-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
341 lines (279 loc) · 13.6 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
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
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
import torch
import torch.optim as optim
import torch.nn as nn
import model_factory as model_no
import adversarial as ad
import numpy as np
import argparse
from data_list import ImageList
import pre_process as prep
import criterion_factory as cf
import sampler
def test_target(loader, model):
with torch.no_grad():
start_test = True
iter_val = [iter(loader['val' + str(i)]) for i in range(10)]
for i in range(len(loader['val0'])):
data = [iter_val[j].next() for j in range(10)]
inputs = [data[j][0] for j in range(10)]
labels = data[0][1]
for j in range(10):
inputs[j] = inputs[j].cuda()
labels = labels.cuda()
outputs = []
for j in range(10):
_, output = model(inputs[j])
outputs.append(output)
outputs = sum(outputs)
if start_test:
all_output = outputs.data.float()
all_label = labels.data.float()
start_test = False
else:
all_output = torch.cat((all_output, outputs.data.float()), 0)
all_label = torch.cat((all_label, labels.data.float()), 0)
_, predict = torch.max(all_output, 1)
accuracy = torch.sum(torch.squeeze(predict).float() == all_label).item() / float(all_label.size()[0])
return accuracy
def inv_lr_scheduler(param_lr, optimizer, iter_num, gamma, power, init_lr=0.001):
lr = init_lr * (1 + gamma * iter_num) ** (-power)
i = 0
for param_group in optimizer.param_groups:
param_group['lr'] = lr * param_lr[i]
i += 1
return optimizer
class discriminatorDANN(nn.Module):
def __init__(self, feature_len):
super(discriminatorDANN, self).__init__()
self.relu = nn.ReLU()
self.dropout = nn.Dropout(0.5)
self.ad_layer1 = nn.Linear(feature_len, 1024)
self.ad_layer1.weight.data.normal_(0, 0.01)
self.ad_layer1.bias.data.fill_(0.0)
self.fc1 = nn.Sequential(self.ad_layer1, nn.ReLU(), nn.Dropout(0.5))
self.ad_layer2 = nn.Linear(1024, 1024)
self.ad_layer2.weight.data.normal_(0, 0.01)
self.ad_layer2.bias.data.fill_(0.0)
self.ad_layer3 = nn.Linear(1024, 1)
self.ad_layer3.weight.data.normal_(0, 0.3)
self.ad_layer3.bias.data.fill_(0.0)
self.fc2_3 = nn.Sequential(self.ad_layer2, nn.ReLU(), nn.Dropout(0.5), self.ad_layer3, nn.Sigmoid())
def forward(self, x, y):
f2 = self.fc1(x)
f = self.fc2_3(f2)
return f
class discriminatorCDAN(nn.Module):
def __init__(self, feature_len):
super(discriminatorCDAN, self).__init__()
self.ad_layer1 = nn.Linear(feature_len * 31, 1024)
self.ad_layer1.weight.data.normal_(0, 0.01)
self.ad_layer1.bias.data.fill_(0.0)
self.fc1 = nn.Sequential(self.ad_layer1, nn.ReLU(), nn.Dropout(0.5))
self.ad_layer2 = nn.Linear(1024, 1024)
self.ad_layer2.weight.data.normal_(0, 0.01)
self.ad_layer2.bias.data.fill_(0.0)
self.ad_layer3 = nn.Linear(1024, 1)
self.ad_layer3.weight.data.normal_(0, 0.3)
self.ad_layer3.bias.data.fill_(0.0)
self.fc2_3 = nn.Sequential(self.ad_layer2, nn.ReLU(), nn.Dropout(0.5), self.ad_layer3, nn.Sigmoid())
def forward(self, x, y):
op_out = torch.bmm(y.unsqueeze(2), x.unsqueeze(1))
ad_in = op_out.view(-1, y.size(1) * x.size(1))
f2 = self.fc1(ad_in)
f = self.fc2_3(f2)
return f
class predictor(nn.Module):
def __init__(self, feature_len, cate_num):
super(predictor, self).__init__()
self.classifier = nn.Linear(feature_len, cate_num)
self.classifier.weight.data.normal_(0, 0.01)
self.classifier.bias.data.fill_(0.0)
def forward(self, features):
activations = self.classifier(features)
return (activations)
class net(nn.Module):
def __init__(self, feature_len):
super(net, self).__init__()
self.model_fc = model_no.Resnet50Fc()
self.bottleneck_0 = nn.Linear(feature_len, 256)
self.bottleneck_0.weight.data.normal_(0, 0.005)
self.bottleneck_0.bias.data.fill_(0.1)
self.bottleneck_layer = nn.Sequential(self.bottleneck_0, nn.ReLU(), nn.Dropout(0.5))
self.classifier_layer = predictor(256, num_categories)
def forward(self, x):
features = self.model_fc(x)
