-
Notifications
You must be signed in to change notification settings - Fork 7
/
linear_eval.py
175 lines (134 loc) · 5.91 KB
/
linear_eval.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
import torch
from data.imagenet import *
from data.augmentation import *
from network.head import *
from network.resnet import *
from torch.nn.parallel import DistributedDataParallel
import torch.nn.functional as F
from util.meter import *
import time
from util.torch_dist_sum import *
from util.dist_init import dist_init
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--checkpoint', type=str, default='')
args = parser.parse_args()
print(args)
def accuracy(output, target, topk=(1,)):
with torch.no_grad():
maxk = max(topk)
batch_size = target.size(0)
_, pred = output.topk(maxk, 1, True, True)
pred = pred.t()
correct = pred.eq(target.view(1, -1).expand_as(pred))
res = []
for k in topk:
correct_k = correct[:k].view(-1).float().sum(0, keepdim=True)
res.append(correct_k.mul_(100.0 / batch_size))
return res
def main():
rank, local_rank, world_size = dist_init()
epochs = 80
batch_size = 4096 // world_size
num_workers = 8
lr = 0.2 * 4096 / 256
pre_train = resnet50()
pre_train = nn.SyncBatchNorm.convert_sync_batchnorm(pre_train)
state_dict = torch.load('checkpoints/' + args.checkpoint, map_location='cpu')['model']
prefix = 'module.net.'
for k in list(state_dict.keys()):
if not k.startswith(prefix):
del state_dict[k]
if k.startswith(prefix):
state_dict[k[len(prefix):]] = state_dict[k]
del state_dict[k]
pre_train.load_state_dict(state_dict)
model = LinearHead(pre_train)
model = DistributedDataParallel(model.to(local_rank), device_ids=[local_rank], output_device=local_rank, find_unused_parameters=True)
optimizer = torch.optim.SGD(model.module.fc.parameters(), lr=lr, momentum=0.9, weight_decay=0, nesterov=True)
torch.backends.cudnn.benchmark = True
train_dataset = Imagenet(max_class=1000)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=batch_size, shuffle=(train_sampler is None),
num_workers=num_workers, pin_memory=True, sampler=train_sampler)
test_dataset = Imagenet(mode='val', aug=eval_aug, max_class=1000)
test_sampler = torch.utils.data.distributed.DistributedSampler(test_dataset)
test_loader = torch.utils.data.DataLoader(
test_dataset, batch_size=batch_size, shuffle=(test_sampler is None),
num_workers=num_workers, pin_memory=True, sampler=test_sampler)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=epochs * train_loader.__len__())
best_acc = 0
best_acc5 = 0
checkpoint_path = 'checkpoints/eval_' + args.checkpoint
print('checkpoint_path:', checkpoint_path)
if os.path.exists(checkpoint_path):
checkpoint = torch.load(checkpoint_path, map_location='cpu')
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
start_epoch = checkpoint['epoch']
else:
start_epoch = 0
model.eval()
for epoch in range(start_epoch, epochs):
train_sampler.set_epoch(epoch)
batch_time = AverageMeter('Time', ':6.3f')
data_time = AverageMeter('Data', ':6.3f')
losses = AverageMeter('Loss', ':.4e')
progress = ProgressMeter(
train_loader.__len__(),
[batch_time, data_time, losses],
prefix="Epoch: [{}]".format(epoch)
)
end = time.time()
for i, (image, label) in enumerate(train_loader):
data_time.update(time.time() - end)
image = image.cuda(local_rank, non_blocking=True)
label = label.cuda(local_rank, non_blocking=True)
out = model(image)
loss = F.cross_entropy(out, label)
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_time.update(time.time() - end)
end = time.time()
losses.update(loss.item())
if i % 20 == 0 and rank == 0:
progress.display(i)
scheduler.step()
# ---------------------- Test --------------------------
top1 = AverageMeter('Acc@1', ':6.2f')
top5 = AverageMeter('Acc@5', ':6.2f')
with torch.no_grad():
end = time.time()
for i, (image, label) in enumerate(test_loader):
image = image.cuda(local_rank, non_blocking=True)
label = label.cuda(local_rank, non_blocking=True)
# compute output
output = model(image)
# measure accuracy and record loss
acc1, acc5 = accuracy(output, label, topk=(1, 5))
top1.update(acc1[0], image.size(0))
top5.update(acc5[0], image.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
sum1, cnt1, sum5, cnt5 = torch_dist_sum(local_rank, top1.sum, top1.count, top5.sum, top5.count)
top1_acc = sum(sum1.float()) / sum(cnt1.float())
top5_acc = sum(sum5.float()) / sum(cnt5.float())
best_acc = max(top1_acc, best_acc)
best_acc5 = max(top5_acc, best_acc5)
if rank == 0:
print('Epoch:{} * Acc@1 {top1_acc:.3f} Acc@5 {top5_acc:.3f} Best_Acc@1 {best_acc:.3f} Best_Acc@5 {best_acc5:.3f}'.format(epoch, top1_acc=top1_acc, top5_acc=top5_acc, best_acc=best_acc, best_acc5=best_acc5))
torch.save(
{
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'epoch': epoch+1,
},
checkpoint_path
)
if __name__ == "__main__":
main()