ConvNN/withPoolDim.py at master · MDProject/ConvNN · GitHub

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
import torch
import torch.nn as nn
from torch.autograd import Variable
import torchvision.datasets as dset
import torchvision.transforms as transforms
import torch.nn.functional as F
import torch.optim as optim
import os.path as IO
import numpy as np

conv1 = []
conv2 = []
conv3 = []
FC = []
class Net(nn.Module):

    def __init__(self):
        super(Net, self).__init__()
        # 1 input image channel, 6 output channels, 5x5 square convolution
        # kernel
        self.conv1 = nn.Conv2d(1, 5, 5) # 28*28 -- 24*24 + pooling 24*24 -- 12*12
        self.conv2 = nn.Conv2d(5, 10, 5) # 12*12 -- 8*8
        self.conv3 = nn.Conv2d(10, 20, 5) # 8*8 -- 4*4
        # an affine operation: y = Wx + b
        self.fc1 = nn.Linear(20 * 4 * 4, 64)
        self.fc2 = nn.Linear(64, 10)

    def forward(self, x):
        tmp = F.relu(self.conv1(x))
        x = F.avg_pool2d(tmp,(2,2))
        conv1.append(tmp)
        x = F.relu(self.conv2(x))
        conv2.append(x)
        x = F.relu(self.conv3(x))
        conv3.append(x)
        x = x.view(-1, self.num_flat_features(x))
        x = F.relu(self.fc1(x))
        x = self.fc2(x)
        return x

    def num_flat_features(self, x):
        size = x.size()[1:]  # all dimensions except the batch dimension
        num_features = 1
        for s in size:
            num_features *= s
        return num_features

net = Net()
for layer in list(net._modules.items()):
    print(layer[1])

"""
def for_hook(module, input, output):
    print(module)
    for val in input:
        print("input val:",val.size())
    for out_val in output:
        print("output val:", out_val.size())
"""
# net.conv2.register_forward_hook(for_hook)
# net.conv3.register_forward_hook(for_hook)
batchSize = 50
trans = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (1.0,))])
# print(IO.exists('../MNIST_DATA/train-labels-idx1-ubyte'))
test_set = dset.MNIST('../MNIST_DATA/', train=False, transform=trans, download=True)
# img, label = train_set[0]

test_loader = torch.utils.data.DataLoader(dataset=test_set,
                                          batch_size=batchSize,
                                          shuffle=False)

criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(),lr=1e-3)
loadPath = './Project/Pool/param'
appendix = '.t7'
#   print(IO.exists(savePath))
loss_bound = 0.01
Ensemble_num = 10
n = 0   # w index
net = Net()
path = loadPath + str(n) + appendix
model_dict=net.load_state_dict(torch.load(path))
for images, labels in test_loader:
    outputs = net(images)
print(len(conv1))
print(len(conv2)) # 200
h_act1 = torch.zeros(len(conv1)*Ensemble_num, 24*24)
h_act2 = torch.zeros(len(conv2)*Ensemble_num, 8*8)
h_act3 = torch.zeros(len(conv3)*Ensemble_num, 4*4)

D1 = np.zeros(conv1[0][0].size(0))
for cindex in range(conv1[0][0].size(0)):
    for i in range(len(conv1)):
        for j in range(Ensemble_num): # channel idx = 1
            h_act1[i*Ensemble_num+j,:] = conv1[i][j,cindex].reshape(1,24*24)
    h_act1_T = torch.transpose(h_act1,0,1)
    CMat1 = torch.mm(h_act1_T,h_act1)/h_act1.size()[0]
    h_act1_ = torch.mean(h_act1,0)
    h_act1_ = h_act1_.view(1,h_act1_.size(0))
    CMat1_ = torch.mm(torch.transpose(h_act1_,0,1),h_act1_)
    CovMatrix1 = CMat1 - CMat1_
    D1[cindex] = np.square(torch.trace(CovMatrix1).detach().numpy())/torch.trace(torch.mm(CovMatrix1,CovMatrix1)).detach().numpy()
print(D1)
"""
"""
D2 = np.zeros(conv2[0][0].size(0))
for cindex in range(conv2[0][0].size(0)):
    for i in range(len(conv2)):
        for j in range(Ensemble_num): # channel idx = 1
            h_act2[i*Ensemble_num+j,:] = conv2[i][j,cindex].reshape(1,8*8)
    h_act2_T = torch.transpose(h_act2,0,1)
    CMat2 = torch.mm(h_act2_T,h_act2)/h_act2.size()[0]
    h_act2_ = torch.mean(h_act2,0)
    h_act2_ = h_act2_.view(1,h_act2_.size(0))
    CMat2_ = torch.mm(torch.transpose(h_act2_,0,1),h_act2_)
    CovMatrix2 = CMat2 - CMat2_
    D2[cindex] = np.square(torch.trace(CovMatrix2).detach().numpy())/torch.trace(torch.mm(CovMatrix2,CovMatrix2)).detach().numpy()
print(D2)
"""
"""
D3 = np.zeros(conv3[0][0].size(0))
for cindex in range(conv3[0][0].size(0)):
    for i in range(len(conv3)):
        for j in range(Ensemble_num): # channel idx = 1
            h_act3[i*Ensemble_num+j,:] = conv3[i][j,cindex].reshape(1,4*4)
    h_act3_T = torch.transpose(h_act3,0,1)
    CMat3 = torch.mm(h_act3_T,h_act3)/h_act3.size()[0]
    h_act3_ = torch.mean(h_act3,0)
    h_act3_ = h_act3_.view(1,h_act3_.size(0))
    CMat3_ = torch.mm(torch.transpose(h_act3_,0,1),h_act3_)
    CovMatrix3 = CMat3 - CMat3_
    D3[cindex] = np.square(torch.trace(CovMatrix3).detach().numpy())/torch.trace(torch.mm(CovMatrix3,CovMatrix3)).detach().numpy()
print(D3)