文章目录
实现【卷积-池化-激活】代码,并分析总结
使用不同的卷积核进行卷积
再进行最大池化,池化采用的是2×2的窗口,间隔为2
再使用relu函数进行激活,就是把负数都变成0,正数不变
以上就是进行一次卷积-池化-激活的过程。如果图片有卷积核的特征的话,明显了一些。通过多次的卷积-池化-激活(三者的顺序可以不固定),最后通过全连接层学习得到输入数据的特征表达,可以诸如图像分类,视觉对象等定位任务。
1. For循环版本:手工实现 卷积-池化-激活
import numpy as np
x = np.array([[-1, -1, -1, -1, -1, -1, -1, -1, -1],
[-1, 1, -1, -1, -1, -1, -1, 1, -1],
[-1, -1, 1, -1, -1, -1, 1, -1, -1],
[-1, -1, -1, 1, -1, 1, -1, -1, -1],
[-1, -1, -1, -1, 1, -1, -1, -1, -1],
[-1, -1, -1, 1, -1, 1, -1, -1, -1],
[-1, -1, 1, -1, -1, -1, 1, -1, -1],
[-1, 1, -1, -1, -1, -1, -1, 1, -1],
[-1, -1, -1, -1, -1, -1, -1, -1, -1]])
print("x=\n", x)
# 初始化 三个 卷积核
Kernel = [[0 for i in range(0, 3)] for j in range(0, 3)]
Kernel[0] = np.array([[1, -1, -1],
[-1, 1, -1],
[-1, -1, 1]])
Kernel[1] = np.array([[1, -1, 1],
[-1, 1, -1],
[1, -1, 1]])
Kernel[2] = np.array([[-1, -1, 1],
[-1, 1, -1],
[1, -1, -1]])
# --------------- 卷积 ---------------
stride = 1 # 步长
feature_map_h = 7 # 特征图的高
feature_map_w = 7 # 特征图的宽
feature_map = [0 for i in range(0, 3)] # 初始化3个特征图
for i in range(0, 3):
feature_map[i] = np.zeros((feature_map_h, feature_map_w)) # 初始化特征图
for h in range(feature_map_h): # 向下滑动,得到卷积后的固定行
for w in range(feature_map_w): # 向右滑动,得到卷积后的固定行的列
v_start = h * stride # 滑动窗口的起始行(高)
v_end = v_start + 3 # 滑动窗口的结束行(高)
h_start = w * stride # 滑动窗口的起始列(宽)
h_end = h_start + 3 # 滑动窗口的结束列(宽)
window = x[v_start:v_end, h_start:h_end] # 从图切出一个滑动窗口
for i in range(0, 3):
feature_map[i][h, w] = np.divide(np.sum(np.multiply(window, Kernel[i][:, :])), 9)
print("feature_map:\n", np.around(feature_map, decimals=2))
# --------------- 池化 ---------------
pooling_stride = 2 # 步长
pooling_h = 4 # 特征图的高
pooling_w = 4 # 特征图的宽
feature_map_pad_0 = [[0 for i in range(0, 8)] for j in range(0, 8)]
for i in range(0, 3): # 特征图 补 0 ,行 列 都要加 1 (因为上一层是奇数,池化窗口用的偶数)
feature_map_pad_0[i] = np.pad(feature_map[i], ((0, 1), (0, 1)), 'constant', constant_values=(0, 0))
# print("feature_map_pad_0 0:\n", np.around(feature_map_pad_0[0], decimals=2))
pooling = [0 for i in range(0, 3)]
for i in range(0, 3):
pooling[i] = np.zeros((pooling_h, pooling_w)) # 初始化特征图
for h in range(pooling_h): # 向下滑动,得到卷积后的固定行
for w in range(pooling_w): # 向右滑动,得到卷积后的固定行的列
v_start = h * pooling_stride # 滑动窗口的起始行(高)
v_end = v_start + 2 # 滑动窗口的结束行(高)
h_start = w * pooling_stride # 滑动窗口的起始列(宽)
h_end = h_start + 2 # 滑动窗口的结束列(宽)
for i in range(0, 3):
pooling[i][h, w] = np.max(feature_map_pad_0[i][v_start:v_end, h_start:h_end])
print("pooling:\n", np.around(pooling[0], decimals=2))
