-
Notifications
You must be signed in to change notification settings - Fork 3
/
main.py
274 lines (234 loc) · 11.7 KB
/
main.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
import argparse
import warnings
from models import *
from layers import *
from loss import *
import torch
import scipy.io as sio
warnings.filterwarnings("ignore")
parser = argparse.ArgumentParser(description='CVCLNet')
parser.add_argument('--load_model', default=False, help='Testing if True or training.')
parser.add_argument('--save_model', default=False, help='Saving the model after training.')
parser.add_argument('--db', type=str, default='MSRCv1',
choices=['MSRCv1', 'MNIST-USPS', 'COIL20', 'scene', 'hand', 'Fashion', 'BDGP'],
help='dataset name')
parser.add_argument('--seed', type=int, default=10, help='Initializing random seed.')
parser.add_argument("--mse_epochs", default=200, help='Number of epochs to pre-training.')
parser.add_argument("--con_epochs", default=100, help='Number of epochs to fine-tuning.')
parser.add_argument('-lr', '--learning_rate', type=float, default=0.0005, help='Initializing learning rate.')
parser.add_argument('--weight_decay', type=float, default=0., help='Initializing weight decay.')
parser.add_argument("--temperature_l", type=float, default=1.0)
parser.add_argument('--batch_size', default=100, type=int,
help='The total number of samples must be evenly divisible by batch_size.')
parser.add_argument('--normalized', type=bool, default=False)
parser.add_argument('--gpu', default='0', type=str, help='GPU device idx.')
args = parser.parse_args()
print("==========\nArgs:{}\n==========".format(args))
# torch.cuda.set_device(0)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
device = 'cuda' if torch.cuda.is_available() else 'cpu'
def set_seed(seed):
np.random.seed(seed)
random.seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
if __name__ == "__main__":
if args.db == "MSRCv1":
# db checked 97.62
args.learning_rate = 0.0005
args.batch_size = 35
args.con_epochs = 400
args.seed = 10
args.normalized = False
dim_high_feature = 2000
dim_low_feature = 1024
dims = [256, 512]
lmd = 0.01
beta = 0.005
elif args.db == "MNIST-USPS":
# db checked 99.7
args.learning_rate = 0.0001
args.batch_size = 50
args.seed = 10
args.con_epochs = 200
args.normalized = False
dim_high_feature = 1500
dim_low_feature = 1024
dims = [256, 512, 1024]
lmd = 0.05
beta = 0.05
elif args.db == "COIL20":
# db checked 84.65
args.learning_rate = 0.0005
args.batch_size = 180
args.seed = 50
args.con_epochs = 400
args.normalized = False
dim_high_feature = 768
dim_low_feature = 200
dims = [256, 512, 1024, 2048]
lmd = 0.01
beta = 0.01
elif args.db == "scene":
# db checked 44.59
args.learning_rate = 0.0005
args.con_epochs = 100
args.batch_size = 69
args.seed = 10
args.normalized = False
dim_high_feature = 1500
dim_low_feature = 256
dims = [256, 512, 1024, 2048]
lmd = 0.01
beta = 0.05
elif args.db == "hand":
# db checked 96.85
args.learning_rate = 0.0001
args.batch_size = 200
args.seed = 50
args.con_epochs = 200
args.normalized = True
dim_high_feature = 1024
dim_low_feature = 1024
dims = [256, 512, 1024]
lmd = 0.005
beta = 0.001
elif args.db == "Fashion":
# db checked 99.31
args.learning_rate = 0.0005
args.batch_size = 100
args.con_epochs = 100
args.seed = 20
args.normalized = True
args.temperature_l = 0.5
dim_high_feature = 2000
dim_low_feature = 500
dims = [256, 512]
lmd = 0.005
beta = 0.005
elif args.db == "BDGP":
# db checked 99.2
args.learning_rate = 0.0001
args.batch_size = 250
args.seed = 10
args.con_epochs = 100
args.normalized = True
dim_high_feature = 2000
dim_low_feature = 1024
dims = [256, 512]
lmd = 0.01
beta = 0.01
set_seed(args.seed)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
mv_data = MultiviewData(args.db, device)
num_views = len(mv_data.data_views)
num_samples = mv_data.labels.size
num_clusters = np.unique(mv_data.labels).size
input_sizes = np.zeros(num_views, dtype=int)
for idx in range(num_views):
input_sizes[idx] = mv_data.data_views[idx].shape[1]
t = time.time()
# neural network architecture
mnw = CVCLNetwork(num_views, input_sizes, dims, dim_high_feature, dim_low_feature, num_clusters)
# filling it into GPU
mnw = mnw.to(device)
mvc_loss = DeepMVCLoss(args.batch_size, num_clusters)
optimizer = torch.optim.Adam(mnw.parameters(), lr=args.learning_rate, weight_decay=args.weight_decay)
if args.load_model:
state_dict = torch.load('./models/CVCL_pytorch_model_%s.pth' % args.db)
mnw.load_state_dict(state_dict)
else:
pre_train_loss_values = pre_train(mnw, mv_data, args.batch_size, args.mse_epochs, optimizer)
# sio.savemat('pre_train_loss_%s.mat' % args.db, {'data': pre_train_loss_values})
t = time.time()
fine_tuning_loss_values = np.zeros(args.con_epochs, dtype=np.float64)
for epoch in range(args.con_epochs):
total_loss = contrastive_train(mnw, mv_data, mvc_loss, args.batch_size, lmd, beta,
args.temperature_l, args.normalized, epoch, optimizer)
fine_tuning_loss_values[epoch] = total_loss
# if epoch > 0 and (epoch % 50 == 0 or epoch == args.con_epochs - 1):
# acc, nmi, pur, ari = valid(mnw, mv_data, args.batch_size)
# with open('result_%s.txt' % args.db, 'a+') as f:
# f.write('{} \t {} \t {} \t {} \t {} \t {} \t {} \t {:.6f} \t {:.4f} \n'.format(
# dim_high_feature, dim_low_feature, args.seed, args.batch_size,
# args.learning_rate, args.temperature_l, lmd, acc, (time.time() - t)))
# f.flush()
# sio.savemat('fine_tuning_loss_%s.mat' % args.db, {'data': fine_tuning_loss_values})
print("contrastive_train finished.")
