pytorch学习记录02——多分支网络

本文使用pytorch框架来搭建一个多分支的神经网络，编程时借鉴了Inception的编程思想。
准备实现的网络结构如下图所示

（图画的有点丑）

每个分支的输入图片大小设置为64*64，卷积层和池化层的参数设置如下表所示

层	参数
CONV1	3，16，kernel_size=3, stride=1, padding=1
Pooling1	kernel_size=2, stride=2
CONV2	16，32，kernel_size=3, stride=1, padding=1
Pooling2	kernel_size=2, stride=2
CONV3	32，64，kernel_size=3, stride=1, padding=1
Pooling3	kernel_size=2, stride=2
CONV4	64，128，kernel_size=3, stride=1, padding=1
Pooling4	kernel_size=2, stride=2

首先，导入需要的库

# 导入库
import torch
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
import torchvision.transforms as transforms
import torchvision.datasets as datasets
from torch import nn, optim

接下来，定义网络模型

# 定义模型结构
class ThreeInputsNet(nn.Module):
    def __init__(self):
        super(ThreeInputsNet, self).__init__()
        # 3, 64, 64
        self.conv1_1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)
        self.pooling1_1 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv1_2 = nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1)
        self.conv1_3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)
        self.conv1_4 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
        # 3, 64, 64
        self.conv2_1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)
        self.pooling2_1 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv2_2 = nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1)
        self.conv2_3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)
        self.conv2_4 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
        # 3, 64, 64
        self.conv3_1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)
        self.pooling3_1 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv3_2 = nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1)
        self.conv3_3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)
        self.conv3_4 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
        # 128, 4, 4
        # 三个通道的channel合并
        # 128*5, 4, 4
        self.outlayer1 = nn.Linear(3 * 128 * 4 * 4, 128 * 3)
        self.outlayer2 = nn.Linear(128 * 3, 256)
        self.outlayer3 = nn.Linear(256, 3)
	# 此处的输入为三个，对应三个分支
    def forward(self, input1, input2, input3):
        out1 = self.pooling1_1(self.conv1_1(input1))
        out1 = self.pooling1_1(self.conv1_2(out1))
        out1 = self.pooling1_1(self.conv1_3(out1))
        out1 = self.pooling1_1(self.conv1_4(out1))

        out2 = self.pooling2_1(self.conv2_1(input2))
        out2 = self.pooling2_1(self.conv2_2(out2))
        out2 = self.pooling2_1(self.conv2_3(out2))
        out2 = self.pooling2_1(self.conv2_4(out2))

        out3 = self.pooling3_1(self.conv3_1(input3))
        out3 = self.pooling3_1(self.conv3_2(out3))
        out3 = self.pooling3_1(self.conv3_3(out3))
        out3 = self.pooling3_1(self.conv3_4(out3))
		# 将三个分支的结果在channel维度上合并
        out = torch.cat((out1, out2, out3), dim=1)
        out = out.view(out.size(0), -1)  # [B, C, H, W] --> [B, C*H*W]
        out = self.outlayer1(out)
        out = self.outlayer2(out)
        out = self.outlayer3(out)
        return out

输入一些数据测试一下网络能否跑通

if __name__ == '__main__':
    input1 = torch.ones(8, 3, 64, 64)
    input2 = torch.ones(8, 3, 64, 64)
    input3 = torch.ones(8, 3, 64, 64)
    net = ThreeInputsNet()
    output = net(input1, input2, input3)
    print("out.shape:{}".format(output.shape))

完整代码

# 导入库
import torch
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
import torchvision.transforms as transforms
import torchvision.datasets as datasets
from torch import nn, optim


# 定义模型结构
class ThreeInputsNet(nn.Module):
    def __init__(self):
        super(ThreeInputsNet, self).__init__()
       # 3, 64, 64
        self.conv1_1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)
        self.pooling1_1 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv1_2 = nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1)
        self.conv1_3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)
        self.conv1_4 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
        # 3, 64, 64
        self.conv2_1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)
        self.pooling2_1 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv2_2 = nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1)
        self.conv2_3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)
        self.conv2_4 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
        # 3, 64, 64
        self.conv3_1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)
        self.pooling3_1 = nn.MaxPool2d(kernel_size=2, stride=2)
        self.conv3_2 = nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1)
        self.conv3_3 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)
        self.conv3_4 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
        # 128, 4, 4
        # 三个个通道的channel合并
        # 128*5, 4, 4
        self.outlayer1 = nn.Linear(3 * 128 * 4 * 4, 128 * 5)
        self.outlayer2 = nn.Linear(128 * 5, 256)
        self.outlayer3 = nn.Linear(256, 3)

    def forward(self, input1, input2, input3):
        out1 = self.pooling1_1(self.conv1_1(input1))
        out1 = self.pooling1_1(self.conv1_2(out1))
        out1 = self.pooling1_1(self.conv1_3(out1))
        out1 = self.pooling1_1(self.conv1_4(out1))

        out2 = self.pooling2_1(self.conv2_1(input2))
        out2 = self.pooling2_1(self.conv2_2(out2))
        out2 = self.pooling2_1(self.conv2_3(out2))
        out2 = self.pooling2_1(self.conv2_4(out2))

        out3 = self.pooling3_1(self.conv3_1(input3))
        out3 = self.pooling3_1(self.conv3_2(out3))
        out3 = self.pooling3_1(self.conv3_3(out3))
        out3 = self.pooling3_1(self.conv3_4(out3))
        # 将三个分支的结果在channel维度上合并
        out = torch.cat((out1, out2, out3), dim=1)
        out = out.view(out.size(0), -1)  # [B, C, H, W] --> [B, C*H*W]
        out = self.outlayer1(out)
        out = self.outlayer2(out)
        out = self.outlayer3(out)
        return out


if __name__ == '__main__':
    input1 = torch.ones(8,3,64,64)
    input2 = torch.ones(8, 3, 64, 64)
    input3 = torch.ones(8, 3, 64, 64)
    net = ThreeInputsNet()
    output = net(input1, input2, input3)
    print("out.shape:{}".format(output.shape))