【Pytorch】神经网络卷积层学习笔记

视频地址
打开torch官方文档的卷积层页面

最常用的是nn.conv2d&＃xff0c;点击

CLASStorch.nn.Conv2d(in_channels, out_channels, kernel_size, stride&＃61;1, padding&＃61;0, dilation&＃61;1, groups&＃61;1, bias&＃61;True, padding_mode&＃61;&＃39;zeros&＃39;, device&＃61;None, dtype&＃61;None)


一些不太清楚的

• groups&＃xff1a;分组卷积&＃xff0c;几乎用不到
• bias&＃xff1a;通常设置为True

写代码初始化一下model

import torch
import torchvision
from torch import nn
from torch.nn import Conv2d
from torch.utils.data import DataLoaderdataset &＃61; torchvision.datasets.CIFAR10("./dataset", train&＃61;False, transform&＃61;torchvision.transforms.ToTensor())dataloader &＃61; DataLoader(dataset, batch_size&＃61;64)class Model(nn.Module):def __init__(self):super(Model, self).__init__()self.conv1 &＃61; Conv2d(in_channels&＃61;3, out_channels&＃61;6, kernel_size&＃61;3, stride&＃61;1, padding&＃61;0) # 把self后面的变量在其它函数里也能够用到def forward(self, x):x &＃61; self.conv1(x)return xmodel &＃61; Model()
print(model)

输出结果为

D:\Anaconda3\envs\pytorch\python.exe D:/研究生/代码尝试/nn_conv2d.py
Model((conv1): Conv2d(3, 6, kernel_size&＃61;(3, 3), stride&＃61;(1, 1))
)进程已结束&＃xff0c;退出代码为 0


再把dataloder装进去&＃xff0c;即图片输入

for data in dataloader:imgs, targets &＃61; dataoutput &＃61; model(imgs)print(imgs.shape)print(output.shape)


输出结果为&＃xff08;前两行&＃xff09;

torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])


可以看出&＃xff0c;这里的batch_size是64&＃xff0c;输入channel有3层&＃xff0c;卷积后有6个channel&＃xff0c;卷积核不需要初始化&＃xff0c;是自动初始化的

可视化&＃xff0c;用Tensorboard显示一下&＃xff0c;顺便把代码贴全

import torch
import torchvision
from torch import nn
from torch.nn import Conv2d
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriterdataset &＃61; torchvision.datasets.CIFAR10("./dataset", train&＃61;False, transform&＃61;torchvision.transforms.ToTensor())dataloader &＃61; DataLoader(dataset, batch_size&＃61;64)class Model(nn.Module):def __init__(self):super(Model, self).__init__()self.conv1 &＃61; Conv2d(in_channels&＃61;3, out_channels&＃61;6, kernel_size&＃61;3, stride&＃61;1, padding&＃61;0) # 把self后面的变量在其它函数里也能够用到def forward(self, x):x &＃61; self.conv1(x)return xmodel &＃61; Model()writer &＃61; SummaryWriter("./logs")
step &＃61; 0
for data in dataloader:imgs, targets &＃61; dataoutput &＃61; model(imgs)print(imgs.shape)print(output.shape)# 输入大小 torch.Size([64, 3, 32, 32])writer.add_images("input", imgs, step)# 输出大小 torch.Size([64, 6, 30, 30]) 但是六个channel没法显示&＃xff0c;所以要转换一下# ([64, 6, 30, 30]) ->([xxx, 3, 30, 30])output &＃61; torch.reshape(output, (-1, 3, 30, 30))writer.add_images("output", output, step)step &＃61; step &＃43; 1

还是老样子

tensorboard --logdir&＃61;logs


点开后。。。麦艾斯。。。


因为通道从6个压缩成了3个&＃xff0c;所以这里有128张图