-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathstudent_v0.py
255 lines (206 loc) · 9.41 KB
/
student_v0.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
import os
import os.path as osp
import argparse
import time
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import MultiStepLR
import torchvision.transforms as transforms
from tensorboardX import SummaryWriter
from models import model_dict
from zoo import Attention, Similarity
from dataset import CIFAR100
from utils import accuracy, val, AverageMeter, cal_center
items = ['acc', 'loss', \
's_select_confidence', 's_select_margin', 's_select_entropy', \
's_else_confidence', 's_else_margin', 's_else_entropy', \
's_all_confidence', 's_all_margin', 's_all_entropy', \
't_confidence', 't_margin', 't_entropy', \
'center_dist']
parser = argparse.ArgumentParser(description='train student network.')
parser.add_argument('--epoch', type=int, default=240)
parser.add_argument('--batch-size', type=int, default=64)
parser.add_argument('--k', type=int, default=48)
parser.add_argument('--b', type=int, default=32)
parser.add_argument('--w', type=float, default=1000)
parser.add_argument('--lr', type=float, default=0.05)
parser.add_argument('--momentum', type=float, default=0.9)
parser.add_argument('--weight-decay', type=float, default=5e-4)
parser.add_argument('--milestones', type=float, nargs='+', default=[150, 180, 210])
parser.add_argument('--teacher-path', type=str, default='./experiments/teacher_resnet32x4')
parser.add_argument('--teacher-ckpt', type=str, default='best')
parser.add_argument('--student-arch', type=str, default='resnet8x4')
parser.add_argument('--ce-weight', type=float, default=0.0)
parser.add_argument('--kd-weight', type=float, default=1.0)
parser.add_argument('--other-distill', type=str, choices=['AT', 'SP'], default=None)
parser.add_argument('--T', type=float, default=4.0)
parser.add_argument('--strategy', type=int, choices=[0,1,2,3], default=3)
# 0: random, 1: least confidence, 2: margin, 3: entropy
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--gpu-id', type=int, default=0)
args = parser.parse_args()
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_id)
torch.backends.cudnn.benchmark = True
teacher_arch = '_'.join(args.teacher_path.split('/')[-1].split('_')[1:])
exp_name = '{}_student_{}_teacher_{}-{}_strategy{}_k{}_b{}_w{}_seed{}'.format(\
__file__.split('.')[0].split('_')[-1], \
args.student_arch, teacher_arch, args.teacher_ckpt, \
args.strategy, \
args.k, args.b, args.w, \
args.seed)
exp_path = './experiments/{}'.format(exp_name)
os.makedirs(exp_path, exist_ok=True)
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5071, 0.4866, 0.4409], std=[0.2675, 0.2565, 0.2761]),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.5071, 0.4866, 0.4409], std=[0.2675, 0.2565, 0.2761]),
])
trainset = CIFAR100('./data', train=True, transform=transform_train)
valset = CIFAR100('./data', train=False, transform=transform_test)
num_classes = 100
train_loader = DataLoader(trainset, batch_size=args.batch_size, \
shuffle=True, num_workers=3, pin_memory=True)
val_loader = DataLoader(valset, batch_size=args.batch_size, \
shuffle=False, num_workers=3, pin_memory=True)
ckpt_path = osp.join('{}/ckpt/{}.pth'.format( \
args.teacher_path, args.teacher_ckpt))
t_model = model_dict[teacher_arch](num_classes=num_classes).cuda()
state_dict = torch.load(ckpt_path)['state_dict']
t_model.load_state_dict(state_dict)
t_model.eval()
logger = SummaryWriter(osp.join(exp_path, 'events'))
s_model = model_dict[args.student_arch](num_classes=num_classes).cuda()
optimizer = optim.SGD(s_model.parameters(), lr=args.lr, \
momentum=args.momentum, weight_decay=args.weight_decay)
scheduler = MultiStepLR(optimizer, milestones=args.milestones)
if args.other_distill is not None:
if args.other_distill == 'AT':
criterion = Attention()
weight = 1000
elif args.other_distill == 'SP':
criterion = Similarity()
weight = 3000
best_acc = 0
counter = torch.zeros(args.epoch, 50000).cuda()
epoch = 0
for epoch in range(args.epoch):
record = {name:AverageMeter() for name in items}
center = cal_center(val_loader, args, s_model)
for fuck, (x, y, k) in enumerate(train_loader):
s_model.train()
x = x.cuda()
y = y.cuda()
k = k.cuda()
with torch.no_grad():
s_feats, logits = s_model(x, is_feat=True)
probs = F.softmax(logits, dim=1)
# confidence
conf = probs.max(dim=1)[0]
# margin
rank = torch.argsort(probs, dim=1)
top2 = torch.gather(probs, dim=1, index=rank[:,-2:])
margin = top2[:,-1] - top2[:,-2]
# entropy
entropy = -torch.sum(probs * torch.log(probs), dim=1)
if args.strategy == 0:
scores = torch.rand(x.size(0)).cuda()
elif args.strategy == 1:
scores = 1 - conf
elif args.strategy == 2:
scores = -margin
elif args.strategy == 3:
scores = entropy
else:
raise ValueError('Invalid strategy.')