out_bottleneck = self.bottleneck_layer(features)
logits = self.classifier_layer(out_bottleneck)
return (out_bottleneck, logits)
def get_parameters(self):
parameter_list = [{"params": self.model_fc.parameters(), "lr_mult": 0.1}, \
{"params": self.bottleneck_layer.parameters(), "lr_mult": 1}, \
{"params": self.classifier_layer.parameters(), "lr_mult": 1}]
return parameter_list
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Transfer Learning')
parser.add_argument('--source', type=str, nargs='?', default='c', help="source dataset")
parser.add_argument('--target', type=str, nargs='?', default='p', help="target dataset")
parser.add_argument('--lr', type=float, nargs='?', default=0.03, help="target dataset")
parser.add_argument('--max_iteration', type=float, nargs='?', default=12500, help="target dataset")
parser.add_argument('--k', type=int, default=3, help="k")
parser.add_argument('--test_batch_size', type=int, default=4, help="test batch size")
parser.add_argument('--n_samples', type=int, default=2, help='number of samples from each src class')
parser.add_argument('--multi_gpu', type=int, default=0, help="use dataparallel if 1")
parser.add_argument('--msc_coeff', type=float, default=1.0, help="coeff for similarity loss")
parser.add_argument('--pre_train', type=int, default=0, help="number of iterations to pretrain for")
parser.add_argument('--ila_switch_iter', type=int, default=0,
help="number of iterations when only DA loss works and sim doesn't")
parser.add_argument('--method', type=str, nargs='?', default='DANN', choices=['DANN', 'CDAN'])
parser.add_argument('--mu', type=int, default=80,
help="these many target samples are used finally, eg. 2/3 of batch") #mu in number
args = parser.parse_args()
args.multi_gpu = bool(args.multi_gpu)
num_categories = 31
file_path = {
"i": "/vulcan-pvc1/ml_for_da_pan_base/fg_dataset_list/bird31_ina_list_2017.txt",
"n": "/vulcan-pvc1/ml_for_da_pan_base/fg_dataset_list/bird31_nabirds_list.txt",
"c": "/vulcan-pvc1/ml_for_da_pan_base/fg_dataset_list/bird31_cub2011.txt",
}
dataset_source = file_path[args.source]
dataset_target = dataset_test = file_path[args.target]
batch_size = {"train": args.n_samples * num_categories, "val": args.n_samples * num_categories, "test": args.test_batch_size}
for i in range(10):
batch_size["val" + str(i)] = 4
src_train_sampler = sampler.get_sampler({
'path' : dataset_source,
'n_classes' : num_categories,
'n_samples' : args.n_samples,
})
dataset_loaders = {}
dataset_list = ImageList(open(dataset_source).readlines(),
transform=prep.image_train(resize_size=256, crop_size=224))
dataset_loaders["train"] = torch.utils.data.DataLoader(dataset_list, batch_sampler=src_train_sampler, \
shuffle=False, num_workers=16)
dataset_list = ImageList(open(dataset_target).readlines(),
transform=prep.image_train(resize_size=256, crop_size=224))
dataset_loaders["val"] = torch.utils.data.DataLoader(dataset_list, batch_size=batch_size['train'], shuffle=True,
num_workers=16, drop_last=True)
prep_dict_test = prep.image_test_10crop(resize_size=256, crop_size=224)
for i in range(10):
dataset_list = ImageList(open(dataset_test).readlines(), transform=prep_dict_test["val" + str(i)])
dataset_loaders["val" + str(i)] = torch.utils.data.DataLoader(dataset_list,
batch_size=batch_size["val" + str(i)],
shuffle=False, num_workers=16)
# network construction
feature_len = 2048
my_net = net(feature_len)
my_net = my_net.cuda()
if args.multi_gpu:
print('USING MULTI GPU')
my_net = nn.DataParallel(my_net)
my_net_accr = my_net.module
else:
print('NOT USING MULTI GPU')
my_net_accr = my_net
my_net.train(True)
if args.method == 'DANN':
my_discriminator = discriminatorDANN(256)
elif args.method == 'CDAN':
my_discriminator = discriminatorCDAN(256)
else:
raise Exception('{} not implemented'.format(args.method))
# domain discriminator
my_discriminator = my_discriminator.cuda()
my_discriminator.train(True)
# gradient reversal layer
my_grl = ad.AdversarialLayer()
msc_config = {
