学习pytorch9 神经网络-卷积层

官网

https://pytorch.org/docs/stable/nn.html#convolution-layers
图像识别常用conv2d 二维卷积 nn.Conv2d
https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html#torch.nn.Conv2d

卷积数据公式

参数说明

Parameters:

• in_channels (int) – Number of channels in the input image 输入通道

• out_channels (int) – Number of channels produced by the convolution 输出通道

• kernel_size (int or tuple) – Size of the convolving kernel 卷积核大小

• stride (int or tuple, optional) – Stride of the convolution. Default: 1 每次卷积走多少步，横向纵向的步径大小

• padding (int, tuple or str, optional) – Padding added to all four sides of the input. Default: 0 是否在卷积过程中对输入图像的边缘进行填充

• padding_mode (str, optional) – ‘zeros’, ‘reflect’, ‘replicate’ or ‘circular’. Default: ‘zeros’ 填充数据的模式是什么，默认为zeros，填充的都是0

• dilation (int or tuple, optional) – Spacing between kernel elements. Default: 1 卷积核中间的距离? 一般不改 不常用 称为空洞卷积 https://zhuanlan.zhihu.com/p/453558537

• groups (int, optional) – Number of blocked connections from input channels to output channels. Default: 1 一般不改 不常用

• bias (bool, optional) – If True, adds a learnable bias to the output. Default: True 添加偏置值，默认为True添加偏置值
  

卷积运算演示

https://github.com/vdumoulin/conv_arithmetic/blob/master/README.md

卷积

绿色方格：表示输出图像
蓝色方格：表示输入图像
蓝色方格中的深色阴影部分：表示kernel 卷积核
白色虚线：表示padding填充
动画中深色阴影上下左右整体移动的方格数，表示stride的大小

空洞卷积

https://zhuanlan.zhihu.com/p/453558537

输入输出channel

两个卷积核做两次卷积，叠加输出一起是out_channel=2

代码

注意点：

1. super()括号里面没有内容，自动填充的self应该去掉

super().__init__() 
1

code

import torch
import torchvision
from torch import nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

test_set = torchvision.datasets.CIFAR10(root='./dataset', train=False, transform=torchvision.transforms.ToTensor(), download=True)

dataloader = DataLoader(test_set, batch_size=64, shuffle=False)

class NnConv2d(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=1, padding=0)

    def forward(self, x):
        x = self.conv1(x)
        return x

nnconvd = NnConv2d()
writer = SummaryWriter('./logs')
step = 0
for data in dataloader:
    imgs, targets = data
    output = nnconvd(imgs)
    print(imgs.shape)
    print(output.shape)
    writer.add_images("input", imgs)
    output = output.reshape([-1, 3, 30, 30])
    writer.add_images("output", output)
    step += 1

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

执行结果

......
torch.Size([64, 6, 30, 30])
torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])
torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])
torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])
torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])
torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])
torch.Size([16, 3, 32, 32])
torch.Size([16, 6, 30, 30])  # 最后一个batch16是因为drop_last默认为False，最后没除尽的也要保留，参与训练
1
2
3
4
5
6
7
8
9
10
11
12
13
14