叠加卷积层似乎很有用！ # 从0实现 

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
import torch
from torch import nn

def corr2d(X, K):
    # 位相关运算
    h, w = X.shape[-2], X.shape[-1]
    kh, kw = K.shape[0], K.shape[1]
    Y = torch.zeros(h-kh+1, w-kw+1)
    for i in range(Y.shape[0]):
        for j in range(Y.shape[1]):
            Y[i, j] = (X[i:i+kh, j:j+kh]*K).sum()
    
    return Y

class Conv2D(nn.Module):
    def __init__(self, kernel_size) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.rand(kernel_size))
        self.bias = nn.Parameter(torch.zeros(1))

    def forward(self, x):
        return corr2d(x, self.weight) + self.bias

# 下面y相当于是边缘检查核，是早期的图像特征提取的方法
# 但是注意卷积神经网络的核都是训练出来的
x = torch.cat([torch.ones((6, 3)), torch.zeros((6, 3))], axis = 1)
k = torch.tensor([[-1,1], [-1,1]])

y = corr2d(x, k)


# train 得到上面的k
net = Conv2D((2,2))
lr = 0.02

for i in range(10):
    y_hat = net(x)
    l = (y-y_hat)**2
    net.zero_grad()
    l.sum().backward()
    net.weight.data[:] -= lr*net.weight.grad

    print('epoch:{}, loss:{}'.format(i, l.sum()))

train FashionMNIST

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
import torch
from torch import nn
import torchvision
from torchvision import transforms

# 定义net来train FashionMNIST
mnist_train = torchvision.datasets.FashionMNIST(
    root='../data', train=True, transform=transforms.ToTensor(), download=True
)
mnist_test = torchvision.datasets.FashionMNIST(
    root='../data', train=False, transform=transforms.ToTensor(), download=True
)

class CnnNet(nn.Module):
    def __init__(self, input_channels=1) -> None:
        super().__init__()
        self.is_training = True
        self.drop_layer = nn.Dropout(0.5)
        self.conn2d1 = nn.Sequential(
            nn.Conv2d(1, 16, kernel_size=(3,3), stride=1, padding=1),
            nn.MaxPool2d(kernel_size=(2,2), stride=2),
            nn.BatchNorm2d(16),
            nn.ReLU()
        )
        self.conn2d2 = nn.Sequential(
            nn.Conv2d(16, 32, kernel_size=(5,5), stride=1, padding=1),
            nn.MaxPool2d(kernel_size=(2,2), stride=2),
            nn.BatchNorm2d(32),
            nn.ReLU()
        )
        self.conn2d3 = nn.Sequential(
            nn.Conv2d(32, 64, kernel_size=(5,5), stride=1, padding=1),
            nn.MaxPool2d(kernel_size=(2,2), stride=2),
            nn.BatchNorm2d(64),
            nn.ReLU()
        )
        self.conn2d4 = nn.Sequential(
            nn.Conv2d(64, 128, kernel_size=(2,2), stride=1, padding=1),
            # nn.MaxPool2d(kernel_size=(2,2), stride=2),
            nn.BatchNorm2d(128),
            nn.ReLU()
        )
        # self.conv1 = nn.Conv2d(1, 6, kernel_size=(3,3), stride=1, padding=1)
        # self.maxpool = nn.MaxPool2d(kernel_size=(2,2), stride=2)
        # self.conv2 = nn.Conv2d(6, 16, kernel_size=(3,3), stride=1, padding=1)
        self.flatten = nn.Flatten()
        self.relu = nn.ReLU()
        # self.line1 = nn.Sequential(
        #     nn.Linear(784, 120),
        #     nn.ReLU(),
        #     nn.Linear(120, 84),
        #     nn.ReLU(),
        #     nn.Linear(84, 10)
        # )
        self.line1 = nn.Linear(1152, 256)
        self.line2 = nn.Linear(256, 10)
        # self.line1 = nn.Linear(1152, 240)
        # self.line2 = nn.Linear(240, 84)
        # self.line3 = nn.Linear(84, 10)

    
    def forward(self, x):
        # c1 = self.conv1(x)
        # c1 = self.relu(self.maxpool(c1))
        # c2 = self.conv2(c1)
        # c2 = self.relu(self.maxpool(c2))
        c1 = self.conn2d1(x)
        c2 = self.conn2d2(c1)
        c3 = self.conn2d3(c2)
        c4 = self.conn2d4(c3)
        o1 = self.flatten(c4)
        # return o1
        # l = self.line(o1)
        l1 = self.relu(self.line1(o1))
        if self.is_training:
            l1 = self.drop_layer(l1)
        
