热门标签
热门文章
- 1一个心塞的手忙脚乱的项目结束后的总结_项目结束各奔天涯
- 2解决：IDEA无法下载源码，Cannot download sources, sources not found for: xxxx_idea sources not found for: cn.newgrand.sup:sup-co
- 3对java技术提升的定级总结_java提升技术的级别
- 4正则表达式-限定位数的正数_正则表达式限制位数
- 5SpringBoot 简单开发流程（详解）_springboot开发
- 6mitmdump 脚本使用python第三方包方法（报错：in script xxx.py: No module named ‘xxx‘）_mitmdump in script e:\work_code\appnium_frame\util
- 7CMake错误: Imported targets not available for * version, Could NOT find *, ... 解决_调试时target not available
- 8⭐算法入门⭐《二叉树》简单04 —— LeetCode 104. 二叉树的最大深度
- 9centos文件和目录权限（umask）_centos umask
- 10LLM微调过程中灾难性遗忘问题解决方法_大模型lora微调如何避免灾难性遗忘
当前位置: article > 正文
理解深度学习的卷积层nn.Conv2d()、激活层nn.ReLU()、池化层nn.MaxPool2d()
作者：你好赵伟 | 2024-02-10 18:51:31
赞
踩
nn.relu
理解深度学习的卷积层、激活层、池化层
网上看了些教程，感觉还是来点实例可能更好理解一些。
下面分别把卷积层nn.Conv2d()、激活层nn.ReLU()、池化层nn.MaxPool2d()用代码表示出来，并且用图片显示加以理解。
读取图片
先读取一张图片，用来做卷积等操作的示范。
用cv2来读取读片，并转换成深度学习网络可使用的数据格式。
因为神经网络中有一个batch size的功能，但我们只有一张图片，所以需要image.unsqueeze(0) 一下增加一个数量
# 将图片转换为模型可用的数据
def getImageData(imgFile):
image = cv2.imread(imgFile, 0)
image = F.to_tensor(image) # shape为 [1, 768, 1024], [通道，高，宽]
image = image.unsqueeze(0) # shape为 [1, 1, 768, 1024], [数量，通道，高，宽]
return image
- 1
- 2
- 3
- 4
- 5
- 6
显示图片
在图片做完卷积、激活、池化后，需要将图片显示出来，用来观察这些操作到底做了什么事情。
显示图片，用的是matplotlib.pyplot。
这里只显示三个图片做示例。
# 将三个图片显示在plt上
def showTensorImg(image, title=""):
plt.figure(title) # 创建第一个图形
plt.sca(plt.subplot(1, 3, 1)) # 分成1行2列，选择第1个位置
plt.imshow(image[0].data, cmap=plt.cm.gray)
plt.sca(plt.subplot(1, 3, 2)) # 分成1行2列，选择第2个位置
plt.imshow(image[1].data, cmap=plt.cm.gray)
plt.sca(plt.subplot(1, 3, 3)) # 分成1行2列，选择第2个位置
plt.imshow(image[2].data, cmap=plt.cm.gray)
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
定义一个神经网络
用一个神经网络，在其中加入卷积层。
这里构建网络，用的是pytorch的 torch.nn。
# 定义模型
class MyModel(nn.Module):
def __init__(self):
super(MyModel, self).__init__()
self.conv0 = nn.Conv2d(in_channels=1, out_channels=3, kernel_size=(3, 3), stride=1, padding=1)
# 把模型里的卷积层和激活层后的效果，显示出来看一看
def forward(self, image):
img = self.conv0(image)
print("conv0:", img, img.shape, "\n")
showTensorImg(img.squeeze(), "conv0")
return img
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
查看卷积层结果
拿一张图片来做示范，先显示原图，再显示卷积层的结果。
if __name__ == "__main__":
# 显示原图
img = cv2.imread("Data/2.jpeg", 0)
plt.imshow(img)
print(img.shape)
net = MyModel()
image = getImageData("Data/2.jpeg")
img = net(image)
plt.show()
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
看看显示结果：
看看输出结果：