print("pooling:\n", np.around(pooling[1], decimals=2))
print("pooling:\n", np.around(pooling[2], decimals=2))
# --------------- 激活 ---------------
def relu(x):
return (abs(x) + x) / 2
relu_map_h = 7 # 特征图的高
relu_map_w = 7 # 特征图的宽
relu_map = [0 for i in range(0, 3)] # 初始化3个特征图
for i in range(0, 3):
relu_map[i] = np.zeros((relu_map_h, relu_map_w)) # 初始化特征图
for i in range(0, 3):
relu_map[i] = relu(feature_map[i])
print("relu map :\n",np.around(relu_map[0], decimals=2))
print("relu map :\n",np.around(relu_map[1], decimals=2))
print("relu map :\n",np.around(relu_map[2], decimals=2))
实现结果就是上图,有些地方四舍五入不同。仅展示池化结果
2. Pytorch版本:调用函数完成 卷积-池化-激活
# https://blog.csdn.net/qq_26369907/article/details/88366147
# https://zhuanlan.zhihu.com/p/405242579
import numpy as np
import torch
import torch.nn as nn
x = torch.tensor([[[[-1, -1, -1, -1, -1, -1, -1, -1, -1],
[-1, 1, -1, -1, -1, -1, -1, 1, -1],
[-1, -1, 1, -1, -1, -1, 1, -1, -1],
[-1, -1, -1, 1, -1, 1, -1, -1, -1],
[-1, -1, -1, -1, 1, -1, -1, -1, -1],
[-1, -1, -1, 1, -1, 1, -1, -1, -1],
[-1, -1, 1, -1, -1, -1, 1, -1, -1],
[-1, 1, -1, -1, -1, -1, -1, 1, -1],
[-1, -1, -1, -1, -1, -1, -1, -1, -1]]]], dtype=torch.float)
print(x.shape)
print(x)
print("--------------- 卷积 ---------------")
conv1 = nn.Conv2d(1, 1, (3, 3), 1) # in_channel , out_channel , kennel_size , stride
conv1.weight.data = torch.Tensor([[[[1, -1, -1],
[-1, 1, -1],
[-1, -1, 1]]
]])
conv2 = nn.Conv2d(1, 1, (3, 3), 1) # in_channel , out_channel , kennel_size , stride
conv2.weight.data = torch.Tensor([[[[1, -1, 1],
[-1, 1, -1],
[1, -1, 1]]
]])
conv3 = nn.Conv2d(1, 1, (3, 3), 1) # in_channel , out_channel , kennel_size , stride
conv3.weight.data = torch.Tensor([[[[-1, -1, 1],
[-1, 1, -1],
[1, -1, -1]]
]])
feature_map1 = conv1(x)
feature_map2 = conv2(x)
feature_map3 = conv3(x)
print(feature_map1 / 9)
print(feature_map2 / 9)
print(feature_map3 / 9)
print("--------------- 池化 ---------------")
max_pool = nn.MaxPool2d(2, padding=0, stride=2) # Pooling
zeroPad = nn.ZeroPad2d(padding=(0, 1, 0, 1)) # pad 0 , Left Right Up Down
feature_map_pad_0_1 = zeroPad(feature_map1)
feature_pool_1 = max_pool(feature_map_pad_0_1)
feature_map_pad_0_2 = zeroPad(feature_map2)
feature_pool_2 = max_pool(feature_map_pad_0_2)
feature_map_pad_0_3 = zeroPad(feature_map3)
feature_pool_3 = max_pool(feature_map_pad_0_3)
print(feature_pool_1.size())
print(feature_pool_1 / 9)
print(feature_pool_2 / 9)
print(feature_pool_3 / 9)
print("--------------- 激活 ---------------")
