PyTorch实现卷积和池化（kaggle环境）

import numpy as np
import torch
from torch import nn
from torch.autograd import Variable
from PIL import Image
import matplotlib.pylab as plt
%matplotlib inline


卷积层

导入图片并且转化为灰度图

kaggle环境中所有本地数据都有上传才能加载&＃xff01;

im &＃61; Image.open(&＃39;../input/cat.jpg&＃39;).convert(&＃39;L&＃39;)


将图片转化为矩阵

im &＃61; np.array(im, dtype&＃61;&＃39;float32&＃39;)


查看图片&＃xff0c;将图片矩阵转化为uint8类型&＃xff0c;uint8是无符号八位整型&＃xff0c;表示范围是[0, 255]的整数&＃xff0c;colormap为gray

plt.imshow(im.astype(&＃39;uint8&＃39;), cmap&＃61;&＃39;gray&＃39;)


im.shape


(121, 121)


将图片矩阵转化为Tensor&＃xff0c;并将图片大小转化为卷积输入的要求大小

print(im.shape)
im &＃61; torch.from_numpy(im.reshape(1,1,im.shape[0], im.shape[1]))
print(im.shape)


(121, 121)
torch.Size([1, 1, 121, 121])


使用nn.Conv2d定义卷积层。参数&＃xff1a;输入通道数&＃xff0c;输出通道数&＃xff0c;kernel大小&＃xff0c;偏置为0

conv1 &＃61; nn.Conv2d(1, 1, 3,bias&＃61;False)


定义kernel矩阵内容&＃xff0c;为3X3矩阵&＃xff0c;中间为8&＃xff0c;其余为-1

sobel_kernel &＃61; np.array([[-1,-1,-1],[-1,8,-1],[-1,-1,-1]],dtype&＃61;&＃39;float32&＃39;)
print(sobel_kernel.shape)


(3, 3)


将卷积核大小设置与卷积层通道相一致

sobel_kernel &＃61; sobel_kernel.reshape((1,1,3,3))
print(sobel_kernel.shape)


(1, 1, 3, 3)


将定义好得卷积核转化为Tensor&＃xff0c;并赋值给卷积层权重

conv1.weight.data &＃61; torch.from_numpy(sobel_kernel)


使用卷积层对图片进行卷积

edge1 &＃61; conv1(Variable(im))
print(edge1.shape)


torch.Size([1, 1, 119, 119])


用函数squeeze去掉卷积操作后的其他为1的维数&＃xff0c;并且将Tensor转化为numpy

print(edge1.data.squeeze().shape)
edge1 &＃61; edge1.data.squeeze().numpy()


torch.Size([119, 119])


将经过卷积后的图片显示出来

plt.imshow(edge1,cmap&＃61;&＃39;gray&＃39;)


池化层

导入最大池化层&＃xff0c;并定义池化大小

pool1 &＃61; nn.MaxPool2d(2, 2)


打印出张量格式的图片大小

print(&＃39;before max pool, image shape: {} x {}&＃39;.format(im.shape[2], im.shape[3]))


before max pool, image shape: 121 x 121


将Tensor转化为Variable&＃xff0c;进行最大池化操作。并将结果转化为numpy

small_im1 &＃61; pool1(Variable(im))
small_im1 &＃61; small_im1.data.squeeze().numpy()


打印出经过池化层后的图片大小

print(&＃39;after max pool, image shape: {} X {}&＃39;.format(small_im1.shape[0],small_im1.shape[1]))


after max pool, image shape: 60 X 60


打印出图片

plt.imshow(small_im1,cmap&＃61;&＃39;gray&＃39;)