(8, 8) conv0: tensor([[[[ 0.4108, 0.6527, 0.8689, 0.8689, 0.8689, 0.8689, 0.8303, 0.5918], [ 0.2540, 0.6494, 0.6211, 0.6211, 0.6211, 0.6211, 0.6613, 0.2530], [ 0.0873, 0.0837, -0.3083, -0.0584, 0.1192, -0.1307, -0.1400, -0.1493], [ 0.1661, -0.0434, -0.2697, 0.1860, 0.1980, -0.2577, -0.1014, 0.0951], [ 0.2942, 0.2339, 0.0656, 0.3721, 0.3261, 0.0195, 0.2339, 0.2812], [ 0.2942, 0.2339, 0.0656, 0.3721, 0.3261, 0.0195, 0.2339, 0.2812], [ 0.2942, 0.2339, 0.0270, 0.0837, -0.1400, -0.1967, 0.2339, 0.2812], [ 0.2130, 0.0919, -0.0362, -0.1854, -0.2434, -0.0942, 0.0919, 0.1501]], [[-0.4264, -0.3058, -0.0771, -0.0771, -0.0771, -0.0771, 0.1542, 0.2827], [-0.5278, -0.4299, 0.0381, 0.0381, 0.0381, 0.0381, 0.4555, 0.6393], [-0.1677, 0.1134, 0.3660, 0.3549, 0.8149, 0.8260, 0.6546, 0.6552], [ 0.0183, 0.3910, 0.2182, 0.1155, 1.0009, 1.1036, 0.5068, 0.4159], [-0.1105, 0.0901, -0.1985, -0.1290, 0.8722, 0.8027, 0.0901, 0.1713], [-0.1105, 0.0901, -0.1985, -0.1290, 0.8722, 0.8027, 0.0901, 0.1713], [-0.1105, 0.0901, 0.0328, 0.1134, 0.6546, 0.5740, 0.0901, 0.1713], [-0.0422, 0.0939, 0.2227, 0.3948, 0.5107, 0.3385, 0.0939, 0.1101]], [[-0.0304, -0.1294, 0.0825, 0.0825, 0.0825, 0.0825, 0.1528, 0.3564], [-0.0499, -0.0802, 0.2031, 0.2031, 0.2031, 0.2031, 0.2914, 0.4559], [-0.2361, 0.1632, 0.0132, -0.1656, 0.1166, 0.2954, 0.3580, 0.0929], [-0.2182, 0.0699, -0.1696, -0.2371, 0.1346, 0.2021, 0.1752, 0.0214], [ 0.0384, 0.0756, -0.2692, -0.0859, 0.3911, 0.2077, 0.0756, 0.1726], [ 0.0384, 0.0756, -0.2692, -0.0859, 0.3911, 0.2077, 0.0756, 0.1726], [ 0.0384, 0.0756, -0.1989, 0.1632, 0.3580, -0.0041, 0.0756, 0.1726], [ 0.1085, 0.0860, -0.1705, 0.0804, 0.1857, -0.0651, 0.0860, 0.1633]]]], grad_fn=<ThnnConv2DBackward0>) torch.Size([1, 3, 8, 8])
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
我的输入是图片的灰度单通道，但经过卷积后可以输出3通道甚至更多。
再看输出的shape，图片宽高均为(8,8)，没有改变。
卷积的原理：
（图片我是从这里搬过来的，感谢作者！）
卷积核为(3,3)，每(3,3)的方块中通过卷积运算得到一个像素值。因为在卷积时使用了一个padding=1，所以在原图四周多出一个像素，也就导致卷积图和原图保持同一个大小。
卷积的原理看这儿。
神经网络中添加一个激活层
这里使用的激活函数为nn.ReLU()，先添加运行看看结果。
# 定义模型 class MyModel(nn.Module): def __init__(self): super(MyModel, self).__init__() self.conv0 = nn.Conv2d(in_channels=1, out_channels=3, kernel_size=(3, 3), stride=1, padding=1) self.relu0 = nn.ReLU(inplace=True) # 把模型里的卷积层和激活层后的效果，显示出来看一看 def forward(self, image): img = self.conv0(image) print("conv0:", img, img.shape, "\n") showTensorImg(img.squeeze(), "conv0") img = self.relu0(img) print("relu0:", img, img.shape, "\n") showTensorImg(img.squeeze(), "relu0") return img
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
再看看输出结果：