print("Total time elapsed: {:.2f}s".format(time.time() - t))
if args.save_model:
torch.save(mnw.state_dict(), './models/CVCL_pytorch_model_%s.pth' % args.db)
acc, nmi, pur, ari = valid(mnw, mv_data, args.batch_size)
with open('result_%s.txt' % args.db, 'a+') as f:
f.write('{} \t {} \t {} \t {} \t {} \t {} \t {} \t {:.6f} \t {:.6f} \t {:.6f} \t {:.4f} \n'.format(
dim_high_feature, dim_low_feature, args.seed, args.batch_size,
args.learning_rate, lmd, beta, acc, nmi, pur, (time.time() - t)))
f.flush()
# dim_high_features = np.array([2000, 1500, 1024, 1000, 768, 512, 500, 256, 200], dtype=np.int32)
# dim_low_features = np.array([2000, 1500, 1024, 1000, 768, 512, 500, 256, 200], dtype=np.int32)
# seeds = np.array([10, 20, 50], dtype=np.int32)
# # dims_layers = np.array([[256, 512, 1024]])
# # dims_layers = np.array([[256, 512], [256, 512, 1024], [256, 512, 1024, 2048]])
# dims_layers = [[256, 512], [256, 512, 1024], [256, 512, 1024, 2048]]
# batch_sizes = np.array([20, 30, 50, 60], dtype=np.int32)
# lambdas = np.array([0.005, 0.01, 0.05], dtype=np.float32)
# betas = np.array([0.005, 0.01, 0.05], dtype=np.float32)
# learning_rates = np.array([0.0001, 0.0005], dtype=np.float32)
# for dh_idx in range(dim_high_features.shape[0]):
# dim_high_feature = dim_high_features[dh_idx]
# for dl_idx in range(dh_idx, dim_low_features.shape[0]):
# dim_low_feature = dim_low_features[dl_idx]
# for sd_idx in range(seeds.shape[0]):
# seed = seeds[sd_idx]
# for dim_idx in range(len(dims_layers)):
# dims = np.array(dims_layers[dim_idx])
# for bs_idx in range(batch_sizes.shape[0]):
# batch_size = int(batch_sizes[bs_idx])
# for lmd_idx in range(lambdas.shape[0]):
# lmd = lambdas[lmd_idx]
# for beta_idx in range(betas.shape[0]):
# beta = betas[beta_idx]
# for lr_idx in range(learning_rates.shape[0]):
# learning_rate = learning_rates[lr_idx]
#
# set_seed(args.seed)
# device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# mv_data = MultiviewData(args.db, device)
# num_views = len(mv_data.data_views)
# num_samples = mv_data.labels.size
# num_clusters = np.unique(mv_data.labels).size
#
# input_sizes = np.zeros(num_views, dtype=int)
# for idx in range(num_views):
# input_sizes[idx] = mv_data.data_views[idx].shape[1]
#
# t = time.time()
# # neural network architecture
# mnw = CVCLNetwork(num_views, input_sizes, dims, dim_high_feature,
# dim_low_feature, num_clusters)
# # filling it into GPU
# mnw = mnw.to(device)
#
# mvc_loss = DeepMVCLoss(batch_size, num_clusters)
# optimizer = torch.optim.Adam(mnw.parameters(), lr=learning_rate,
# weight_decay=args.weight_decay)
# pre_train(mnw, mv_data, batch_size, args.mse_epochs, optimizer)
#
# for epoch in range(args.con_epochs):
# total_loss = contrastive_train(mnw, mv_data, mvc_loss, batch_size, lmd,
# beta, args.temperature_l, args.normalized,
# epoch, optimizer)
#
# print("contrastive_train finished.")
# print("Total time elapsed: {:.2f}s".format(time.time() - t))
#
# acc, nmi, pur, ari = valid(mnw, mv_data, batch_size)
# with open(args.db + '_result.txt', 'a+') as f:
# f.write('{} \t {} \t {} \t {} \t {} \t {:.4f} \t {:.3f} \t {:.3f} \t {:.6f} '
# '\t {:.6f} \t {:.6f} \t {:.6f} \t {:.4f} \n'.format(
# dim_idx, dim_high_feature, dim_low_feature, seed, batch_size,
# learning_rate, lmd, beta, acc, nmi, pur, ari, (time.time() - t)))
# f.flush()