-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathensv.py
315 lines (278 loc) · 14.6 KB
/
ensv.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
import argparse
import os
import random
import math
import os.path as osp
import copy
import time
from collections import OrderedDict
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 sklearn.metrics import confusion_matrix
import pre_process as prep
from data_list import ImageList
from classifier import ImageClassifier, ImageClassifierMDD, ImageClassifierAFN
from backbone import get_model
def entropy(input_):
epsilon = 1e-5
entropy = -input_ * torch.log(input_ + epsilon)
entropy = torch.sum(entropy, dim=-1)
return entropy
def init_model(args):
backbone = get_model(args.net, pretrain=True)
#pool_layer = nn.Identity() if args.no_pool else None
if args.method == 'MDD':
classifier = ImageClassifierMDD(backbone, args.class_num, bottleneck_dim=args.bottleneck_dim, width=args.width)
elif args.method == 'SAFN':
classifier = ImageClassifierAFN(backbone, args.class_num, bottleneck_dim=args.bottleneck_dim)
else:
classifier = ImageClassifier(backbone, args.class_num, bottleneck_dim=args.bottleneck_dim)
return classifier
def load_ckpt(args):
ckpt_path = args.ckpt_path
ckpt_dict = torch.load(ckpt_path)
filtered_state_dict = OrderedDict()
for k in ckpt_dict:
if 'backbone.fc' in k:
pass
else:
filtered_state_dict[k] = ckpt_dict[k]
return filtered_state_dict
def load_ckpt_mdd(args):
ckpt_path = args.ckpt_path
ckpt_dict = torch.load(ckpt_path)
filtered_state_dict = OrderedDict()
for k in ckpt_dict:
if 'fc' in k or 'adv' in k:
pass
elif 'bottleneck.1' in k:
newk = k.replace('bottleneck.1', 'bottleneck.0')
filtered_state_dict[newk] = ckpt_dict[k]
elif 'bottleneck.2' in k:
newk = k.replace('bottleneck.2', 'bottleneck.1')
filtered_state_dict[newk] = ckpt_dict[k]
else:
filtered_state_dict[k] = ckpt_dict[k]
return filtered_state_dict
def test(config, args, net, dset_path):
prep_dict = {}
prep_dict["test_tgt"] = prep.image_test(**config["prep"]["params"])
dsets = {}
dset_loaders = {}
data_config = config["data"]
test_bs = data_config["test_tgt"]["batch_size"]
dsets["test_tgt"] = ImageList(open(dset_path).readlines(), transform=prep_dict["test_tgt"])
dset_loaders["test_tgt"] = DataLoader(dsets["test_tgt"], batch_size=test_bs, shuffle=False, num_workers=4, drop_last=False)
net = net.cuda()
net.eval()
start_test = True
with torch.no_grad():
iter_test = iter(dset_loaders["test_tgt"])
for i in range(len(dset_loaders["test_tgt"])):
data = iter_test.next()
inputs = data[0]
labels = data[1]
inputs = inputs.cuda()
outputs = net(inputs)
if start_test:
all_output = outputs.float().cpu()
all_label = labels.float()
start_test = False
else:
all_output = torch.cat((all_output, outputs.float().cpu()), 0)
all_label = torch.cat((all_label, labels.float()), 0)
_, predict = torch.max(all_output, 1)
accuracy = torch.sum(torch.squeeze(predict).float() == all_label).item() / float(all_label.size()[0])
all_label = all_label.long()
all_sfmx = nn.Softmax(dim=-1)(all_output)
mean_sfmx = all_sfmx.mean(dim=0)
# entropy: min is best
mean_ent = torch.mean(entropy(all_sfmx)).item()
