用pytorch搭建简单CNN前馈网络处理三通道彩色图像

﹏ヽ暗。殇╰゛Y 2022-12-10 08:48 157阅读 0赞

最近想重新学习简单CNN网络前向传播的搭建，发现网上好像没有用pytorch对三通道彩色图像进行处理的。

## 1.自定义初始化4\*4的三个卷积核，然后经过各一次的relu激活和池化，pytorch中，处理图片必须一个batch一个batch的操作，所以我们要准备的数据的格式是 \[batch\_size, n\_channels, hight, width\]，记录一下代码： ##

import cv2
    import matplotlib.pyplot as plt
    
    
    img_path = '1.jpg'
    
    bgr_img = cv2.imread(img_path)
    
    # bgr_img = bgr_img.transpose(2,0,1)
    print(bgr_img.shape)
    b,g,r = cv2.split(bgr_img)
    img_rgb = cv2.merge([r,g,b])
    gray_img = cv2.cvtColor(img_rgb, cv2.COLOR_BGR2GRAY)
    img_rgb1 = img_rgb.transpose(2,0,1)#将通道数放在最前面的位置
    print(img_rgb.shape)
    
    
    plt.figure()
    plt.imshow(gray_img)
    print(gray_img.shape)
    # Normalise
    gray_img = gray_img.astype("float32")/255
    # print(gray_img.shape)
    plt.figure()
    plt.imshow(gray_img)
    plt.figure()
    plt.imshow(img_rgb)
    
    # plt.show()
    
    import numpy as np
    
    filter_vals = np.array([[-1, -1, 1, 1], [-1, -1, 1, 1], [-1, -1, 1, 1], [-1, -1, 1, 1]],dtype=np.float64)
    filter_vals=filter_vals
    print('Filter shape: ', filter_vals.shape)
    filter_1 = filter_vals
    filter_2 = -filter_1
    filter_3 = filter_1.T
    # filter_4 = -filter_3
    filters = np.array([filter_1, filter_2, filter_3])#, filter_4])
    import torch
    import torch.nn as nn
    import torch.nn.functional as F
    
    
    class Net(nn.Module):
    
        def __init__(self,weight):
            super(Net, self).__init__()
            # initializes the weights of the convolutional layer to be the weights of the 4 defined filters
            # k_height, k_width = weight.shape[2:]
            # assumes there are 4 grayscale filters
            self.conv = nn.Conv2d(111,111,4, bias=False)
            self.conv.weight = torch.nn.Parameter(weight)
            # define a pooling layer
            self.pool = nn.MaxPool2d(2, 2)
    
        def forward(self, x):
            # calculates the output of a convolutional layer
            # pre- and post-activation
            conv_x = self.conv(x)
            activated_x = F.relu(conv_x)
            print('activated_x-shape',activated_x.shape)
            # applies pooling layer
            pooled_x = self.pool(activated_x)
    
            # returns all layers
            return conv_x, activated_x, pooled_x
    
    
    # instantiate the model and set the weights
    weight = torch.from_numpy(filters).unsqueeze(0)#.type(torch.FloatTensor)
    print("weight shape",weight.shape)
    model = Net(weight)
    
    # print out the layer in the network
    print(model)
    
    
    def viz_layer(layer, n_filters=1):
        fig = plt.figure(figsize=(20, 20))
        for i in range(n_filters):
            ax = fig.add_subplot(1, n_filters, i + 1)
            ax.imshow(np.squeeze(layer[0, i].data.numpy()))
            ax.set_title('Output %s' % str(i + 1))
    