activation_function = nn.ReLU()
feature_relu1 = activation_function(feature_map1)
feature_relu2 = activation_function(feature_map2)
feature_relu3 = activation_function(feature_map3)
print(feature_relu1 / 9)
print(feature_relu2 / 9)
print(feature_relu3 / 9)
conv1 = nn.Conv2d(1, 1, (3, 3), 1)
这里我们用的都是输入输出通道是1,3×3卷积核,步长是1
max_pool = nn.MaxPool2d(2, padding=0, stride=2)
这里是2×2,步长是2。
运行结果小数位保留的多更加精确。
--------------- 卷积 ---------------
tensor([[[[ 0.7677, -0.1212, 0.1011, 0.3233, 0.5455, -0.1212, 0.3233],
[-0.1212, 0.9900, -0.1212, 0.3233, -0.1212, 0.1011, -0.1212],
[ 0.1011, -0.1212, 0.9900, -0.3434, 0.1011, -0.1212, 0.5455],
[ 0.3233, 0.3233, -0.3434, 0.5455, -0.3434, 0.3233, 0.3233],
[ 0.5455, -0.1212, 0.1011, -0.3434, 0.9900, -0.1212, 0.1011],
[-0.1212, 0.1011, -0.1212, 0.3233, -0.1212, 0.9900, -0.1212],
[ 0.3233, -0.1212, 0.5455, 0.3233, 0.1011, -0.1212, 0.7677]]]],
grad_fn=<DivBackward0>)
tensor([[[[ 0.3403, -0.5486, 0.1181, -0.1041, 0.1181, -0.5486, 0.3403],
[-0.5486, 0.5626, -0.5486, 0.3403, -0.5486, 0.5626, -0.5486],
[ 0.1181, -0.5486, 0.5626, -0.7708, 0.5626, -0.5486, 0.1181],
[-0.1041, 0.3403, -0.7708, 1.0070, -0.7708, 0.3403, -0.1041],
[ 0.1181, -0.5486, 0.5626, -0.7708, 0.5626, -0.5486, 0.1181],
[-0.5486, 0.5626, -0.5486, 0.3403, -0.5486, 0.5626, -0.5486],
[ 0.3403, -0.5486, 0.1181, -0.1041, 0.1181, -0.5486, 0.3403]]]],
grad_fn=<DivBackward0>)
tensor([[[[ 0.3355, -0.1090, 0.5577, 0.3355, 0.1132, -0.1090, 0.7799],
[-0.1090, 0.1132, -0.1090, 0.3355, -0.1090, 1.0021, -0.1090],
[ 0.5577, -0.1090, 0.1132, -0.3312, 1.0021, -0.1090, 0.1132],
[ 0.3355, 0.3355, -0.3312, 0.5577, -0.3312, 0.3355, 0.3355],
[ 0.1132, -0.1090, 1.0021, -0.3312, 0.1132, -0.1090, 0.5577],
[-0.1090, 1.0021, -0.1090, 0.3355, -0.1090, 0.1132, -0.1090],
[ 0.7799, -0.1090, 0.1132, 0.3355, 0.5577, -0.1090, 0.3355]]]],
grad_fn=<DivBackward0>)
--------------- 池化 ---------------
torch.Size([1, 1, 4, 4])
tensor([[[[0.9900, 0.3233, 0.5455, 0.3233],
[0.3233, 0.9900, 0.3233, 0.5455],
[0.5455, 0.3233, 0.9900, 0.1011],
[0.3233, 0.5455, 0.1011, 0.7677]]]], grad_fn=<DivBackward0>)
tensor([[[[0.5626, 0.3403, 0.5626, 0.3403],
[0.3403, 1.0070, 0.5626, 0.1181],
[0.5626, 0.5626, 0.5626, 0.1181],
[0.3403, 0.1181, 0.1181, 0.3403]]]], grad_fn=<DivBackward0>)
tensor([[[[0.3355, 0.5577, 1.0021, 0.7799],
[0.5577, 0.5577, 1.0021, 0.3355],
[1.0021, 1.0021, 0.1132, 0.5577],
[0.7799, 0.3355, 0.5577, 0.3355]]]], grad_fn=<DivBackward0>)
--------------- 激活 ---------------
