-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_rein_ours_stage2.py
156 lines (130 loc) · 6.11 KB
/
train_rein_ours_stage2.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
import os
import torch
import torch.nn as nn
import argparse
import timm
import numpy as np
import utils
import random
import rein
import dino_variant
def symmetric_cross_entropy(x, x_ema):# -> torch.Tensor:
return -0.5*(x_ema.softmax(1) * x.log_softmax(1)).sum(1)-0.5*(x.softmax(1) * x_ema.log_softmax(1)).sum(1)
def train():
parser = argparse.ArgumentParser()
parser.add_argument('--data', '-d', type=str)
parser.add_argument('--gpu', '-g', default = '0', type=str)
parser.add_argument('--netsize', default='s', type=str)
parser.add_argument('--save_path', '-s', type=str)
parser.add_argument('--noise_rate', '-n', type=float, default=0.2)
parser.add_argument('--teacher', '-t', type=str)
args = parser.parse_args()
config = utils.read_conf('conf/'+args.data+'.json')
device = 'cuda:'+args.gpu
save_path = os.path.join(config['save_path'], args.save_path)
data_path = config['id_dataset']
batch_size = int(config['batch_size'])
max_epoch = int(config['epoch'])
noise_rate = args.noise_rate
teacher_path = os.path.join(config['save_path'], args.teacher, 'last.pth.tar')
if not os.path.exists(save_path):
os.mkdir(save_path)
lr_decay = [int(0.5*max_epoch), int(0.75*max_epoch), int(0.9*max_epoch)]
if args.data == 'ham10000':
train_loader, valid_loader = utils.get_noise_dataset(data_path, noise_rate=noise_rate, batch_size = batch_size)
elif args.data == 'aptos':
train_loader, valid_loader = utils.get_aptos_noise_dataset(data_path, noise_rate=noise_rate, batch_size = batch_size)
elif args.data == 'idrid':
train_loader, valid_loader = utils.get_idrid_noise_dataset(data_path, noise_rate=noise_rate, batch_size = batch_size)
elif args.data == 'chaoyang':
train_loader, valid_loader = utils.get_chaoyang_dataset(data_path, batch_size = batch_size)
elif 'mnist' in args.data:
train_loader, valid_loader = utils.get_mnist_noise_dataset(args.data, noise_rate=noise_rate, batch_size = batch_size)
if args.netsize == 's':
model_load = dino_variant._small_dino
variant = dino_variant._small_variant
elif args.netsize == 'b':
model_load = dino_variant._base_dino
variant = dino_variant._base_variant
elif args.netsize == 'l':
model_load = dino_variant._large_dino
variant = dino_variant._large_variant
# model = timm.create_model(network, pretrained=True, num_classes=2)
model = torch.hub.load('facebookresearch/dinov2', model_load)
dino_state_dict = model.state_dict()
model = rein.ReinsDinoVisionTransformer(
**variant
)
model.load_state_dict(dino_state_dict, strict=False)
model.requires_grad_(False)
model.linear_rein = nn.Linear(variant['embed_dim'], config['num_classes'])
model.to(device)
# print(model.state_dict()['blocks.11.mlp.fc2.weight'])
criterion = torch.nn.CrossEntropyLoss(reduction='none')
model.eval()
teacher = rein.ReinsDinoVisionTransformer(
**variant
)
teacher.linear_rein = nn.Linear(variant['embed_dim'], config['num_classes'])
teacher.load_state_dict(torch.load(teacher_path, map_location='cpu')['state_dict'], strict=False)
teacher.to(device)
teacher.eval()
valid_accuracy = utils.validation_accuracy_ours(teacher, valid_loader, device)
print('Teacher accuracy on test set: ', valid_accuracy)
# optimizer = torch.optim.SGD(model.parameters(), lr = 0.01, momentum=0.9, weight_decay = 1e-05)
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-3, weight_decay = 1e-5)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, lr_decay)
saver = timm.utils.CheckpointSaver(model, optimizer, checkpoint_dir= save_path, max_history = 1)
print(train_loader.dataset[0][0].shape)
avg_accuracy = 0.0
for epoch in range(max_epoch):
## training
model.train()
total_loss = 0
total = 0
correct = 0
correct_linear = 0
for batch_idx, (inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(device), targets.to(device)
optimizer.zero_grad()
features_rein = model.forward_features(inputs)
features_rein = features_rein[:, 0, :]
outputs = model.linear_rein(features_rein)
# print(outputs.shape, outputs_.shape)
with torch.no_grad():
features_rein = teacher.forward_features(inputs)
features_rein = features_rein[:, 0, :]
outputs_ = teacher.linear_rein(features_rein)
pred = outputs_.max(1).indices
linear_accurate = (pred==targets)
# print(pred.shape, targets.shape)
# print(model.reins.state_dict())
loss_rein = linear_accurate*criterion(outputs, targets)
loss = loss_rein.mean()
loss.backward()
optimizer.step()
total_loss += loss
total += targets.size(0)
_, predicted = outputs[:len(targets)].max(1)
correct += predicted.eq(targets).sum().item()
print('\r', batch_idx, len(train_loader), 'Loss: %.3f | Acc: %.3f%% | (%d/%d)'
% (total_loss/(batch_idx+1), 100.*correct/total, correct, total), end = '')
train_accuracy = correct/total
train_avg_loss = total_loss/len(train_loader)
print()
## validation
model.eval()
total_loss = 0
total = 0
correct = 0
valid_accuracy = utils.validation_accuracy_ours(model, valid_loader, device)
scheduler.step()
if epoch >= max_epoch-10:
avg_accuracy += valid_accuracy
saver.save_checkpoint(epoch, metric = valid_accuracy)
print('EPOCH {:4}, TRAIN [loss - {:.4f}, acc - {:.4f}], VALID [acc - {:.4f}]\n'.format(epoch, train_avg_loss, train_accuracy, valid_accuracy))
print(scheduler.get_last_lr())
with open(os.path.join(save_path, 'avgacc.txt'), 'w') as f:
f.write(str(avg_accuracy/10))
if __name__ =='__main__':
train()