(8, 8) conv0: tensor([[[[-0.4327, -0.4364, -0.5815, -0.5815, -0.5815, -0.5815, -0.3790, -0.3967], [-0.1099, -0.2190, -0.2511, -0.2511, -0.2511, -0.2511, -0.2947, -0.2357], [ 0.2893, 0.2509, 0.2563, 0.2617, 0.3190, 0.3136, 0.0159, 0.0043], [ 0.0433, 0.1370, 0.2754, 0.1487, 0.0730, 0.1997, 0.0351, -0.1086], [-0.1535, -0.1358, 0.1045, 0.0538, -0.1238, -0.0732, -0.1358, -0.2035], [-0.1535, -0.1358, 0.1045, 0.0538, -0.1238, -0.0732, -0.1358, -0.2035], [-0.1535, -0.1358, 0.3070, 0.2509, 0.0159, 0.0720, -0.1358, -0.2035], [-0.0980, -0.0395, 0.1572, 0.2333, 0.1314, 0.0553, -0.0395, -0.1642]], [[-0.2516, -0.4649, -0.5853, -0.5853, -0.5853, -0.5853, -0.6295, -0.4837], [-0.1780, -0.2942, -0.4109, -0.4109, -0.4109, -0.4109, -0.5343, -0.4370], [-0.4042, -0.1781, -0.1395, -0.3199, -0.4845, -0.3040, -0.0809, -0.3259], [-0.4835, -0.3571, -0.2429, -0.3237, -0.5638, -0.4830, -0.1843, -0.3296], [-0.3014, -0.3369, -0.3955, -0.3145, -0.3818, -0.4628, -0.3369, -0.3204], [-0.3014, -0.3369, -0.3955, -0.3145, -0.3818, -0.4628, -0.3369, -0.3204], [-0.3014, -0.3369, -0.4396, -0.1781, -0.0809, -0.3423, -0.3369, -0.3204], [-0.2631, -0.2646, -0.4467, -0.2849, -0.1120, -0.2739, -0.2646, -0.2357]], [[ 0.5655, 0.6888, 0.4497, 0.4497, 0.4497, 0.4497, 0.3205, 0.0511], [ 0.6213, 0.8629, 0.5083, 0.5083, 0.5083, 0.5083, 0.3436, 0.0312], [ 0.3276, 0.5826, 0.6993, 0.9224, 0.5541, 0.3310, 0.6991, 0.3734], [ 0.2920, 0.4577, 0.7255, 1.0379, 0.5186, 0.2062, 0.7253, 0.4888], [ 0.4566, 0.3858, 0.3860, 0.9348, 0.6831, 0.1342, 0.3858, 0.3858], [ 0.4566, 0.3858, 0.3860, 0.9348, 0.6831, 0.1342, 0.3858, 0.3858], [ 0.4566, 0.3858, 0.2568, 0.5826, 0.6991, 0.3734, 0.3858, 0.3858], [ 0.3575, 0.3540, 0.1894, 0.4259, 0.6935, 0.4570, 0.3540, 0.3903]]]], grad_fn=<ThnnConv2DBackward0>) torch.Size([1, 3, 8, 8]) relu0: tensor([[[[0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.2893, 0.2509, 0.2563, 0.2617, 0.3190, 0.3136, 0.0159, 0.0043], [0.0433, 0.1370, 0.2754, 0.1487, 0.0730, 0.1997, 0.0351, 0.0000], [0.0000, 0.0000, 0.1045, 0.0538, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.1045, 0.0538, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.3070, 0.2509, 0.0159, 0.0720, 0.0000, 0.0000], [0.0000, 0.0000, 0.1572, 0.2333, 0.1314, 0.0553, 0.0000, 0.0000]], [[0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000]], [[0.5655, 0.6888, 0.4497, 0.4497, 0.4497, 0.4497, 0.3205, 0.0511], [0.6213, 0.8629, 0.5083, 0.5083, 0.5083, 0.5083, 0.3436, 0.0312], [0.3276, 0.5826, 0.6993, 0.9224, 0.5541, 0.3310, 0.6991, 0.3734], [0.2920, 0.4577, 0.7255, 1.0379, 0.5186, 0.2062, 0.7253, 0.4888], [0.4566, 0.3858, 0.3860, 0.9348, 0.6831, 0.1342, 0.3858, 0.3858], [0.4566, 0.3858, 0.3860, 0.9348, 0.6831, 0.1342, 0.3858, 0.3858], [0.4566, 0.3858, 0.2568, 0.5826, 0.6991, 0.3734, 0.3858, 0.3858], [0.3575, 0.3540, 0.1894, 0.4259, 0.6935, 0.4570, 0.3540, 0.3903]]]], grad_fn=<ReluBackward0>) torch.Size([1, 3, 8, 8])