r = torch.arange(x.size(0)).float()
m = (2*args.b-1) / (2*args.batch_size)
mask_proto = 1 / (1 + torch.exp(-args.w * (r/args.batch_size - m) ))
mask_proto = mask_proto.cuda()
lamb = np.random.beta(1, 1)
mask = lamb * mask_proto.view(-1, 1, 1, 1)
rank = torch.argsort(scores, descending=True)
index = torch.randperm(x.size(0)).cuda()
x = (1-mask) * x[rank] + mask * x[index]
x = x[:args.k]
counter[epoch, k[rank[:args.b]] ] += 1
s_feats, s_logits = s_model(x, is_feat=True)
with torch.no_grad():
t_feats, t_logits = t_model(x, is_feat=True)
## for statistics
t_probs = F.softmax(t_logits, dim=1)
# confidence
t_conf = t_probs.max(dim=1)[0]
# margin
t_rank = torch.argsort(t_probs, dim=1)
t_top2 = torch.gather(t_probs, dim=1, index=t_rank[:,-2:])
t_margin = t_top2[:,-1] - t_top2[:,-2]
# entropy
t_entropy = -torch.sum(t_probs * torch.log(t_probs), dim=1)
# compute loss
log_s_probs = F.log_softmax(s_logits / args.T, dim=1)
t_probs = F.softmax(t_logits / args.T, dim=1)
tmp = mask.squeeze()[:args.k]
loss_ce = F.cross_entropy(s_logits, y[rank][:args.k], reduction='none') * (1-tmp) + \
F.cross_entropy(s_logits, y[index][:args.k], reduction='none') * tmp
loss_kd = F.kl_div(log_s_probs, t_probs, reduction='batchmean') * args.T * args.T
if args.other_distill is not None:
loss_other = sum(criterion(s_feats[1:-1], t_feats[1:-1])) if args.other_distill == 'AT' \
else sum(criterion(s_feats[-2], t_feats[-2]))
loss = args.ce_weight * loss_ce.mean() + args.kd_weight * loss_kd + weight * loss_other
else:
loss = args.ce_weight * loss_ce.mean() + args.kd_weight * loss_kd
# BP
optimizer.zero_grad()
loss.backward()
optimizer.step()
# compute distance between samples and center
C = center[y[rank[:args.k]]]
S = s_feats[-1]
D = torch.pow(C-S, 2).sum(dim=1).sqrt().mean()
record['center_dist'].update(D.item(), rank[:args.k].size(0))
batch_acc = accuracy(logits, y, topk=(1,))[0]
record['acc'].update(batch_acc.item(), logits.size(0))
record['loss'].update(loss.item(), s_logits.size(0))
i = rank[:args.k].size(0)
record['s_select_confidence'].update(conf[rank[:args.k]].mean().item(), i)
record['s_select_margin'].update(margin[rank[:args.k]].mean().item(), i)
record['s_select_entropy'].update(entropy[rank[:args.k]].mean().item(), i)
i = rank[args.k:].size(0)
if i > 0:
record['s_else_confidence'].update(conf[rank[args.k:]].mean().item(), i)
record['s_else_margin'].update(margin[rank[args.k:]].mean().item(), i)
record['s_else_entropy'].update(entropy[rank[args.k:]].mean().item(), i)
i = conf.size(0)
record['s_all_confidence'].update(conf.mean().item(), i)
record['s_all_margin'].update(margin.mean().item(), i)
record['s_all_entropy'].update(entropy.mean().item(), i)
i = t_conf.size(0)
record['t_confidence'].update(t_conf.mean().item(), i)
record['t_margin'].update(t_margin.mean().item(), i)
record['t_entropy'].update(t_entropy.mean().item(), i)
for item in items:
logger.add_scalar('train/{}'.format(item), record[item].avg, epoch+1)
# val
acc = val(val_loader, args, t_model, s_model, logger, epoch)
if acc > best_acc:
best_acc = acc
state_dict = dict(state_dict=s_model.state_dict(), best_acc=best_acc)
name = osp.join(exp_path, 'ckpt/student_best.pth')
os.makedirs(osp.dirname(name), exist_ok=True)
torch.save(state_dict, name)
scheduler.step()
if args.seed ==0 :
counter = counter.cpu().numpy()
np.save(osp.join(exp_path, 'counter.npy'), counter)