CNN - nn. Conv1d use

---

One 、Conv1d Parameter setting

torch.nn.Conv1d(in_channels,       " Enter the number of channels in the image "
                out_channels,      " The number of channels generated by convolution "
                kernel_size,       " The size of the convolution kernel "
                stride,            " The pace of convolution . The default value is ：1"
                padding,           " Fill added to both sides of the input . The default value is ：0"
                dilation,          " The spacing between kernel elements . The default value is ：1"
                groups,            " The number of blocked connections from the input channel to the output channel . The default value is ：1"
                bias,              "If True, Add learnable deviations to the output . Default ：True"
                padding_mode       "'zeros', 'reflect', 'replicate'  or  'circular'.  Default ：'zeros'"
                )

in_channels ( int ) – Enter the number of channels in the image
out_channels ( int ) – The number of channels generated by convolution
kernel_size ( int or tuple ) – The size of the convolution kernel
stride ( int or tuple , optional ) – The pace of convolution . The default value is ：1
padding ( int , tuple or str , optional ) – Fill added to both sides of the input . The default value is ：0
dilation ( int or tuple , optional ) – The spacing between kernel elements . The default value is ：1
groups ( int , optional ) – The number of blocked connections from the input channel to the output channel . The default value is ：1
bias ( bool , optional ) – If True, Add learnable deviations to the output . Default ：True
padding_mode ( character string , Optional ) – ‘zeros’, ‘reflect’, ‘replicate’ or ’circular’. Default ：‘zeros’

Two 、Conv1d I / O and convolution kernel dimension

input – $(\text{minibatch},\text{in\_channels},iW)$ （ Batch size , Number of data channels , Data length ）
output – $(\text{minibatch},\text{out\_channels },iW)$ （ Batch size , Number of channels generated , Length after convolution ）

Convoluted dimension ：(n - k + 2 * p ) / s + 1
k: Convolution kernel size ,p: Use boundary fill ,s: step .

Convolution kernel dimension ：$(\text{in\_channels},kernel\_size,\text{out\_channels })$

among ：out_channels dimension , Represents the number of convolution kernels , Used to extract multidimensional features .

3、 ... and 、Conv1d The calculation process

1. Test one ：in_channels=1, out_channels=1

The convolution is defined as follows ： Input channel ：1, Output channel ：1, Convolution kernel ：1 * 3 * 1, step ：1, fill ：0
Input ： Batch size ：1, Number of data channels ：1, Data length ： 5

import torch
import torch.nn as nn
input = torch.randn(1, 1, 5)
conv = nn.Conv1d(in_channels=1, out_channels=1, kernel_size=3, stride=1, padding=0)
out = conv(input)

The first convolution is calculated as follows ：

Move the calculation backward according to the step size ：

Output ： Batch size ：1, Number of data channels ：1, Data length ： 3

Running results ：

2. Test two ：in_channels=1, out_channels=2

The convolution is defined as follows ： Input channel ：1, Output channel ：2, Convolution kernel ：1 3 2 （ Two 13, Extracting two-dimensional features ）*, step ：1, fill ：0
Input ： Batch size ：1, Number of data channels ：1, Data length ： 5

import torch
import torch.nn as nn
input = torch.randn(1, 1, 5)
conv = nn.Conv1d(in_channels=1, out_channels=2, kernel_size=3, stride=1, padding=0)
out = conv(input)

Output ： Batch size ：1, Number of data channels ：2, Data length ： 3

3. Test three ：in_channels=8, out_channels=1

The convolution is defined as follows ： Input channel ：8, Output channel ：1, Convolution kernel ：8 * 3 * 1 , step ：1, fill ：0
Input ： Batch size ：1, Number of data channels ：1, Data length ： 7

import torch
import torch.nn as nn
input = torch.randn(1, 8, 7)
conv = nn.Conv1d(in_channels=8, out_channels=1, kernel_size=3, stride=1, padding=0)
out = conv(input)

Output ： Batch size ：1, Number of data channels ：1, Data length ： 5

3. Test three ：in_channels=8, out_channels=2

The convolution is defined as follows ： Input channel ：8, Output channel ：1, Convolution kernel ：8 * 3 * 2 , step ：1, fill ：0
Input ： Batch size ：1, Number of data channels ：1, Data length ： 7

import torch
import torch.nn as nn
input = torch.randn(1, 8, 7)
conv = nn.Conv1d(in_channels=8, out_channels=2, kernel_size=3, stride=1, padding=0)
out = conv(input)

Output ： Batch size ：1, Number of data channels ：2, Data length ： 5

Four 、Conv1d Application in text – TextCNN

The paper ：Convolutional Neural Networks for Sentence Classification

The model framework is shown in the figure below .

Suppose we need to classify sentences . Every word in a sentence is made up of n It's made up of vector words , That is to say, the size of the input matrix is m*n, among m For sentence length . stay pytorch in , Convolution from left to right , Therefore, you need to exchange the input dimension , namely （n, m）— ( Word vector dimension , Sentence length )

As shown in the figure above ： The input dimension is ：（5, 7）：

import torch
import torch.nn as nn
input = torch.randn(1, 5, 7)

CNN Need to convolute the input sample , For text data , It's kind of like N-gram In extracting the local correlation between words . There are three types of pictures kernel_size Convolution of , Namely 2,3,4, Every kernel_size There are two filter（ In practice filter There will be a lot of ）. Apply different words in different windows filter, The resulting 6 A convoluted vector .

With kernel_size = 4 For example ： Convolution kernel dimension （5, 4）

import torch
import torch.nn as nn
input = torch.randn(1, 5, 7)
conv = nn.Conv1d(in_channels=5, out_channels=1, kernel_size=4, stride=1, padding=0)
out = conv(input)

Output ：

Then maximize the pooling of each vector and splice the pooling values , Finally, we get the feature representation of this sentence , Throw this sentence vector to the classifier for classification , This completes the whole process .