【pytorch笔记】第六篇 卷积原理和卷积层

1. 卷积原理

① 卷积核不停的在原图上进行滑动，对应元素相乘再相加。

② 下图为每次滑动移动1格，然后再利用原图与卷积核上的数值进行计算得到缩略图矩阵的数据，如下图右所示。

import torch
import torch.nn.functional as F

input = torch.tensor([[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]])

kernel = torch.tensor([[1, 2, 1],
                       [0, 1, 0],
                       [2, 1, 0]])

print(input.shape)
print(kernel.shape)
input = torch.reshape(input, (1,1,5,5))
kernel = torch.reshape(kernel, (1,1,3,3))
print(input.shape)
print(kernel.shape)

output = F.conv2d(input, kernel, stride=1)
print(output)

torch.Size([5, 5])
torch.Size([3, 3])
torch.Size([1, 1, 5, 5])
torch.Size([1, 1, 3, 3])
tensor([[[[10, 12, 12],
          [18, 16, 16],
          [13,  9,  3]]]])

import torch
import torch.nn.functional as F

input = torch.tensor([[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]])

kernel = torch.tensor([[1, 2, 1],
                       [0, 1, 0],
                       [2, 1, 0]])

print(input.shape)
print(kernel.shape)
input = torch.reshape(input, (1,1,5,5))
kernel = torch.reshape(kernel, (1,1,3,3))
print(input.shape)
print(kernel.shape)

output2 = F.conv2d(input, kernel, stride=2)  # 步伐为2
print(output2)

torch.Size([5, 5])
torch.Size([3, 3])
torch.Size([1, 1, 5, 5])
torch.Size([1, 1, 3, 3])
tensor([[[[10, 12],
          [13,  3]]]])

import torch
import torch.nn.functional as F

input = torch.tensor([[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]])

kernel = torch.tensor([[1, 2, 1],
                       [0, 1, 0],
                       [2, 1, 0]])

print(input.shape)
print(kernel.shape)
input = torch.reshape(input, (1,1,5,5))
kernel = torch.reshape(kernel, (1,1,3,3))
print(input.shape)
print(kernel.shape)

output3 = F.conv2d(input, kernel, stride=1, padding=1)  # 周围只填充一层
print(output3)

torch.Size([5, 5])
torch.Size([3, 3])
torch.Size([1, 1, 5, 5])
torch.Size([1, 1, 3, 3])
tensor([[[[10, 12],
          [13,  3]]]])

2. 步幅、填充原理

① 步幅：卷积核经过输入特征图的采样间隔。设置步幅的目的：希望减小输入参数的数目，减少计算量。

② 填充：在输入特征图的每一边添加一定数目的行列。设置填充的目的：希望每个输入方块都能作为卷积窗口的中心，或使得输出的特征图的长、宽 = 输入的特征图的长、宽。

③ 一个尺寸 a * a 的特征图，经过 b * b 的卷积层，步幅（stride）= c，填充（padding）= d，若d等于0，也就是不填充，输出的特征图的尺寸 =（a-b）/ c+1；若d不等于0，也就是填充，输出的特征图的尺寸 =（a+2d-b）/ c+1。
例子1：一个特征图尺寸为4 * 4的输入，使用3 * 3的卷积核，步幅=1，填充=0，输出的尺寸=(4 - 3)/1 + 1 = 2。

例子2： 一个特征图尺寸为5 * 5的输入，使用3 * 3的卷积核，步幅=1，填充=1，输出的尺寸=(5 + 2 * 1 - 3)/1 + 1 = 5。

例子3：一个特征图尺寸为5 * 5的输入， 使用3 * 3的卷积核，步幅=2，填充=0，输出的尺寸=(5-3)/2 + 1 = 2。

例子4：一个特征图尺寸为6 * 6的输入， 使用3 * 3的卷积核，步幅=2，填充=1，输出的尺寸=(6 + 2 * 1 - 3)/2 + 1 = 2.5 + 1 = 3.5 向下取整=3（降采样：边长减少1/2）。