# im: max is best
mean_div = -torch.sum(mean_sfmx * torch.log(mean_sfmx + 1e-5)).item()
mean_im = mean_div - mean_ent
# corr-c: min is best
ori_corr = torch.mm(all_sfmx.t(), all_sfmx)
sfmxcorr = ori_corr.diag().sum().item() / ((ori_corr**2).sum()**0.5).item()
# snd: max is best
ori_normalized = F.normalize(all_output)
ori_mat = torch.matmul(ori_normalized, ori_normalized.t()) / 0.05
ori_mask = torch.eye(ori_mat.size(0), ori_mat.size(0)).bool()
ori_mat.masked_fill_(ori_mask, -1 / 0.05)
snd = entropy(ori_mat.softmax(dim=-1)).mean().item()
log_str = "Testing accuracy: {:.4f}, entropy is {:.4f}, im is {:.4f}, corr is {:.4f}, snd is {:.4f}.\n".format(accuracy, mean_ent, mean_im, sfmxcorr, snd)
acc_score = accuracy
if config["dataset"] == "visda":
matrix = confusion_matrix(all_label, torch.squeeze(predict).float())
acc = matrix.diagonal()/matrix.sum(axis=1)
cls_acc_str = ' '.join(['{:.4f}'.format(x) for x in acc.tolist()])
log_str = "Testing per-class accuracy: {}, mean-class accuracy: {:.4f}, mean accuracy: {:.4f}, mean entropy is {:.4f}.\n".format(cls_acc_str, np.mean(acc), accuracy, mean_ent)
acc_score = np.mean(acc)
config["task_hyper_log"].write(log_str)
config["task_hyper_log"].flush()
print(log_str)
if config["task_pred_file"] is not None:
torch.save(all_output, config["task_pred_file"])
return mean_ent, all_output, all_label, acc_score
def load_candidate(config, args, hyper_str):
args.ckpt_path = os.path.join('./ckpts/'+args.da, args.method, config['name'], hyper_str+'_final.pt')
if args.da == '2d_uda':
args.bottleneck_dim = int(hyper_str.split('_')[1])
args.width = args.bottleneck_dim
net = init_model(args)
if args.method == 'MDD':
net.load_state_dict(load_ckpt_mdd(args))
else:
net.load_state_dict(load_ckpt(args))
return net
def ensv(config, args):
if args.da == '2d_uda':
hyper_list = config['2d_hyper_list'][args.method]
else:
hyper_list = config['hyper_list'][args.method]
config["task_path"] = os.path.join(config["save_dir"], args.method, config['name'])
if not osp.exists(config["task_path"]):
os.system('mkdir -p '+config["task_path"])
acc_list = []
for idx in range(len(hyper_list)):
hyper_str = hyper_list[idx]
if not osp.exists(osp.join(config["task_path"], hyper_str)):
os.system('mkdir -p '+osp.join(config["task_path"], hyper_str))
config["task_hyper_log"] = open(osp.join(config["task_path"], hyper_str, hyper_str+"_log.txt"), "a+")
config["task_pred_file"] = osp.join(config["task_path"], hyper_str, hyper_str+"_pred.pt")
setting_str = "dset: {}, src: {}, tgt: {}, method: {}, net: {}, hyperparameter: {}.\n"\
.format(args.dset, names[args.s], names[args.t], args.method, args.net, hyper_str)
config["task_hyper_log"].write(setting_str)
config["task_hyper_log"].flush()
net = load_candidate(config, args, hyper_str)
ent, logits, labels, acc_score = test(config, args, net, args.t_dset_path)
acc_list.append(acc_score)
if idx == 0:
ensem_pred = nn.Softmax(dim=-1)(logits)
else:
ensem_pred += nn.Softmax(dim=-1)(logits)
print('********************Model Ensembling***************')
ensem_pred /= len(hyper_list)
_, ensem_pl = torch.max(ensem_pred, 1)
