【22-23 春学期】AI作业9-卷积3-XO识别

1. For循环版本：手工实现 卷积-池化-激活

代码：

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))

运行结果：

D:\python\python.exe D:\pythonprojects\text\main.py

x=

 [[-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]]

feature_map:

 [[[ 0.78 -0.11  0.11  0.33  0.56 -0.11  0.33]

  [-0.11  1.   -0.11  0.33 -0.11  0.11 -0.11]

  [ 0.11 -0.11  1.   -0.33  0.11 -0.11  0.56]

  [ 0.33  0.33 -0.33  0.56 -0.33  0.33  0.33]

  [ 0.56 -0.11  0.11 -0.33  1.   -0.11  0.11]

  [-0.11  0.11 -0.11  0.33 -0.11  1.   -0.11]

  [ 0.33 -0.11  0.56  0.33  0.11 -0.11  0.78]]

 [[ 0.33 -0.56  0.11 -0.11  0.11 -0.56  0.33]

  [-0.56  0.56 -0.56  0.33 -0.56  0.56 -0.56]

  [ 0.11 -0.56  0.56 -0.78  0.56 -0.56  0.11]

  [-0.11  0.33 -0.78  1.   -0.78  0.33 -0.11]

  [ 0.11 -0.56  0.56 -0.78  0.56 -0.56  0.11]

  [-0.56  0.56 -0.56  0.33 -0.56  0.56 -0.56]

  [ 0.33 -0.56  0.11 -0.11  0.11 -0.56  0.33]]

 [[ 0.33 -0.11  0.56  0.33  0.11 -0.11  0.78]

  [-0.11  0.11 -0.11  0.33 -0.11  1.   -0.11]

  [ 0.56 -0.11  0.11 -0.33  1.   -0.11  0.11]

  [ 0.33  0.33 -0.33  0.56 -0.33  0.33  0.33]

  [ 0.11 -0.11  1.   -0.33  0.11 -0.11  0.56]

  [-0.11  1.   -0.11  0.33 -0.11  0.11 -0.11]

  [ 0.78 -0.11  0.11  0.33  0.56 -0.11  0.33]]]

pooling:

 [[1.   0.33 0.56 0.33]

 [0.33 1.   0.33 0.56]

 [0.56 0.33 1.   0.11]

 [0.33 0.56 0.11 0.78]]

pooling:

 [[0.56 0.33 0.56 0.33]

 [0.33 1.   0.56 0.11]

 [0.56 0.56 0.56 0.11]

 [0.33 0.11 0.11 0.33]]

pooling:

 [[0.33 0.56 1.   0.78]

 [0.56 0.56 1.   0.33]

 [1.   1.   0.11 0.56]

 [0.78 0.33 0.56 0.33]]

relu map :

 [[0.78 0.   0.11 0.33 0.56 0.   0.33]

 [0.   1.   0.   0.33 0.   0.11 0.  ]

 [0.11 0.   1.   0.   0.11 0.   0.56]

 [0.33 0.33 0.   0.56 0.   0.33 0.33]

 [0.56 0.   0.11 0.   1.   0.   0.11]

 [0.   0.11 0.   0.33 0.   1.   0.  ]

 [0.33 0.   0.56 0.33 0.11 0.   0.78]]

relu map :

 [[0.33 0.   0.11 0.   0.11 0.   0.33]

 [0.   0.56 0.   0.33 0.   0.56 0.  ]

 [0.11 0.   0.56 0.   0.56 0.   0.11]

 [0.   0.33 0.   1.   0.   0.33 0.  ]

 [0.11 0.   0.56 0.   0.56 0.   0.11]

 [0.   0.56 0.   0.33 0.   0.56 0.  ]

 [0.33 0.   0.11 0.   0.11 0.   0.33]]

relu map :

 [[0.33 0.   0.56 0.33 0.11 0.   0.78]

 [0.   0.11 0.   0.33 0.   1.   0.  ]

 [0.56 0.   0.11 0.   1.   0.   0.11]

 [0.33 0.33 0.   0.56 0.   0.33 0.33]

 [0.11 0.   1.   0.   0.11 0.   0.56]

 [0.   1.   0.   0.33 0.   0.11 0.  ]

 [0.78 0.   0.11 0.33 0.56 0.   0.33]]

Process finished with exit code 0

2. Pytorch版本：调用函数完成 卷积-池化-激活

代码：

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)

运行结果

D:\python\python.exe D:\pythonprojects\1\main.py

torch.Size([1, 1, 9, 9])

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.]]]])

--------------- 卷积  ---------------

tensor([[[[ 0.7555, -0.1333,  0.0889,  0.3111,  0.5333, -0.1333,  0.3111],

          [-0.1333,  0.9778, -0.1333,  0.3111, -0.1333,  0.0889, -0.1333],

          [ 0.0889, -0.1333,  0.9778, -0.3556,  0.0889, -0.1333,  0.5333],

          [ 0.3111,  0.3111, -0.3556,  0.5333, -0.3556,  0.3111,  0.3111],

          [ 0.5333, -0.1333,  0.0889, -0.3556,  0.9778, -0.1333,  0.0889],

          [-0.1333,  0.0889, -0.1333,  0.3111, -0.1333,  0.9778, -0.1333],

          [ 0.3111, -0.1333,  0.5333,  0.3111,  0.0889, -0.1333,  0.7555]]]],

       grad_fn=<DivBackward0>)

