers=8, pin_memory=True, drop_last=False)
tokenizer = BertTokenizer(vocab_file=args.vocab_filename, do_lower_case=False, do_basic_tokenize=False)
tokenizer.wordpiece_tokenizer = WordpieceTokenizer(vocab=tokenizer.vocab, unk_token=tokenizer.unk_token, max_input_chars_per_word=250)
# fix the seed for reproducibility
seed = args.seed if args.seed else random.randint(0, 100)
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
cudnn.benchmark = True
# === Model === #
print("Creating model")
model = SPMM_classifier(config=config, tokenizer=tokenizer, n_output=dataset_train.n_output)
print('#parameters:', sum(p.numel() for p in model.parameters() if p.requires_grad))
if args.checkpoint:
print('LOADING PRETRAINED MODEL..')
checkpoint = torch.load(args.checkpoint, map_location='cpu')
state_dict = checkpoint['state_dict']
msg = model.load_state_dict(state_dict, strict=False)
print('load checkpoint from %s' % args.checkpoint)
# print(msg)
model = model.to(device)
arg_opt = config['optimizer']
optimizer = optim.AdamW(model.parameters(), lr=arg_opt['lr'], weight_decay=arg_opt['weight_decay'])
arg_sche = AttrDict(config['schedular'])
lr_scheduler, _ = create_scheduler(arg_sche, optimizer)
max_epoch = config['schedular']['epochs']
warmup_steps = config['schedular']['warmup_epochs']
best_valid = 0.
best_test = 0.
start_time = time.time()
for epoch in range(0, max_epoch):
print('TRAIN', epoch)
train(model, train_loader, optimizer, tokenizer, epoch, warmup_steps, device, lr_scheduler)
val_stats = evaluate(model, val_loader, tokenizer, device)
print('VALID AUROC: %.4f' % val_stats)
test_stats = evaluate(model, test_loader, tokenizer, device)
print('TEST AUROC: %.4f' % test_stats)
if val_stats > best_valid:
best_valid = val_stats
best_test = test_stats
lr_scheduler.step(epoch + warmup_steps + 1)
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
print('DATASET:', args.name, '\tTest set AUROC of the checkpoint with best validation AUROC:', best_test)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint', default='./Pretrain/checkpoint_SPMM_20m.ckpt')
parser.add_argument('--vocab_filename', default='./vocab_bpe_300.txt')
parser.add_argument('--device', default='cuda')
parser.add_argument('--seed', default=37, type=int)
parser.add_argument('--name', default='clintox', type=str)
parser.add_argument('--lr', default=2e-5, type=float)
parser.add_argument('--min_lr', default=5e-6, type=float)
parser.add_argument('--epoch', default=10, type=int)
parser.add_argument('--batch_size', default=8, type=int)
args = parser.parse_args()
cls_config = {
'batch_size_train': args.batch_size,
'batch_size_test': 16,
'embed_dim': 256,
'bert_config_text': './config_bert.json',
'bert_config_property': './config_bert_property.json',
'schedular': {'sched': 'cosine', 'lr': args.lr, 'epochs': args.epoch, 'min_lr': args.min_lr,
'decay_rate': 1, 'warmup_lr': 0.5e-5, 'warmup_epochs': 1, 'cooldown_epochs': 0},
'optimizer': {'opt': 'adamW', 'lr': args.lr, 'weight_decay': 0.02}
}
main(args, cls_config)