tensor([[[[0.7677, 0.0000, 0.1011, 0.3233, 0.5455, 0.0000, 0.3233],
[0.0000, 0.9900, 0.0000, 0.3233, 0.0000, 0.1011, 0.0000],
[0.1011, 0.0000, 0.9900, 0.0000, 0.1011, 0.0000, 0.5455],
[0.3233, 0.3233, 0.0000, 0.5455, 0.0000, 0.3233, 0.3233],
[0.5455, 0.0000, 0.1011, 0.0000, 0.9900, 0.0000, 0.1011],
[0.0000, 0.1011, 0.0000, 0.3233, 0.0000, 0.9900, 0.0000],
[0.3233, 0.0000, 0.5455, 0.3233, 0.1011, 0.0000, 0.7677]]]],
grad_fn=<DivBackward0>)
tensor([[[[0.3403, 0.0000, 0.1181, 0.0000, 0.1181, 0.0000, 0.3403],
[0.0000, 0.5626, 0.0000, 0.3403, 0.0000, 0.5626, 0.0000],
[0.1181, 0.0000, 0.5626, 0.0000, 0.5626, 0.0000, 0.1181],
[0.0000, 0.3403, 0.0000, 1.0070, 0.0000, 0.3403, 0.0000],
[0.1181, 0.0000, 0.5626, 0.0000, 0.5626, 0.0000, 0.1181],
[0.0000, 0.5626, 0.0000, 0.3403, 0.0000, 0.5626, 0.0000],
[0.3403, 0.0000, 0.1181, 0.0000, 0.1181, 0.0000, 0.3403]]]],
grad_fn=<DivBackward0>)
tensor([[[[0.3355, 0.0000, 0.5577, 0.3355, 0.1132, 0.0000, 0.7799],
[0.0000, 0.1132, 0.0000, 0.3355, 0.0000, 1.0021, 0.0000],
[0.5577, 0.0000, 0.1132, 0.0000, 1.0021, 0.0000, 0.1132],
[0.3355, 0.3355, 0.0000, 0.5577, 0.0000, 0.3355, 0.3355],
[0.1132, 0.0000, 1.0021, 0.0000, 0.1132, 0.0000, 0.5577],
[0.0000, 1.0021, 0.0000, 0.3355, 0.0000, 0.1132, 0.0000],
[0.7799, 0.0000, 0.1132, 0.3355, 0.5577, 0.0000, 0.3355]]]],
3. 可视化:了解数字与图像之间的关系
# https://blog.csdn.net/qq_26369907/article/details/88366147
# https://zhuanlan.zhihu.com/p/405242579
import torch
import torch.nn as nn
import matplotlib.pyplot as plt
plt.rcParams['font.sans-serif']=['SimHei'] #用来正常显示中文标签
plt.rcParams['axes.unicode_minus']=False #用来正常显示负号 #有中文出现的情况,需要u'内容
x = torch.tensor([[[[-1, -1, -1, -1, -1, -1, -1, -1, -1],
[-1, 1, -1, -1, -1, -1, -1, 1, -1],
[-1, -1, 1, -1, -1, -1, 1, -1, -1],
[-1, -1, -1, 1, -1, 1, -1, -1, -1],
[-1, -1, -1, -1, 1, -1, -1, -1, -1],
[-1, -1, -1, 1, -1, 1, -1, -1, -1],
[-1, -1, 1, -1, -1, -1, 1, -1, -1],
[-1, 1, -1, -1, -1, -1, -1, 1, -1],
[-1, -1, -1, -1, -1, -1, -1, -1, -1]]]], dtype=torch.float)
print(x.shape)
print(x)
img = x.data.squeeze().numpy() # 将输出转换为图片的格式
plt.imshow(img, cmap='gray')
plt.title('原图')
plt.show()
print("--------------- 卷积 ---------------")
conv1 = nn.Conv2d(1, 1, (3, 3), 1) # in_channel , out_channel , kennel_size , stride
conv1.weight.data = torch.Tensor([[[[1, -1, -1],
[-1, 1, -1],
[-1, -1, 1]]
]])
img = conv1.weight.data.squeeze().numpy() # 将输出转换为图片的格式
plt.imshow(img, cmap='gray')
plt.title('Kernel 1')
plt.show()
conv2 = nn.Conv2d(1, 1, (3, 3), 1) # in_channel , out_channel , kennel_size , stride