tensor([[[[ 0.3014, -0.5875,  0.0792, -0.1430,  0.0792, -0.5875,  0.3014],

          [-0.5875,  0.5237, -0.5875,  0.3014, -0.5875,  0.5237, -0.5875],

          [ 0.0792, -0.5875,  0.5237, -0.8097,  0.5237, -0.5875,  0.0792],

          [-0.1430,  0.3014, -0.8097,  0.9681, -0.8097,  0.3014, -0.1430],

          [ 0.0792, -0.5875,  0.5237, -0.8097,  0.5237, -0.5875,  0.0792],

          [-0.5875,  0.5237, -0.5875,  0.3014, -0.5875,  0.5237, -0.5875],

          [ 0.3014, -0.5875,  0.0792, -0.1430,  0.0792, -0.5875,  0.3014]]]],

       grad_fn=<DivBackward0>)

tensor([[[[ 0.3468, -0.0976,  0.5691,  0.3468,  0.1246, -0.0976,  0.7913],

          [-0.0976,  0.1246, -0.0976,  0.3468, -0.0976,  1.0135, -0.0976],

          [ 0.5691, -0.0976,  0.1246, -0.3198,  1.0135, -0.0976,  0.1246],

          [ 0.3468,  0.3468, -0.3198,  0.5691, -0.3198,  0.3468,  0.3468],

          [ 0.1246, -0.0976,  1.0135, -0.3198,  0.1246, -0.0976,  0.5691],

          [-0.0976,  1.0135, -0.0976,  0.3468, -0.0976,  0.1246, -0.0976],

          [ 0.7913, -0.0976,  0.1246,  0.3468,  0.5691, -0.0976,  0.3468]]]],

       grad_fn=<DivBackward0>)

--------------- 池化  ---------------

torch.Size([1, 1, 4, 4])

tensor([[[[0.9778, 0.3111, 0.5333, 0.3111],

          [0.3111, 0.9778, 0.3111, 0.5333],

          [0.5333, 0.3111, 0.9778, 0.0889],

          [0.3111, 0.5333, 0.0889, 0.7555]]]], grad_fn=<DivBackward0>)

tensor([[[[0.5237, 0.3014, 0.5237, 0.3014],

          [0.3014, 0.9681, 0.5237, 0.0792],

          [0.5237, 0.5237, 0.5237, 0.0792],

          [0.3014, 0.0792, 0.0792, 0.3014]]]], grad_fn=<DivBackward0>)

tensor([[[[0.3468, 0.5691, 1.0135, 0.7913],

          [0.5691, 0.5691, 1.0135, 0.3468],

          [1.0135, 1.0135, 0.1246, 0.5691],

          [0.7913, 0.3468, 0.5691, 0.3468]]]], grad_fn=<DivBackward0>)

--------------- 激活  ---------------

tensor([[[[0.7555, 0.0000, 0.0889, 0.3111, 0.5333, 0.0000, 0.3111],

          [0.0000, 0.9778, 0.0000, 0.3111, 0.0000, 0.0889, 0.0000],

          [0.0889, 0.0000, 0.9778, 0.0000, 0.0889, 0.0000, 0.5333],

          [0.3111, 0.3111, 0.0000, 0.5333, 0.0000, 0.3111, 0.3111],

          [0.5333, 0.0000, 0.0889, 0.0000, 0.9778, 0.0000, 0.0889],

          [0.0000, 0.0889, 0.0000, 0.3111, 0.0000, 0.9778, 0.0000],

          [0.3111, 0.0000, 0.5333, 0.3111, 0.0889, 0.0000, 0.7555]]]],

       grad_fn=<DivBackward0>)

tensor([[[[0.3014, 0.0000, 0.0792, 0.0000, 0.0792, 0.0000, 0.3014],

          [0.0000, 0.5237, 0.0000, 0.3014, 0.0000, 0.5237, 0.0000],

          [0.0792, 0.0000, 0.5237, 0.0000, 0.5237, 0.0000, 0.0792],

          [0.0000, 0.3014, 0.0000, 0.9681, 0.0000, 0.3014, 0.0000],

          [0.0792, 0.0000, 0.5237, 0.0000, 0.5237, 0.0000, 0.0792],

          [0.0000, 0.5237, 0.0000, 0.3014, 0.0000, 0.5237, 0.0000],

          [0.3014, 0.0000, 0.0792, 0.0000, 0.0792, 0.0000, 0.3014]]]],

       grad_fn=<DivBackward0>)

tensor([[[[0.3468, 0.0000, 0.5691, 0.3468, 0.1246, 0.0000, 0.7913],

          [0.0000, 0.1246, 0.0000, 0.3468, 0.0000, 1.0135, 0.0000],

          [0.5691, 0.0000, 0.1246, 0.0000, 1.0135, 0.0000, 0.1246],

          [0.3468, 0.3468, 0.0000, 0.5691, 0.0000, 0.3468, 0.3468],

          [0.1246, 0.0000, 1.0135, 0.0000, 0.1246, 0.0000, 0.5691],

          [0.0000, 1.0135, 0.0000, 0.3468, 0.0000, 0.1246, 0.0000],

          [0.7913, 0.0000, 0.1246, 0.3468, 0.5691, 0.0000, 0.3468]]]],

       grad_fn=<DivBackward0>)

Process finished with exit code 0

3.可视化：了解数字与图像之间的关系

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()

运行结果：