-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
170 lines (150 loc) · 7.08 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
"""
This demo is for "MULTI VIEW INFORMATION BOTTLENECK WITHOUT VARIATIONAL APPROXIMATION" in ICASSP 2022.
"""
import numpy as np
import scipy.io as sio
import torch
from model import MLP
import torch.backends.cudnn as cudnn
import torch.nn as nn
import argparse
from scipy.spatial.distance import pdist, squareform
from utils import calculate_MI
import os
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
parser = argparse.ArgumentParser(description='MEIB')
parser.add_argument('--lr', type=float, default=1e-4, help='learning rate')
parser.add_argument('--weight_decay', '-wd', type=float, default=0.03, help='weight_decay. Default:5e-4')
parser.add_argument('--epochs', type=int, default=60, help='Number of training epochs. Default: 10')
parser.add_argument('--batchsize', type=int, default=64, help='batch_size')
parser.add_argument('--ks', type=int, default=10, help='kernel size')
args = parser.parse_args()
# ---------------------------------------------- Load data function ---------------------------------------------------
def load_data(validation, iterr):
data = sio.loadmat('synthetic_data/iter' + str(iterr) + '.mat')
Xtrain1 = data['X1_train']
Xtrain2 = data['X2_train']
ytrain = data['ytrain']
if validation:
Xtest1 = data['X1_val']
Xtest2 = data['X2_val']
ytest = data['yval']
else:
Xtest1 = data['X1_test']
Xtest2 = data['X2_test']
ytest = data['ytest']
ytrain = np.squeeze(ytrain)
ytest = np.squeeze(ytest)
num_examples, _ = Xtrain1.shape
idx = np.random.permutation(num_examples)
Xtrain1_shuffle = Xtrain1[idx, :]
Xtrain2_shuffle = Xtrain2[idx, :]
ytrain_shuffle = ytrain[idx]
Xtrain1, Xtrain2, ytrain = torch.from_numpy(Xtrain1_shuffle), torch.from_numpy(Xtrain2_shuffle), torch.from_numpy(
ytrain_shuffle)
Xtest1, Xtest2, ytest = torch.from_numpy(Xtest1), torch.from_numpy(Xtest2), torch.from_numpy(ytest)
return Xtrain1, Xtrain2, ytrain, Xtest1, Xtest2, ytest
# ---------------------------------------------- Model -----------------------------------------------------------------
print('==> Building model..')
net = MLP(10)
net = net.to(device)
if device == 'cuda':
net = torch.nn.DataParallel(net)
cudnn.benchmark = True
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=args.lr, weight_decay=args.weight_decay, amsgrad=True)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.01)
# ---------------------------------------------- Train and test---------------------------------------------------------
def train(epochs, iterr):
for epoch in range(epochs):
train_loss = 0
IX1Z1_loss = 0
IX2Z2_loss = 0
correct = 0
total = 0
net.train()
inputs1, inputs2, targets, _, _, _ = load_data(True, iterr) # train data
num_examples, d1 = inputs1.shape
inputs1, inputs2, targets = inputs1.float().to(device), inputs2.float().to(device), targets.long().to(device)
batch_size = args.batchsize
batch_num = int(num_examples / batch_size)
for step in range(batch_num):
x1 = inputs1[step * batch_size: (step + 1) * batch_size, :]
x2 = inputs2[step * batch_size: (step + 1) * batch_size, :]
y = targets[step * batch_size: (step + 1) * batch_size]
x1, x2, y = x1.to(device), x2.to(device), y.to(device)
optimizer.zero_grad()
Z1, Z2, outputs = net(x1, x2)
loss = criterion(outputs, y)
with torch.no_grad():
x1_numpy = x1.cpu().detach().numpy()
k_x1 = squareform(pdist(x1_numpy, 'euclidean'))
sigma_x1 = np.mean(np.mean(np.sort(k_x1[:, :args.ks], 1)))
x2_numpy = x2.cpu().detach().numpy()
k_x2 = squareform(pdist(x2_numpy, 'euclidean'))
sigma_x2 = np.mean(np.mean(np.sort(k_x2[:, :args.ks], 1)))
Z1_numpy = Z1.cpu().detach().numpy()
Z2_numpy = Z2.cpu().detach().numpy()
k_z1 = squareform(pdist(Z1_numpy, 'euclidean'))
k_z2 = squareform(pdist(Z2_numpy, 'euclidean'))
sigma_z1 = np.mean(np.mean(np.sort(k_z1[:, :args.ks], 1)))
sigma_z2 = np.mean(np.mean(np.sort(k_z2[:, :args.ks], 1)))
IX1Z1 = calculate_MI(x1, Z1, s_x=sigma_x1 ** 2, s_y=sigma_z1 ** 2)
IX2Z2 = calculate_MI(x2, Z2, s_x=sigma_x2 ** 2, s_y=sigma_z2 ** 2)
beta1 = 0.0001
beta2 = 0.0001
total_loss = loss + beta1 * IX1Z1 + beta2 * IX2Z2
total_loss.backward()
optimizer.step()
train_loss += loss.item()
IX1Z1_loss += IX1Z1.item()
IX2Z2_loss += IX2Z2.item()
_, predicted = outputs.max(1)
total += y.size(0)
correct += predicted.eq(y).sum().item()
acc = correct / total
print("Training: epoch {}: err={:.4f}".format(epoch, 1 - acc))
net.eval()
validate_loss = 0
test_loss = 0
v_correct = 0
t_correct = 0
v_total = 0
t_total = 0
with torch.no_grad():
_, _, _, v_inputs1, v_input2, v_targets = load_data(True, iterr) # validation data
v_inputs1, v_input2, v_targets = v_inputs1.float().to(device), v_input2.float().to(device), \
v_targets.long().to(device)
_, _, v_outputs = net(v_inputs1, v_input2)
v_loss = criterion(v_outputs, v_targets)
validate_loss += v_loss.item()
_, v_predicted = v_outputs.max(1)
v_total += v_targets.size(0)
v_correct += v_predicted.eq(v_targets).sum().item()
v_acc = v_correct / v_total
_, _, _, t_inputs1, t_input2, t_targets = load_data(False, iterr) # test data
t_inputs1, t_input2, t_targets = t_inputs1.float().to(device), t_input2.float().to(device), \
t_targets.long().to(device)
_, _, t_outputs = net(t_inputs1, t_input2)
t_loss = criterion(t_outputs, t_targets)
test_loss += t_loss.item()
_, t_predicted = t_outputs.max(1)
t_total += t_targets.size(0)
t_correct += t_predicted.eq(t_targets).sum().item()
t_acc = t_correct / t_total
print("Validation: epoch {}: acc={:.4f}\terr={:.4f}".format(epoch, v_acc, 1 - v_acc))
print('Testing: epoch{}: acc={:.4f}\terr={:.4f}'.format(epoch, t_acc, 1 - t_acc))
return t_acc
# ---------------------------------------------- main ------------------------------------------------------------------
if __name__ == "__main__":
t_err_list = []
result_dir = 'multiview'
if not os.path.exists(result_dir):
os.makedirs(result_dir)
result_file = open(result_dir + '/result.txt', 'w')
for iterr in range(1, 6):
t_err = train(args.epochs, iterr)
t_err_list.append(1-t_err)
result_file.write("t_err_list: %s\n" % (t_err_list))
result_file.close()
print('t_err_list:\n', t_err_list)