Pytorch convolution operation

pytorch Convolution operation official document
Here we use nn.conv2d To explain the convolution operation .

What is convolution ？

The convolution kernel moves on the input image , Then multiply and sum the values on the convolution kernel and the corresponding position on the input image .Stride=1 To control the moving step of convolution kernel .

Convolution operation example code ：

import torch.nn.functional as F
import torch

#  The input image (5X5)
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]])

#  Convolution kernel (3X3)
kernel = torch.tensor([[1,2,1],
                       [0,1,0],
                       [2,1,0]])


# input: torch.Size([5, 5])
print("input:\n",input.shape)
# kernel:torch.Size([3, 3])
print("kernel:\n",kernel.shape)

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


# input:torch.Size([1, 1, 5, 5])
print("input:\n",input.shape)
# kernel:torch.Size([1, 1, 3, 3])
print("kernel:\n",kernel.shape)

#  Convolution operation   Observe stride Influence on convolution results 
output = F.conv2d(input,kernel,stride=1)
print('output\n',output)


output2 = F.conv2d(input,kernel,stride=2)
print('output2\n',output2)

#  Perform volume and operation   Expand and fill the boundary of the input image   Observe padding Influence on convolution results 
output3 = F.conv2d(input,kernel,stride=1,padding=1)
print("output\n",output3)

Part of the code explanation ：

1.reshape The role of

# reshape front 
# input: torch.Size([5, 5]) kernel:torch.Size([3, 3])
input = torch.reshape(input,(1,1,5,5))
kernel = torch.reshape(kernel,(1,1,3,3))
# reshape after 
# input:torch.Size([1, 1, 5, 5]) kernel:torch.Size([1, 1, 3, 3])

Why do I need to be right input and kernel Conduct reshape This operation ？
Because use torch.nn.functional.conv2d The input parameters are limited , You can see conv2d Requirements for input parameters , requirement input The input is (minibatch,in_channels,iH,iW), among in_channels Indicates the number of channels ,iH Indicates the height of the input image ,iW Indicates the width of the input image .weigt The input is kernel( Convolution kernel ), You can see that it's right weight The parameter requirements of are similar to input, among outchannels Indicates the number of output channels ,in_channels Indicates the number of input channels (groups Default equal to 1),kH Represents the height of the convolution kernel ,kW Represents the width of the convolution kernel . So you need to input and kernel Conduct reshape operation .

2.stride Parameters

#  Convolution operation   Observe stride Influence on convolution results 
output = F.conv2d(input,kernel,stride=1)
print('output\n',output)

output2 = F.conv2d(input,kernel,stride=2)
print('output2\n',output2)

Running results ：

You can see Official documents Yes Stride The explanation of ：

• stride – the stride of the convolving kernel. Can be a single number or a tuple (sH, sW). Default: 1
  When stride What you enter is a number , Then this number is the horizontal and vertical moving steps of the convolution kernel , When stride When you enter a tuple , The steps of the convolution kernel moving horizontally and vertically can be set respectively .
  
  3.padding Parameters

#  Perform volume and operation   Expand and fill the boundary of the input image   Observe padding Influence on convolution results 
output3 = F.conv2d(input,kernel,stride=1,padding=1)
print("output\n",output3)

In the above code padding The function of parameters is equivalent to , Expand the horizontal and vertical boundaries of the input image 1 Length and fill 0, Then perform convolution operation .

You can see Official documents Yes Padding The explanation of ：

• padding – implicit paddings on both sides of the input. Can be a string {‘valid’, ‘same’}, single number or a tuple (padH, padW). Default: 0 padding=‘valid’ is the same as no padding. padding=‘same’ pads the input so the output has the same shape as the input. However, this mode doesn’t support any stride values other than 1.
  When padding What you enter is a number , Then this number is the horizontal and vertical boundary expansion filling of the image ( The default filling value is 0) The length of , When padding When you enter a tuple , You can set the length of the horizontal and vertical boundaries of the image respectively .