ensem_acc = torch.sum(torch.squeeze(ensem_pl).float() == labels.float()).item() / float(labels.size()[0])
log_str = "Ensemble accuracy: {:.4f}.\n".format(ensem_acc)
ensem_acc_score = ensem_acc
if config["dataset"] == "visda":
matrix = confusion_matrix(labels.long(), torch.squeeze(ensem_pl).float())
acc = matrix.diagonal()/matrix.sum(axis=1)
cls_acc_str = ' '.join(['{:.4f}'.format(x) for x in acc.tolist()])
log_str = "Ensemble per-class accuracy: {}, mean-class accuracy: {:.4f}, mean accuracy: {:.4f}.\n".format(cls_acc_str, np.mean(acc), ensem_acc)
ensem_acc_score = np.mean(acc)
print(log_str)
print('********************Model Selection***************')
score_list = []
for idx in range(len(hyper_list)):
hyper_str = hyper_list[idx]
config["task_pred_file"] = osp.join(config["task_path"], hyper_str, hyper_str+"_pred.pt")
idx_pred = torch.load(config["task_pred_file"])
idx_pl = torch.argmax(idx_pred, dim=-1)
idx_ensem_acc = (ensem_pl == idx_pl).sum()/idx_pred.shape[0]
score_list.append(idx_ensem_acc.item())
setting_str = "dset: {}, src: {}, tgt: {}, method: {}, net: {}, hyperpara: {}, tgtAcc: {:.4f}, ensAcc: {:.4f}, plAcc: {:.4f}.\n"\
.format(args.dset, names[args.s], names[args.t], args.method, args.net, hyper_str, acc_list[idx], ensem_acc_score, idx_ensem_acc.item())
print(setting_str)
config["task_avg_log"].write(setting_str)
config["task_avg_log"].flush()
best_index = score_list.index(max(score_list))
worst_index = score_list.index(min(score_list))
best_acc = acc_list[best_index]
worst_acc = acc_list[worst_index]
log_str = "dset: {}, src: {}, tgt: {}, method: {}, net: {}, numOfHyper: {}, bestAcc: {:.4f}, bestIdx: {}, worstAcc: {:.4f}, worstIdx: {}.\n"\
.format(args.dset, names[args.s], names[args.t], args.method, args.net, len(hyper_list), best_acc, best_index, worst_acc, worst_index)
print(log_str)
config["task_avg_log"].write(setting_str)
config["task_avg_log"].flush()
return best_index, worst_index, best_acc, worst_acc
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Model selection for unsupervised domain adaptation')
# task parameters
parser.add_argument('--gpu_id', type=str, nargs='?', default='0', help="device id to run")
parser.add_argument('--dset', type=str, default='office-home', choices=['office', 'image-clef', 'visda', 'office-home', 'DomainNet126'], help="dataset")
parser.add_argument('--seed', type=int, default=123, help="seed")
parser.add_argument('--s', type=int, default=0, help="source")
parser.add_argument('--t', type=int, default=1, help="target")
parser.add_argument('--bs', type=int, default=128, help='batch size')
parser.add_argument('--da', type=str, default='uda', choices=['uda', 'pda', '2d_uda', 'opda'])
# model parameters
parser.add_argument('--net', type=str, default='resnet50', choices=["resnet18", "resnet34", "resnet50", "resnet101"])
parser.add_argument('--bottleneck_dim', type=int, default=256)
parser.add_argument('--width', type=int, default=2048, help="for mdd")
parser.add_argument('--method', type=str, default='CDAN', choices=['CDAN', 'MCC', 'BNM', 'MDD', 'ATDOC', 'SAFN', 'PADA'])
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
SEED = args.seed
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