    
    # plt.imshow(gray_img, cmap='gray')
    
    fig = plt.figure(figsize=(12, 6))
    fig.subplots_adjust(left=0, right=1.5, bottom=0.8, top=1, hspace=0.05, wspace=0.05)
    for i in range(3):
        ax = fig.add_subplot(1, 3, i + 1, xticks=[], yticks=[])
        ax.imshow(filters[i], cmap='gray')
        ax.set_title('Filter %s' % str(i + 1))
    
    gray_img_tensor = torch.from_numpy(img_rgb1).unsqueeze(0)#.unsqueeze(1)
    gray_img_tensor=gray_img_tensor.type('torch.DoubleTensor')
    
    print('gray_img_tensor.shape',gray_img_tensor.shape)
    conv_x, activated_layer, pooled_layer = model(gray_img_tensor)
    
    viz_layer(activated_layer)
    #
    # viz_layer(pooled_layer)

## 2.如下系原版代码，对单通道进行卷积（滤波）处理： ##

import cv2
    import matplotlib.pyplot as plt
    
    
    img_path = '1.jpg'
    
    bgr_img = cv2.imread(img_path)
    print(bgr_img.shape)
    b,g,r = cv2.split(bgr_img)
    img_rgb = cv2.merge([r,g,b])
    print(img_rgb.shape)
    gray_img = cv2.cvtColor(img_rgb, cv2.COLOR_BGR2GRAY)
    print(gray_img.shape)
    # Normalise
    gray_img = gray_img.astype("float32")/255
    # print(gray_img.shape)
    pass
    # plt.imshow(gray_img, cmap='gray')
    plt.show()
    
    import numpy as np
    
    filter_vals = np.array([[-1, -1, 1, 1], [-1, -1, 1, 1], [-1, -1, 1, 1], [-1, -1, 1, 1]])
    
    print('Filter shape: ', filter_vals.shape)
    filter_1 = filter_vals
    filter_2 = -filter_1
    filter_3 = filter_1.T
    filter_4 = -filter_3
    filters = np.array([filter_1, filter_2, filter_3, filter_4])
    import torch
    import torch.nn as nn
    import torch.nn.functional as F
    
    
    class Net(nn.Module):
    
        def __init__(self,weight):
            super(Net, self).__init__()
            # initializes the weights of the convolutional layer to be the weights of the 4 defined filters
            # k_height, k_width = weight.shape[2:]
            # assumes there are 4 grayscale filters
            self.conv = nn.Conv2d(3, 3, 4, bias=False)
            self.conv.weight = torch.nn.Parameter(weight)
            # define a pooling layer
            self.pool = nn.MaxPool2d(2, 2)
    
        def forward(self, x):
            # calculates the output of a convolutional layer
            # pre- and post-activation
            conv_x = self.conv(x)
            activated_x = F.relu(conv_x)
    
            # applies pooling layer
            pooled_x = self.pool(activated_x)
    
            # returns all layers
            return conv_x, activated_x, pooled_x
    
    
    # instantiate the model and set the weights
    weight = torch.from_numpy(filters).unsqueeze(1).type(torch.FloatTensor)
    model = Net(weight)
    
    # print out the layer in the network
    print(model)
    
    
    def viz_layer(layer, n_filters=4):
        fig = plt.figure(figsize=(20, 20))
    
        for i in range(n_filters):
            ax = fig.add_subplot(1, n_filters, i + 1)
            ax.imshow(np.squeeze(layer[0, i].data.numpy()), cmap='gray')
            ax.set_title('Output %s' % str(i + 1))
    
    
    # plt.imshow(gray_img, cmap='gray')
    
    fig = plt.figure(figsize=(12, 6))
    fig.subplots_adjust(left=0, right=1.5, bottom=0.8, top=1, hspace=0.05, wspace=0.05)
    for i in range(4):
        ax = fig.add_subplot(1, 4, i + 1, xticks=[], yticks=[])
        ax.imshow(filters[i], cmap='gray')
        ax.set_title('Filter %s' % str(i + 1))
    
    r = torch.from_numpy(gray_img).unsqueeze(0).unsqueeze(1)
    print("gray_img_tensor.shape",gray_img_tensor.shape)
    conv_x, activated_layer, pooled_layer = model(r)#r可替换为gray_img_tensor
    
    viz_layer(activated_layer)
    
    viz_layer(pooled_layer)