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
- 49
- 50
- 51
- 52
- 53
- 54
- 55
- 56
- 57
从输出的结果可以看到，nn.ReLU()激活函数，将卷积的结果图片中，所有小于0的值被被置零了，并不会改变图片的大小。
因为值越大，特征就越是突出，那将小于0的值全部抹掉，可以减少模型的计算。
神经网络中添加一个池化层
这里池化层用的是最大池化，nn.MaxPool2d()，平均池化的话可以自己尝试一下。
# 定义模型 class MyModel(nn.Module): def __init__(self): super(MyModel, self).__init__() self.conv0 = nn.Conv2d(in_channels=1, out_channels=3, kernel_size=(3, 3), stride=1, padding=1) self.relu0 = nn.ReLU(inplace=True) self.pool0 = nn.MaxPool2d(kernel_size=2, stride=1) # 把模型里的卷积层和激活层后的效果，显示出来看一看 def forward(self, image): img = self.conv0(image) print("conv0:", img, img.shape, "\n") showTensorImg(img.squeeze(), "conv0") img = self.relu0(img) print("relu0:", img, img.shape, "\n") showTensorImg(img.squeeze(), "relu0") img = self.pool0(img) print("pool0:", img, img.shape, "\n") showTensorImg(img.squeeze(), "pool0") return img
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
输出结果：
D:\ProgramData\anaconda3\envs\Python3.8.5-Pytorch\python.exe E:/work/Python/Demo/demo.py (8, 8) conv0: tensor([[[[ 0.1296, 0.0942, 0.2273, 0.2273, 0.2273, 0.2273, -0.0275, -0.0733], [ 0.0647, 0.1555, 0.0734, 0.0734, 0.0734, 0.0734, -0.1328, -0.3786], [-0.1575, -0.1451, -0.3350, -0.3473, -0.4690, -0.4568, -0.5740, -0.7413], [-0.1089, -0.2873, -0.3107, -0.1321, -0.4204, -0.5990, -0.5496, -0.5262], [-0.1416, -0.2293, 0.0097, 0.0975, -0.4531, -0.5410, -0.2293, -0.2965], [-0.1416, -0.2293, 0.0097, 0.0975, -0.4531, -0.5410, -0.2293, -0.2965], [-0.1416, -0.2293, -0.2452, -0.1451, -0.5740, -0.6741, -0.2293, -0.2965], [-0.2099, -0.3350, -0.3023, -0.3930, -0.6554, -0.5647, -0.3350, -0.3305]], [[ 0.1986, 0.3047, 0.0893, 0.0893, 0.0893, 0.0893, 0.0789, -0.0405], [ 0.2143, 0.4924, 0.5072, 0.5072, 0.5072, 0.5072, 0.4453, 0.2312], [-0.0329, 0.2532, 0.5312, 0.6332, 0.4074, 0.3054, 0.7061, 0.4840], [-0.0843, 0.0403, 0.1396, 0.4030, 0.3559, 0.0925, 0.3145, 0.2538], [ 0.1525, 0.1671, -0.0078, 0.3656, 0.5927, 0.2193, 0.1671, 0.2164], [ 0.1525, 0.1671, -0.0078, 0.3656, 0.5927, 0.2193, 0.1671, 0.2164], [ 0.1525, 0.1671, -0.0182, 0.2532, 0.7061, 0.4347, 0.1671, 0.2164], [ 0.1208, 0.1961, -0.0406, 0.0693, 0.3435, 0.2335, 0.1961, 0.2483]], [[ 0.7197, 0.8873, 1.0323, 1.0323, 1.0323, 1.0323, 0.8345, 0.5481], [ 0.9719, 1.2985, 1.2925, 1.2925, 1.2925, 1.2925, 0.8764, 0.4069], [ 0.7647, 1.0524, 1.0036, 1.1729, 1.1201, 0.9508, 0.5970, 0.1663], [ 0.5465, 0.6483, 0.9689, 1.3240, 0.9019, 0.5468, 0.5622, 0.3174], [ 0.5559, 0.5273, 0.9340, 1.4195, 0.9113, 0.4258, 0.5273, 0.4129], [ 0.5559, 0.5273, 