np.random.seed(SEED)
random.seed(SEED)
s = args.s
t = args.t
if args.dset == 'office-home':
names = ['Art', 'Clipart', 'Product', 'RealWorld']
args.s_dset_path = './data/' + args.dset + '/' + names[s] + '_list.txt'
if args.da in {'uda', '2d_uda'}:
args.t_dset_path = './data/' + args.dset + '/' + names[t] + '_list.txt'
elif args.da == 'pda':
args.t_dset_path = './data/' + args.dset + '/' + names[t] + '_25_list.txt'
args.class_num = 65
if args.dset == 'office':
names = ['amazon', 'dslr', 'webcam']
args.s_dset_path = './data/' + args.dset + '/' + names[s] + '_list.txt'
if args.da in {'uda', '2d_uda'}:
args.t_dset_path = './data/' + args.dset + '/' + names[t] + '_list.txt'
elif args.da == 'pda':
args.t_dset_path = './data/' + args.dset + '/' + names[t] + '_10_list.txt'
args.class_num = 31
if args.dset == 'visda':
names = ['training', 'validation']
args.s_dset_path = './data/visda17/train_list.txt'
args.t_dset_path = './data/visda17/validation_list.txt'
args.class_num = 12
config = {}
config['hyper_list'] = {
'ATDOC': ['0.02', '0.05', '0.1', '0.2', '0.5', '1.0', '2.0'],
'BNM': ['0.02', '0.05', '0.1', '0.2', '0.5', '1.0', '2.0'],
'CDAN': ['0.05', '0.1', '0.2', '0.5', '1.0', '2.0', '5.0'],
'MCC': ['1.0', '1.5', '2.0', '2.5', '3.0', '3.5', '4.0'],
'MDD': ['0.5', '1.0', '2.0', '3.0', '4.0', '5.0', '6.0'],
'SAFN': ['0.002', '0.005', '0.01', '0.02', '0.05', '0.1', '0.2'],
'PADA': ['0.05', '0.1', '0.2', '0.5', '1.0', '2.0', '5.0']
}
config['2d_hyper_list'] = {
'MCC': ['1.0_256', '1.0_512', '1.0_1024', '1.0_2048', '1.5_256', '1.5_512', '1.5_1024', '1.5_2048',
'2.0_256', '2.0_512', '2.0_1024', '2.0_2048', '2.5_256', '2.5_512', '2.5_1024', '2.5_2048',
'3.0_256', '3.0_512', '3.0_1024', '3.0_2048', '3.5_256', '3.5_512', '3.5_1024', '3.5_2048',
'4.0_256', '4.0_512', '4.0_1024', '4.0_2048'],
'MDD': ['1.0_256', '1.0_512', '1.0_1024', '1.0_2048', '2.0_256', '2.0_512', '2.0_1024', '2.0_2048',
'3.0_256', '3.0_512', '3.0_1024', '3.0_2048', '4.0_256', '4.0_512', '4.0_1024', '4.0_2048',
'5.0_256', '5.0_512', '5.0_1024', '5.0_2048', '6.0_256', '6.0_512', '6.0_1024', '6.0_2048',
'7.0_256', '7.0_512', '7.0_1024', '7.0_2048']
}
config['visda'] = (args.dset == 'visda')
config['method'] = args.method
config["gpu"] = args.gpu_id
config['name'] = args.dset + '/' + names[s][0].upper() + names[t][0].upper()
config["save_dir"] = os.path.join('./logs/'+args.da, args.da)
if not osp.exists(config["save_dir"]):
os.system('mkdir -p '+config["save_dir"])
config["task_avg_log"] = open(osp.join(config["save_dir"], args.da+"_avglog.txt"), "a+")
config["prep"] = {'params':{"resize_size":256, "crop_size":224, "alexnet":False}}
config["dataset"] = args.dset
# 36 100 96
config["data"] = {"source":{"list_path":args.s_dset_path, "batch_size":args.bs}, \
"target":{"list_path":args.t_dset_path, "batch_size":args.bs}, \
"test_tgt":{"list_path":args.t_dset_path, "batch_size":args.bs}, \
"test_src":{"list_path":args.s_dset_path, "batch_size":args.bs}}
setting_str = "dset: {}, src: {}, tgt: {}, method: {}, net: {}.\n"\
.format(args.dset, names[args.s], names[args.t], args.method, args.net)
print(setting_str)
ensv(config, args)