conv2.weight.data = torch.Tensor([[[[1, -1, 1],
[-1, 1, -1],
[1, -1, 1]]
]])
img = conv2.weight.data.squeeze().numpy() # 将输出转换为图片的格式
plt.imshow(img, cmap='gray')
plt.title('Kernel 2')
plt.show()
conv3 = nn.Conv2d(1, 1, (3, 3), 1) # in_channel , out_channel , kennel_size , stride
conv3.weight.data = torch.Tensor([[[[-1, -1, 1],
[-1, 1, -1],
[1, -1, -1]]
]])
img = conv3.weight.data.squeeze().numpy() # 将输出转换为图片的格式
plt.imshow(img, cmap='gray')
plt.title('Kernel 3')
plt.show()
feature_map1 = conv1(x)
feature_map2 = conv2(x)
feature_map3 = conv3(x)
print(feature_map1 / 9)
print(feature_map2 / 9)
print(feature_map3 / 9)
img = feature_map1.data.squeeze().numpy() # 将输出转换为图片的格式
plt.imshow(img, cmap='gray')
plt.title('卷积后的特征图1')
plt.show()
print("--------------- 池化 ---------------")
max_pool = nn.MaxPool2d(2, padding=0, stride=2) # Pooling
zeroPad = nn.ZeroPad2d(padding=(0, 1, 0, 1)) # pad 0 , Left Right Up Down
feature_map_pad_0_1 = zeroPad(feature_map1)
feature_pool_1 = max_pool(feature_map_pad_0_1)
feature_map_pad_0_2 = zeroPad(feature_map2)
feature_pool_2 = max_pool(feature_map_pad_0_2)
feature_map_pad_0_3 = zeroPad(feature_map3)
feature_pool_3 = max_pool(feature_map_pad_0_3)
print(feature_pool_1.size())
print(feature_pool_1 / 9)
print(feature_pool_2 / 9)
print(feature_pool_3 / 9)
img = feature_pool_1.data.squeeze().numpy() # 将输出转换为图片的格式
plt.imshow(img, cmap='gray')
plt.title('卷积池化后的特征图1')
plt.show()
print("--------------- 激活 ---------------")
activation_function = nn.ReLU()
feature_relu1 = activation_function(feature_map1)
feature_relu2 = activation_function(feature_map2)
feature_relu3 = activation_function(feature_map3)
print(feature_relu1 / 9)
print(feature_relu2 / 9)
print(feature_relu3 / 9)
img = feature_relu1.data.squeeze().numpy() # 将输出转换为图片的格式
plt.imshow(img, cmap='gray')
plt.title('卷积 + relu 后的特征图1')
plt.show()
运行效果如下:
原图
卷积核
![](https://img-blog.csdnimg.cn/e633591fb6c64530a3fd072120229281.png)
![](https://img-blog.csdnimg.cn/091ab4d4ddde49f39b0ccfb75049cd72.png)
![](https://i-blog.csdnimg.cn/blog_migrate/0646fc0bf86d7a5baa63e24e76c50a12.png)
特征图
![](https://img-blog.csdnimg.cn/258aeb31443c4561b16c74a7a464f761.png)
![](https://img-blog.csdnimg.cn/12d128121ea449deb669a80fedb03915.png)
![](https://i-blog.csdnimg.cn/blog_migrate/844e17640e8a104a9aeec66edc19e73e.png)
可以尝试改变一下池化的参数
2×2,步长1
3×3 步长是1
发现都不如原来的好,池化的步长还是要合适的。大小最好还是小于卷积核的大小
参考资料
【2021-2022 春学期】人工智能-作业5:卷积-池化-激活
pytorch之torch.nn.Conv2d()函数详解
torch.nn.MaxPool2d