Pytorch-卷积层及其操作

卷积层

1、1d/2d/3d

（1）卷积运算：卷积核在输入信号（图像）上滑动，相应位置上进行乘加
（2）卷积核：又称为滤波器，过滤器，可认为是某种模式，某种特征。
（3）卷积过程类似于用一个模版去图像上寻找与它相似的区域，与卷积核模式越相似，激活值越高，从而实现特征提取。
（4）AlexNet卷积核可视化，发现卷积核学习到的是边缘，条纹，色彩这一些细节模式

卷积操作：

卷积维度：一般情况下，卷积核在几个维度上滑动，就是几维卷积
1d卷积

2d卷积：

https://mlnotebook.github.io/post/CNN1/

3d卷积

2、nn.Conv2d

功能：对多个二维信号进行二维卷积


卷积维度：一般情况下，卷积核在几个维度上滑动，就是几维卷积

import os
import torch.nn as nn
from PIL import Image
from torchvision import transforms
from matplotlib import pyplot as plt
path_tools = os.path.join("tools","common_tools.py")
# print(path_tools)
from tools.common_tools import transform_invert,set_seed
set_seed(3)

# 加载图像
img = Image.open('lena.png').convert('RGB') # 0~255
plt.imshow(img)

# 转换成tensor向量
img_transform = transforms.Compose([transforms.ToTensor()])
img_tensor = img_transform(img)
# 在dim=0维度增加一个批次
# 因为卷积的输入tensor必须是四维的：
# input输入：(N, C, H, W)
img_tensor.unsqueeze_(dim=0) # C*H*W to B*C*H*W

# 创建卷积层
flag = 1
# flag = 0
if flag:
    conv_layer = nn.Conv2d(3,1,3) # input:(i, o, size) weights:(o, i , h, w)
    # 参数初始化
    nn.init.xavier_normal_(conv_layer.weight.data)
    img_conv = conv_layer(img_tensor)

# 可视化
print("卷积前尺寸：{}\n卷积后尺寸：{}".format(img_tensor.shape,img_conv.shape))
img_conv = transform_invert(img_conv[0,0:1,...],img_transform)
img_raw = transform_invert(img_tensor.squeeze(),img_transform)
plt.subplot(122).imshow(img_conv,cmap='gray')
plt.subplot(121).imshow(img_raw)
plt.show()

卷积前尺寸：torch.Size([1, 3, 512, 512])
卷积后尺寸：torch.Size([1, 1, 510, 510])

3、转置卷积-nn.ConvTranspose

转置卷积又称为反卷积(Deconvolution)和部分跨越卷积(Fractionally-strided Convolution) ，用于对图像进行上采样(UpSample)

参考知乎大佬们：怎样通俗理解反卷积？

flag = 1
if flag:
    # 设置一个stride=2
    conv_layer = nn.ConvTranspose2d(3,1,3,stride=2)  # input:(i, o, size)
    # 参数初始化
    nn.init.xavier_normal_(conv_layer.weight.data)
    img_conv = conv_layer(img_tensor)
    
# 可视化
print("卷积前尺寸：{}\n卷积后尺寸：{}".format(img_tensor.shape,img_conv.shape))
img_conv = transform_invert(img_conv[0,0:1,...],img_transform)
img_raw = transform_invert(img_tensor.squeeze(),img_transform)
plt.subplot(122).imshow(img_conv,cmap='gray')
plt.subplot(121).imshow(img_raw)
plt.show()

卷积前尺寸：torch.Size([1, 3, 512, 512])
卷积后尺寸：torch.Size([1, 1, 1025, 1025])

# help(nn.Conv2d)
help(nn.init.xavier_normal_)

Help on function xavier_normal_ in module torch.nn.init:

xavier_normal_(tensor, gain=1.0)
    Fills the input `Tensor` with values according to the method
    described in `Understanding the difficulty of training deep feedforward
    neural networks` - Glorot, X. & Bengio, Y. (2010), using a normal
    distribution. The resulting tensor will have values sampled from
    :math:`\mathcal{N}(0, \text{std}^2)` where
    
    .. math::
        \text{std} = \text{gain} \times \sqrt{\frac{2}{\text{fan\_in} + \text{fan\_out}}}
    
    Also known as Glorot initialization.
    
    Args:
        tensor: an n-dimensional `torch.Tensor`
        gain: an optional scaling factor
    
    Examples:
        >>> w = torch.empty(3, 5)
        >>> nn.init.xavier_normal_(w)

import torch
w = torch.empty(3,5)
print(w)
nn.init.xavier_normal_(w)
print(w)

tensor([[0.1107, 0.1588, 0.7464, 0.8560, 0.2025],
        [0.1505, 0.2469, 0.3596, 0.3466, 0.1254],
        [0.7501, 0.0495, 0.5417, 0.2369, 0.4551]])
tensor([[ 1.1172,  1.2298,  1.2443, -0.0800,  0.0806],
        [ 0.4042, -0.5826,  0.1932, -0.6775,  0.2642],
        [-0.5724,  0.0882, -0.0440, -0.5700,  0.4701]])

上述卷积操作的图来源于：https://mlnotebook.github.io/post/CNN1/