九、Pytorch 卷积和池化实现

刺骨的言语ヽ痛彻心扉 2023-06-13 12:30 28阅读 0赞

# 1、卷积 #

**在pytorch中有两种形式，一种是 torch.nn.Conv2d() ，另一种是 torch.nn.functional.conv2d()**

![20191118133334878.png][]

![20191117161024208.png][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3MyOTQ4NzgzMDQ_size_16_color_FFFFFF_t_70][]

# encoding: utf-8
    
    import numpy as np
    import torch
    from torch import nn
    from torch.autograd import Variable
    import torch.nn.functional as F
    from PIL import  Image
    import  matplotlib.pyplot as plt
    
    def show_img(arr):
        plt.imshow(arr.astype('uint8'),cmap='gray')
        plt.show()
    
    if __name__ == '__main__':
        # PIL读取图片并处理
        img = Image.open('lena.jpg')
        gray = img.convert('L')
        arr = np.array(gray,dtype='float32')
        show_img(arr)
        #reshape成 batch_size,input_channel,图像高度，图像宽度,此处为一张单通道图片
        tensor = torch.from_numpy(arr.reshape((1,1,arr.shape[0],arr.shape[1])))
        print(tensor.shape)
        """=========== 1 ============="""
        #定义卷积
        conv1 = nn.Conv2d(1,1,1,bias=True)
        # 定义轮廓检测算子
        sobel_kernel = np.array([[-1,	-1,	-1],
                                 [-1,	8,	-1],
                                 [-1,	-1,	-1]],	dtype='float32')
        # sobel_kernel.dtype=np.float32
        print(sobel_kernel.shape)
        sobel_kernel = sobel_kernel.reshape((1,1,3,3))
        # 给卷积的kernel赋值
        conv1.weight.data = torch.from_numpy(sobel_kernel)
        # 作用在图片上
        edge1 = conv1(Variable(tensor))
        # 将输出转换为图片的格式
        edge1 = edge1.data.squeeze().numpy()
        show_img(edge1)
        """=========== 2 ============="""
        weight = Variable(torch.from_numpy(sobel_kernel))
        edge2 = F.conv2d((Variable(tensor),weight))
        edge2 = edge2.data.squeeze().numpy()
        show_img(edge2)

# 2、池化 #

**在pytorch中有两种形式，一种是** nn.MaxPool2d() **，另一种是** torch.nn.functional.max\_pool2d()

![20191118133538745.png][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3MyOTQ4NzgzMDQ_size_16_color_FFFFFF_t_70 1][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3MyOTQ4NzgzMDQ_size_16_color_FFFFFF_t_70 2][]

# encoding: utf-8
    
    
    import numpy as np
    import torch
    from torch import nn
    from torch.autograd import Variable
    import torch.nn.functional as F
    from PIL import  Image
    import  matplotlib.pyplot as plt
    
    def show_img(arr):
        plt.imshow(arr.astype('uint8'),cmap='gray')
        plt.show()
    
    if __name__ == '__main__':
        # PIL读取图片并处理
        img = Image.open('lena.jpg')
        gray = img.convert('L')
        arr = np.array(gray,dtype='float32')
        #reshape成 batch_size,input_channel,图像高度，图像宽度,此处为一张单通道图片
        tensor = torch.from_numpy(arr.reshape((1,1,arr.shape[0],arr.shape[1])))
        print(tensor.shape)
        """=========== 1 ============="""
        pool1 = nn.MaxPool2d(2,2)
        print('before	max	pool,	image	shape:	{}	x	{}'
              .format(tensor.shape[2], tensor.shape[3]))
        small_im1 = pool1(Variable(tensor))
        small_im1 = small_im1.data.squeeze().numpy()
        print('after	max	pool,	image	shape:	{}	x	{}	'
              .format(small_im1.shape[0],small_im1.shape[1]))
        show_img(arr)
        show_img(small_im1)
        """=========== 2 ============="""
        print('before	max	pool,	image	shape:	{}	x	{}'
              .format(tensor.shape[2], tensor.shape[3]))
        small_im2 = F.max_pool2d(Variable(tensor), 2, 2)
        small_im2 = small_im2.data.squeeze().numpy()
        print('after	max	pool,	image	shape:	{}	x	{}	'
              .format(small_im1.shape[0],small_im1.shape[1]))
        show_img(small_im2)

[20191118133334878.png]: https://img-blog.csdnimg.cn/20191118133334878.png
[20191117161024208.png]: https://img-blog.csdnimg.cn/20191117161024208.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3MyOTQ4NzgzMDQ_size_16_color_FFFFFF_t_70]: https://img-blog.csdnimg.cn/20191117161047311.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3MyOTQ4NzgzMDQ=,size_16,color_FFFFFF,t_70
[20191118133538745.png]: https://img-blog.csdnimg.cn/20191118133538745.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3MyOTQ4NzgzMDQ_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/20191117161621704.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3MyOTQ4NzgzMDQ=,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3MyOTQ4NzgzMDQ_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/20191117161648883.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3MyOTQ4NzgzMDQ=,size_16,color_FFFFFF,t_70