-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain.py
189 lines (153 loc) · 6.37 KB
/
train.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
import os
import torch
import time
from tqdm import tqdm
from opts import parse_opts
from core.dataset import MMDataLoader
from core.scheduler import get_scheduler
from core.utils import AverageMeter, setup_seed, ConfigLogging, save_print_results, calculate_u_test
from models.OverallModal import build_model
from core.metric import MetricsTop
opt = parse_opts()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def main(parse_args):
opt = parse_args
log_path = os.path.join(opt.log_path, opt.datasetName.upper())
if not os.path.exists(log_path):
os.makedirs(log_path)
log_file = os.path.join(log_path, time.strftime('%Y-%m-%d-%H:%M:%S' + '.log'))
logger = ConfigLogging(log_file)
logger.info(opt) # 保存当前模型参数
setup_seed(opt.seed)
model = build_model(opt).to(device)
model.preprocess_model(pretrain_path={
'T': "./pretrainedModel/KnowledgeInjectPretraining/SIMS/SIMS_T_MAE-0.278_Corr-0.765.pth",
'V': "./pretrainedModel/KnowledgeInjectPretraining/SIMS/SIMS_V_MAE-0.522_Corr-0.520.pth",
'A': "./pretrainedModel/KnowledgeInjectPretraining/SIMS/SIMS_A_MAE-0.516_Corr-0.261.pth"
}) # 加载预训练权重并冻结参数
dataLoader = MMDataLoader(opt)
optimizer = torch.optim.AdamW(
model.parameters(),
lr=opt.lr,
weight_decay=opt.weight_decay
)
loss_fn = torch.nn.MSELoss()
metrics = MetricsTop().getMetics(opt.datasetName)
scheduler_warmup = get_scheduler(optimizer, opt.n_epochs)
for epoch in range(1, opt.n_epochs+1):
train_results = train(model, dataLoader['train'], optimizer, loss_fn, epoch, metrics)
valid_results = evaluate(model, dataLoader['valid'], optimizer, loss_fn, epoch, metrics)
test_results = test(model, dataLoader['test'], optimizer, loss_fn, epoch, metrics)
save_print_results(opt, logger, train_results, valid_results, test_results)
scheduler_warmup.step()
def train(model, train_loader, optimizer, loss_fn, epoch, metrics):
train_pbar = tqdm(train_loader)
losses = AverageMeter()
y_pred, y_true = [], []
model.train()
for data in train_pbar:
inputs = {
'V': data['vision'].to(device),
'A': data['audio'].to(device),
'T': data['text'].to(device),
'mask': {
'V': data['vision_padding_mask'][:, 1:data['vision'].shape[1]+1].to(device),
'A': data['audio_padding_mask'][:, 1:data['audio'].shape[1]+1].to(device),
'T': []
}
}
label = data['labels']['M'].to(device)
label = label.view(-1, 1)
copy_label = label.clone().detach()
batchsize = inputs['V'].shape[0]
output, nce_loss = model(inputs, copy_label)
loss_re = loss_fn(output, label)
loss = loss_re + 0.1 * nce_loss
losses.update(loss.item(), batchsize)
loss.backward()
optimizer.step()
optimizer.zero_grad()
y_pred.append(output.cpu())
y_true.append(label.cpu())
train_pbar.set_description('train')
train_pbar.set_postfix({
'epoch': '{}'.format(epoch),
'loss': '{:.5f}'.format(losses.value_avg),
'lr:': '{:.2e}'.format(optimizer.state_dict()['param_groups'][0]['lr'])
})
pred, true = torch.cat(y_pred), torch.cat(y_true)
train_results = metrics(pred, true)
return train_results
def evaluate(model, eval_loader, optimizer, loss_fn, epoch, metrics):
test_pbar = tqdm(eval_loader)
losses = AverageMeter()
y_pred, y_true = [], []
model.eval()
with torch.no_grad():
for data in test_pbar:
inputs = {
'V': data['vision'].to(device),
'A': data['audio'].to(device),
'T': data['text'].to(device),
'mask': {
'V': data['vision_padding_mask'][:, 1:data['vision'].shape[1]+1].to(device),
'A': data['audio_padding_mask'][:, 1:data['audio'].shape[1]+1].to(device),
'T': []
}
}
label = data['labels']['M'].to(device)
label = label.view(-1, 1)
batchsize = inputs['V'].shape[0]
output, _ = model(inputs, None)
y_pred.append(output.cpu())
y_true.append(label.cpu())
loss = loss_fn(output, label)
losses.update(loss.item(), batchsize)
test_pbar.set_description('eval')
test_pbar.set_postfix({
'epoch': '{}'.format(epoch),
'loss': '{:.5f}'.format(losses.value_avg),
'lr:': '{:.2e}'.format(optimizer.state_dict()['param_groups'][0]['lr'])
})
pred, true = torch.cat(y_pred), torch.cat(y_true)
valid_results = metrics(pred, true)
return valid_results
def test(model, test_loader, optimizer, loss_fn, epoch, metrics):
test_pbar = tqdm(test_loader)
losses = AverageMeter()
y_pred, y_true = [], []
model.eval()
with torch.no_grad():
for data in test_pbar:
inputs = {
'V': data['vision'].to(device),
'A': data['audio'].to(device),
'T': data['text'].to(device),
'mask': {
'V': data['vision_padding_mask'][:, 1:data['vision'].shape[1]+1].to(device),
'A': data['audio_padding_mask'][:, 1:data['audio'].shape[1]+1].to(device),
'T': []
}
}
ids = data['id']
label = data['labels']['M'].to(device)
label = label.view(-1, 1)
batchsize = inputs['V'].shape[0]
output, _ = model(inputs, None)
y_pred.append(output.cpu())
y_true.append(label.cpu())
loss = loss_fn(output, label)
losses.update(loss.item(), batchsize)
test_pbar.set_description('test')
test_pbar.set_postfix({
'epoch': '{}'.format(epoch),
'loss': '{:.5f}'.format(losses.value_avg),
'lr:': '{:.2e}'.format(optimizer.state_dict()['param_groups'][0]['lr'])
})
pred, true = torch.cat(y_pred), torch.cat(y_true)
if epoch == 11:
calculate_u_test(pred, true)
test_results = metrics(pred, true)
return test_results
if __name__ == '__main__':
main(opt)