0.9340, 1.4195, 0.9113, 0.4258, 0.5273, 0.4129], [ 0.5559, 0.5273, 0.7361, 1.0524, 0.5970, 0.2807, 0.5273, 0.4129], [ 0.4526, 0.3832, 0.3738, 0.5042, 0.4181, 0.2877, 0.3832, 0.3244]]]], grad_fn=<ThnnConv2DBackward0>) torch.Size([1, 3, 8, 8]) relu0: tensor([[[[0.1296, 0.0942, 0.2273, 0.2273, 0.2273, 0.2273, 0.0000, 0.0000], [0.0647, 0.1555, 0.0734, 0.0734, 0.0734, 0.0734, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0097, 0.0975, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0097, 0.0975, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000]], [[0.1986, 0.3047, 0.0893, 0.0893, 0.0893, 0.0893, 0.0789, 0.0000], [0.2143, 0.4924, 0.5072, 0.5072, 0.5072, 0.5072, 0.4453, 0.2312], [0.0000, 0.2532, 0.5312, 0.6332, 0.4074, 0.3054, 0.7061, 0.4840], [0.0000, 0.0403, 0.1396, 0.4030, 0.3559, 0.0925, 0.3145, 0.2538], [0.1525, 0.1671, 0.0000, 0.3656, 0.5927, 0.2193, 0.1671, 0.2164], [0.1525, 0.1671, 0.0000, 0.3656, 0.5927, 0.2193, 0.1671, 0.2164], [0.1525, 0.1671, 0.0000, 0.2532, 0.7061, 0.4347, 0.1671, 0.2164], [0.1208, 0.1961, 0.0000, 0.0693, 0.3435, 0.2335, 0.1961, 0.2483]], [[0.7197, 0.8873, 1.0323, 1.0323, 1.0323, 1.0323, 0.8345, 0.5481], [0.9719, 1.2985, 1.2925, 1.2925, 1.2925, 1.2925, 0.8764, 0.4069], [0.7647, 1.0524, 1.0036, 1.1729, 1.1201, 0.9508, 0.5970, 0.1663], [0.5465, 0.6483, 0.9689, 1.3240, 0.9019, 0.5468, 0.5622, 0.3174], [0.5559, 0.5273, 0.9340, 1.4195, 0.9113, 0.4258, 0.5273, 0.4129], [0.5559, 0.5273, 0.9340, 1.4195, 0.9113, 0.4258, 0.5273, 0.4129], [0.5559, 0.5273, 0.7361, 1.0524, 0.5970, 0.2807, 0.5273, 0.4129], [0.4526, 0.3832, 0.3738, 0.5042, 0.4181, 0.2877, 0.3832, 0.3244]]]], grad_fn=<ReluBackward0>) torch.Size([1, 3, 8, 8]) pool0: tensor([[[[0.1555, 0.2273, 0.2273, 0.2273, 0.2273, 0.2273, 0.0000], [0.1555, 0.1555, 0.0734, 0.0734, 0.0734, 0.0734, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000], [0.0000, 0.0097, 0.0975, 0.0975, 0.0000, 0.0000, 0.0000], [0.0000, 0.0097, 0.0975, 0.0975, 0.0000, 0.0000, 0.0000], [0.0000, 0.0097, 0.0975, 0.0975, 0.0000, 0.0000, 0.0000], [0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000]], [[0.4924, 0.5072, 0.5072, 0.5072, 0.5072, 0.5072, 0.4453], [0.4924, 0.5312, 0.6332, 0.6332, 0.5072, 0.7061, 0.7061], [0.2532, 0.5312, 0.6332, 0.6332, 0.4074, 0.7061, 0.7061], [0.1671, 0.1671, 0.4030, 0.5927, 0.5927, 0.3145, 0.3145], [0.1671, 0.1671, 0.3656, 0.5927, 0.5927, 0.2193, 0.2164], [0.1671, 0.1671, 0.3656, 0.7061, 0.7061, 0.4347, 0.2164], [0.1961, 0.1961, 0.2532, 0.7061, 0.7061, 0.4347, 0.2483]], [[1.2985, 1.2985, 1.2925, 1.2925, 1.2925, 1.2925, 0.8764], [1.2985, 1.2985, 1.2925, 1.2925, 1.2925, 1.2925, 0.8764], [1.0524, 1.0524, 1.3240, 1.3240, 1.1201, 0.9508, 0.5970], [0.6483, 0.9689, 1.4195, 1.4195, 0.9113, 0.5622, 0.5622], [0.5559, 0.9340, 1.4195, 1.4195, 0.9113, 0.5273, 0.5273], [0.5559, 0.9340, 1.4195, 1.4195, 0.9113, 0.5273, 0.5273], [0.5559, 0.7361, 1.0524, 1.0524, 0.5970, 0.5273, 0.5273]]]], grad_fn=<MaxPool2DWithIndicesBackward0>) torch.Size([1, 3, 7, 7])