3. pytorch中的卷积

① Conv1d代表一维卷积，Conv2d代表二维卷积，Conv3d代表三维卷积。

② kernel_size在训练过程中不断调整，定义为3就是3 * 3的卷积核，实际我们在训练神经网络过程中其实就是对kernel_size不断调整。

③ 可以根据输入的参数获得输出的情况，如下图所示。

4. 搭建卷积层

import torch
from torch import nn
import torchvision
from torch.nn import Conv2d
from torch.utils.data import DataLoader

dataset = torchvision.datasets.CIFAR10("./dataset",train=False,transform=torchvision.transforms.ToTensor(),download=True)       
dataloader = DataLoader(dataset, batch_size=64)
class MyModule(nn.Module):
    def __init__(self):
        super(MyModule, self).__init__()
        self.conv1 = Conv2d(in_channels=3,out_channels=6,kernel_size=3,stride=1,padding=0) # 彩色图像输入为3层，我们想让它的输出为6层，选3 * 3 的卷积                
    
    def forward(self,x):
        x = self.conv1(x)
        return x
    
myModule= MyModule()
print(myModule)

Files already downloaded and verified
MyModule(
  (conv1): Conv2d(3, 6, kernel_size=(3, 3), stride=(1, 1))
)

5. 卷积层处理图片

import torch
import torchvision
from torch import nn
from torch.nn import Conv2d
from torch.utils.data import DataLoader

dataset = torchvision.datasets.CIFAR10("./dataset",train=False,transform=torchvision.transforms.ToTensor(),download=True)       
dataloader = DataLoader(dataset, batch_size=64)
class MyModule(nn.Module):
    def __init__(self):
        super(MyModule, self).__init__()
        self.conv1 = Conv2d(in_channels=3,out_channels=6,kernel_size=3,stride=1,padding=0) # 彩色图像输入为3层，我们想让它的输出为6层，选3 * 3 的卷积                
    
    def forward(self,x):
        x = self.conv1(x)
        return x
    
myModule= MyModule()
for data in dataloader:
    imgs, targets = data
    output = myModule(imgs)
    print(imgs.shape)   # 输入为3通道32×32的64张图片
    print(output.shape) # 输出为6通道30×30的64张图片

Output exceeds the size limit. Open the full output data in a text editor
Files already downloaded and verified
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([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([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])

6. Tensorboard显示

import torch
import torchvision
from torch import nn
from torch.nn import Conv2d
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

dataset = torchvision.datasets.CIFAR10("./dataset",train=False,transform=torchvision.transforms.ToTensor(),download=True)       
dataloader = DataLoader(dataset, batch_size=64)
class MyModule(nn.Module):
    def __init__(self):
        super(MyModule, self).__init__()
        self.conv1 = Conv2d(in_channels=3,out_channels=6,kernel_size=3,stride=1,padding=0) # 彩色图像输入为3层，我们想让它的输出为6层，选3 * 3 的卷积                
    
    def forward(self,x):
        x = self.conv1(x)
        return x
    
myModule= MyModule()
writer = SummaryWriter("logs")
step = 0
for data in dataloader:
    imgs, targets = data
    output = myModule(imgs)
    print(imgs.shape)   
    print(output.shape)    
    writer.add_images("input", imgs, step)
    output = torch.reshape(output,(-1,3,30,30)) # 把原来6个通道拉为3个通道，为了保证所有维度总数不变，其余的分量分到第一个维度中
    writer.add_images("output", output, step)
    step = step + 1

Output exceeds the size limit. Open the full output data in a text editor
Files already downloaded and verified
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([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([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([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([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])

① 在 Anaconda 终端里面，激活py3.6.3环境，再输入 tensorboard --logdir=C:\Users\qj\Desktop\CV\logs 命令，将网址赋值浏览器的网址栏，回车，即可查看tensorboard显示日志情况。