        # l2 = self.relu(self.line2(l1))
        # if self.is_training:
        #     l2 = self.drop_layer(l2)
        # l3 = self.line3(l2)
        o = self.line2(l1)
        return o

    def predict(self, x):
        self.is_training = False
        o = self.forward(x)
        self.is_training = True
        return o

from torch.utils import data

def ac(data_iter, net, device):
    num_acs = []
    for x, y in data_iter:
        x = x.to(device)
        y = y.to(device)
        y_hat = net.predict(x)
        maxs, indexs = torch.max(y_hat, dim=1)
        num_acs.append(y.eq(indexs).sum()/indexs.shape[0])
    return sum(num_acs)/len(num_acs)

# 参数
batch_size = 256
num_epochs = 20
lr = 0.1

train_iter = data.DataLoader(mnist_train, batch_size, shuffle=True, num_workers=4, pin_memory=True)
test_iter = data.DataLoader(mnist_test, batch_size,shuffle=True, num_workers=4)

net = CnnNet()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
net.to(device)
loss = nn.CrossEntropyLoss()
trainer = torch.optim.SGD(net.parameters(), lr=lr)

# 训练
for i in range(num_epochs):
    for x,y in train_iter:
        x = x.to(device)
        y = y.to(device)
        y_hat = net(x)
        l = loss(y_hat, y)
        trainer.zero_grad()
        l.backward()
        trainer.step()
    print(l)
    print(ac(test_iter, net, device))


from torchinfo import summary

net = CnnNet()
summary(net, input_size=(1,1,28,28))

# 输出net大小
==========================================================================================
Layer (type:depth-idx)                   Output Shape              Param #
==========================================================================================
CnnNet                                   [1, 10]                   --
├─Sequential: 1-1                        [1, 16, 14, 14]           --
│    └─Conv2d: 2-1                       [1, 16, 28, 28]           160
│    └─MaxPool2d: 2-2                    [1, 16, 14, 14]           --
│    └─BatchNorm2d: 2-3                  [1, 16, 14, 14]           32
│    └─ReLU: 2-4                         [1, 16, 14, 14]           --
├─Sequential: 1-2                        [1, 32, 6, 6]             --
│    └─Conv2d: 2-5                       [1, 32, 12, 12]           12,832
│    └─MaxPool2d: 2-6                    [1, 32, 6, 6]             --
│    └─BatchNorm2d: 2-7                  [1, 32, 6, 6]             64
│    └─ReLU: 2-8                         [1, 32, 6, 6]             --
├─Sequential: 1-3                        [1, 64, 2, 2]             --
│    └─Conv2d: 2-9                       [1, 64, 4, 4]             51,264
│    └─MaxPool2d: 2-10                   [1, 64, 2, 2]             --
│    └─BatchNorm2d: 2-11                 [1, 64, 2, 2]             128
│    └─ReLU: 2-12                        [1, 64, 2, 2]             --
├─Sequential: 1-4                        [1, 128, 3, 3]            --
│    └─Conv2d: 2-13                      [1, 128, 3, 3]            32,896
│    └─BatchNorm2d: 2-14                 [1, 128, 3, 3]            256
│    └─ReLU: 2-15                        [1, 128, 3, 3]            --
├─Flatten: 1-5                           [1, 1152]                 --
├─Linear: 1-6                            [1, 256]                  295,168
├─ReLU: 1-7                              [1, 256]                  --
├─Dropout: 1-8                           [1, 256]                  --
├─Linear: 1-9                            [1, 10]                   2,570
==========================================================================================
Total params: 395,370
Trainable params: 395,370
Non-trainable params: 0
Total mult-adds (M): 3.39
==========================================================================================
Input size (MB): 0.00
Forward/backward pass size (MB): 0.20
Params size (MB): 1.58
Estimated Total Size (MB): 1.79
==========================================================================================