'k' : args.k,
'm' : args.n_samples,
'mu' : args.mu,
}
msc_module = cf.MSCLoss(msc_config)
msc_module = msc_module.cuda()
# criterion and optimizer
criterion = {
"classifier" : nn.CrossEntropyLoss(),
"adversarial": nn.BCELoss()
}
optimizer_dict = [
{"params": filter(lambda p: p.requires_grad, my_net_accr.model_fc.parameters()), "lr": 0.1},
{"params": filter(lambda p: p.requires_grad, my_net_accr.bottleneck_layer.parameters()), "lr": 1},
{"params": filter(lambda p: p.requires_grad, my_net_accr.classifier_layer.parameters()), "lr": 1},
{"params": filter(lambda p: p.requires_grad, my_discriminator.parameters()), "lr": 1} # ,
]
optimizer = optim.SGD(optimizer_dict, lr=0.1, momentum=0.9, weight_decay=0.0005)
optimizer_dict_pre = [
{"params": filter(lambda p: p.requires_grad, my_net_accr.model_fc.parameters()), "lr": 0.1},
{"params": filter(lambda p: p.requires_grad, my_net_accr.bottleneck_layer.parameters()), "lr": 1},
{"params": filter(lambda p: p.requires_grad, my_net_accr.classifier_layer.parameters()), "lr": 1}
]
optimizer_pre = optim.SGD(optimizer_dict_pre, lr=0.1, momentum=0.9, weight_decay=0.0005)
param_lr_pre = []
for param_group in optimizer_pre.param_groups:
param_lr_pre.append(param_group["lr"])
param_lr = []
for param_group in optimizer.param_groups:
param_lr.append(param_group["lr"])
len_source = len(dataset_loaders["train"]) - 1
len_target = len(dataset_loaders["val"]) - 1
iter_source = iter(dataset_loaders["train"])
iter_target = iter(dataset_loaders["val"])
# pre_train
for pt in range(1, args.pre_train + 1):
print('pre_train iter {}'.format(pt))
my_net.train(True)
optimizer_pre = inv_lr_scheduler(param_lr_pre, optimizer_pre, pt, init_lr=args.lr, gamma=0.001, power=0.75)
optimizer_pre.zero_grad()
if pt % len_source == 0:
iter_source = iter(dataset_loaders["train"])
data_source = iter_source.next()
inputs_source, labels_source = data_source
inputs_source, labels_source = inputs_source.cuda(), labels_source.cuda()
all_features, logits = my_net(inputs_source)
classifier_loss = criterion["classifier"](logits, labels_source)
classifier_loss.backward()
optimizer_pre.step()
len_source = len(dataset_loaders["train"]) - 1
len_target = len(dataset_loaders["val"]) - 1
iter_source = iter(dataset_loaders["train"])
iter_target = iter(dataset_loaders["val"])
for iter_num in range(1, args.max_iteration + 1):
my_net.train(True)
optimizer = inv_lr_scheduler(param_lr, optimizer, iter_num, init_lr=args.lr, gamma=0.001, power=0.75)
optimizer.zero_grad()
if iter_num % len_source == 0:
iter_source = iter(dataset_loaders["train"])
if iter_num % len_target == 0:
iter_target = iter(dataset_loaders["val"])
data_source = iter_source.next()
data_target = iter_target.next()
inputs_source, labels_source = data_source
inputs_target, labels_target = data_target
inputs = torch.cat((inputs_source, inputs_target), dim=0)
inputs = inputs.cuda()
labels_target = labels_target.cuda()
labels_source = labels_source.cuda()
domain_labels = torch.from_numpy(
np.array([[1], ] * len(inputs_source)+ [[0], ] * len(inputs_target))).float()
domain_labels = domain_labels.cuda()
all_features, logits = my_net(inputs)
src_logits = logits[:len(inputs_source)]
src_features = all_features[:len(inputs_source)]
tgt_features = all_features[len(inputs_source):]
classifier_loss = criterion["classifier"](src_logits, labels_source)
max_iter = args.max_iteration
domain_predicted = my_discriminator(my_grl.apply(all_features), nn.softmax(dim=1)(logits).detach())
transfer_loss = nn.bceloss()(domain_predicted, domain_labels)
total_loss = classifier_loss + transfer_loss
if iter_num > args.ila_switch_iter:
msc_loss = msc_module(src_features, labels_source, tgt_features)
total_loss += (args.msc_coeff * msc_loss)
total_loss.backward()
optimizer.step()
# test
test_interval = 500
if iter_num % test_interval == 0:
my_net.eval()
test_acc = test_target(dataset_loaders, my_net)
stats = 'iter: {:05d}, test_acc:{:.4f}\n'.format(iter_num, test_acc)
print(stats)