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
- 49
- 50
- 51
- 52
- 53
- 54
- 55
- 56
- 57
- 58
- 59
- 60
- 61
- 62
- 63
- 64
- 65
- 66
- 67
- 68
- 69
- 70
- 71
- 72
- 73
- 74
- 75
- 76
- 77
- 78
- 79
- 80
- 81
- 82
可以看到，经过池化层之后原本(8,8)大小的图片，变成了(7,7)。
池化层到底是什么，是为了在一定区域内提取出该区域的关键性信息，起到减少卷积层输出特征量数目的作用，从而能减少模型参数，同时能改善过拟合现象。
虽然图片大小变小了，但关键信息却被提取了出来，还能减少模型计算量。
池化层的原理，例如我的池化层设定是nn.MaxPool2d(kernel_size=2, stride=1)，那么池化层的窗口就为(2,2)，以(2,2)的窗口，在图片中从左往右从上往下移动，每移动步长stride为1。在每个(2,2)的窗口中，寻找其最大值，作为新图片的像素值。
（图片来源在这里，感谢作者！）
全部代码
from torchvision.transforms import functional as F import matplotlib.pyplot as plt import torch.nn as nn import cv2 # 将图片转换为模型可用的数据 def getImageData(imgFile): image = cv2.imread(imgFile, 0) image = F.to_tensor(image) # shape为 [1, 768, 1024], [通道，高，宽] image = image.unsqueeze(0) # shape为 [1, 1, 768, 1024], [数量，通道，高，宽] return image # 将三个图片显示在plt上 def showTensorImg(image, title=""): plt.figure(title) # 创建第一个图形 plt.sca(plt.subplot(1, 3, 1)) # 分成1行2列，选择第1个位置 plt.imshow(image[0].data, cmap=plt.cm.gray) plt.sca(plt.subplot(1, 3, 2)) # 分成1行2列，选择第2个位置 plt.imshow(image[1].data, cmap=plt.cm.gray) plt.sca(plt.subplot(1, 3, 3)) # 分成1行2列，选择第2个位置 plt.imshow(image[2].data, cmap=plt.cm.gray) # 定义模型 class MyModel(nn.Module): def __init__(self): super(MyModel, self).__init__() self.conv0 = nn.Conv2d(in_channels=1, out_channels=3, kernel_size=(3, 3), stride=1, padding=1) self.relu0 = nn.ReLU(inplace=True) self.pool0 = nn.MaxPool2d(kernel_size=2, stride=1) # 把模型里的卷积层和激活层后的效果，显示出来看一看 def forward(self, image): img = self.conv0(image) print("conv0:", img, img.shape, "\n") showTensorImg(img.squeeze(), "conv0") img = self.relu0(img) print("relu0:", img, img.shape, "\n") showTensorImg(img.squeeze(), "relu0") img = self.pool0(img) print("pool0:", img, img.shape, "\n") showTensorImg(img.squeeze(), "pool0") return img if __name__ == "__main__": # 显示原图 img = cv2.imread("Data/T.bmp", 0) plt.imshow(img) print(img.shape) net = MyModel() image = getImageData("Data/T.bmp") img = net(image) plt.show()
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
- 49
- 50
- 51
- 52
- 53
- 54
- 55
- 56
- 57
- 58
- 59
- 60
- 61
- 62
声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/你好赵伟/article/detail/75064
推荐阅读
相关标签
