-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
146 lines (103 loc) · 4.55 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
import torch
import yaml
import torch.nn.functional as functional
import torch.optim as optim
import torch.utils.data as data
import torchvision.transforms as transforms
from models import ContrastModel
from dataset import CaptionDataset
def criterion(image_features, text_features, temperature):
predicts = image_features @ text_features.T
predicts /= temperature
labels = torch.arange(predicts.size(0))
labels = labels.cuda()
i2t_loss = functional.cross_entropy(predicts.permute(0, 1), labels)
t2i_loss = functional.cross_entropy(predicts.permute(1, 0), labels)
return i2t_loss + t2i_loss
def validation(image_features, text_features):
predicts = image_features @ text_features.T
labels = torch.arange(predicts.size(0))
labels = labels.cuda()
correct_i2t = (torch.argmax(predicts.permute(0, 1), dim=1) == labels).sum().item()
correct_t2i = (torch.argmax(predicts.permute(1, 0), dim=1) == labels).sum().item()
return correct_i2t, correct_t2i
with open('configs/train.yaml', 'r') as configs:
configs = yaml.safe_load(configs)
augment_transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.RandomAdjustSharpness(sharpness_factor=0.1),
transforms.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1),
transforms.ToTensor(),
transforms.RandomErasing(scale=(0.02, 0.1), ratio=(0.5, 2.0)),
])
normal_transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
])
if configs['use-augment']:
train_transform = augment_transform
else:
train_transform = normal_transform
valid_transform = normal_transform
train_dataset = CaptionDataset('datasets/train', train_transform)
valid_dataset = CaptionDataset('datasets/valid', valid_transform)
train_loader = data.DataLoader(train_dataset, configs['batch-size'], shuffle=True, num_workers=configs['num-workers'])
valid_loader = data.DataLoader(valid_dataset, configs['batch-size'], shuffle=True, num_workers=configs['num-workers'])
load_path = configs['load-path']
best_path = configs['best-path']
last_path = configs['last-path']
model = ContrastModel(pretrained=configs['load-pretrained'])
model = model.cuda()
if configs['load-checkpoint']:
model.load_state_dict(torch.load(load_path, map_location='cuda', weights_only=True))
scaler = torch.GradScaler()
optimizer = optim.AdamW(model.parameters(), lr=configs['learning-rate'])
best_accuracy = 0.0
temperature = configs['temperature']
print(f'\n---------- Training Start ----------\n')
for epoch in range(configs['epochs']):
model.train()
training_loss = 0.0
for index, (images, texts, masks) in enumerate(train_loader, start=1):
images = images.cuda()
texts = texts.cuda()
masks = masks.cuda()
optimizer.zero_grad()
if configs['use-amp']:
with torch.autocast('cuda'):
loss = criterion(model.encode_image(images), model.encode_text(texts, masks), temperature)
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
else:
loss = criterion(model.encode_image(images), model.encode_text(texts, masks), temperature)
loss.backward()
optimizer.step()
training_loss += loss.item()
print(f'\rBatch Loss: {loss:.5f} [{index}/{len(train_loader)}]', end='')
model.eval()
training_loss /= len(train_loader)
with torch.no_grad():
accuracy_i2t = 0
accuracy_t2i = 0
for images, texts, masks in valid_loader:
images = images.cuda()
texts = texts.cuda()
masks = masks.cuda()
if configs['use-amp']:
with torch.autocast('cuda'):
correct_i2t, correct_t2i = validation(model.encode_image(images), model.encode_text(texts, masks))
else:
correct_i2t, correct_t2i = validation(model.encode_image(images), model.encode_text(texts, masks))
accuracy_i2t += correct_i2t
accuracy_t2i += correct_t2i
accuracy_i2t /= len(valid_dataset)
accuracy_t2i /= len(valid_dataset)
mean_accuracy = 2 * (accuracy_i2t * accuracy_t2i) / (accuracy_i2t + accuracy_t2i)
if mean_accuracy > best_accuracy:
best_accuracy = mean_accuracy
torch.save(model.state_dict(), best_path)
torch.save(model.state_dict(), last_path)
print(f'\tEpoch: {epoch:<6} Loss: {training_loss:<10.5f} I2T: {accuracy_i2t:<6.2f} T2I: {accuracy_t2i:<6.2f}')
print('\n---------- Training Finish ----------\n')
print(f'Best Accuracy: {best_accuracy:.5f}')