【Pytorch教程】Conv3D的推导

torch.nn.Conv3d

Conv3d初始化Parameters：

• in_channels (int) – 输入图片的通道数
• out_channels (int) – 卷积计算后输出的通道数
• kernel_size (int or tuple) – 卷积核大小，如(3,7,7)
• stride (int or tuple, optional) – 步长，默认为1
• padding (int, tuple or str, optional) – Padding added to all six sides of the input. Default: 0
• padding_mode (string, optional) – ‘zeros’, ‘reflect’, ‘replicate’ or ‘circular’. Default: ‘zeros’
• dilation (int or tuple, optional) – Spacing between kernel elements. Default: 1
• groups (int, optional) – Number of blocked connections from input channels to output channels. Default: 1
• bias (bool, optional) – If True, adds a learnable bias to the output. Default: True

输入/输出的shape：

• Input：$(N,C{}_i{}_n,D{}_i{}_n,H{}_i{}_n,W{}_i{}_n)$

$N$- batch size
$C{}_i{}_n$ - 输入图片的通道数
$D{}_i{}_n$ - 输入的帧数
$H{}_i{}_n$ - 输入图片的高
$W{}_i{}_n$ - 输入图片的宽

• Output：$(N,C{}_o{}_u{}_t,D{}_o{}_u{}_t,H{}_o{}_u{}_t,W{}_o{}_u{}_t)$

已知卷积的计算公式：
$$H=\lfloor \frac{H+2\times padding-kernelsize}{stride}+1\rfloor$$
计算各个维度的卷积后的长度：
$$D{}_o{}_u{}_t=\lfloor \frac{D{}_i{}_n+2\times padding[0]-dilation[0]\times (kernelsize[0]-1)-1}{stride[0]}+1\rfloor$$
$$H{}_o{}_u{}_t=\lfloor \frac{H{}_i{}_n+2\times padding[1]-dilation[1]\times (kernelsize[1]-1)-1}{stride[1]}+1\rfloor$$
$$W{}_o{}_u{}_t=\lfloor \frac{W{}_i{}_n+2\times padding[2]-dilation[2]\times (kernelsize[2]-1)-1}{stride[2]}+1\rfloor$$

Demo示例

import torch
import torch.nn as nn
from torch import autograd

# kernel_size的第哥一维度的值是每次处理的图像帧数，后面是卷积核的大小

# 输入/出通道数
# 卷积核大小
# 步长/补位
m = nn.Conv3d(3, 3, (3, 7, 7), stride=1, padding=0)

# 输入:
# batch_size
# Cin 通道数
# 提取时序特征的帧数
# h/w
input = autograd.Variable(torch.randn(1, 3, 7, 60, 40))
# 卷积计算推导:
# w1=(w0+2pad-kernel_size)/stride+1;


print(input.shape)
output = m(input)
print(output.size())
# 输出是 torch.Size([1, 3, 5, 54, 34])