## 3.使用torch默认初始化的卷积核进行三次“卷积激活池化”操作，因为卷积核每次都随机初始化，故每次得出的图像结果都不同 ##

import cv2
    import matplotlib.pyplot as plt
    import torchvision.models as models#就这一句，即可初始化卷积核
    #或者这个import torch.optim as optim
    
    
    img_path = '1.jpg'
    
    bgr_img = cv2.imread(img_path)
    
    # bgr_img = bgr_img.transpose(2,0,1)
    print(bgr_img.shape)
    b,g,r = cv2.split(bgr_img)
    img_rgb = cv2.merge([r,g,b])
    gray_img = cv2.cvtColor(img_rgb, cv2.COLOR_BGR2GRAY)
    img_rgb1 = img_rgb.transpose(2,0,1)#将通道数放在最前面的位置
    print(img_rgb.shape)
    
    
    plt.figure()
    plt.imshow(b)
    print(gray_img.shape)
    # Normalise
    gray_img = gray_img.astype("float32")/255
    # print(gray_img.shape)
    plt.figure()
    plt.imshow(img_rgb)
    plt.figure()
    plt.imshow(b)
    
    # plt.show()
    
    import numpy as np
    
    filter_vals = np.array([[-1, -1, 1, 1], [-1, -1, 1, 1], [-1, -1, 1, 1], [-1, -1, 1, 1]],dtype=np.float64)
    filter_vals=filter_vals
    print('Filter shape: ', filter_vals.shape)
    filter_1 = filter_vals
    filter_2 = -filter_1
    filter_3 = filter_1.T
    # filter_4 = -filter_3
    filters = np.array([filter_1, filter_2, filter_3])#, filter_4])
    import torch
    import torch.nn as nn
    import torch.nn.functional as F
    
    
    class Net(nn.Module):
    
        def __init__(self):
            super(Net, self).__init__()
            # initializes the weights of the convolutional layer to be the weights of the 4 defined filters
            # k_height, k_width = weight.shape[2:]
            # assumes there are 4 grayscale filters
            self.conv1 = nn.Conv2d(3,3,4, bias=False)
            # self.conv1 = torch.nn.Parameter(weight)
            # define a pooling layer
            self.pool1 = nn.MaxPool2d(2, 2)
            self.conv2 = nn.Conv2d(3,3,4, bias=False)
            # self.conv2.weight = torch.nn.Parameter(weight)
            # define a pooling layer
            self.pool2 = nn.MaxPool2d(2, 2)
            self.conv3 = nn.Conv2d(3,3,4, bias=False)
            # self.conv3.weight = torch.nn.Parameter(weight)
            # define a pooling layer
            self.pool3 = nn.MaxPool2d(2, 2)
    
        def forward(self, x):
            # calculates the output of a convolutional layer
            # pre- and post-activation
            conv_x = self.conv1(x)
            activated_x = F.relu(conv_x)
            print('activated_x-shape',activated_x.shape)
            # applies pooling layer
            pooled_x = self.pool1(activated_x)
            conv_x = self.conv2(pooled_x)
            activated_x = F.relu(conv_x)
            print('activated_x-shape',activated_x.shape)
            # applies pooling layer
            pooled_x = self.pool2(activated_x)
            conv_x = self.conv3(pooled_x)
            activated_x = F.relu(conv_x)
            print('activated_x-shape',activated_x.shape)
            # applies pooling layer
            pooled_x = self.pool3(activated_x)
            # returns all layers
            return conv_x, activated_x, pooled_x
    
    
    # instantiate the model and set the weights
    # weight = torch.from_numpy(filters).unsqueeze(0)#.type(torch.FloatTensor)
    # print("weight shape",weight.shape)
    model = Net()
    
    # print out the layer in the network
    print(model)
    
    
    def viz_layer(layer, n_filters=1):
        fig = plt.figure(figsize=(20, 20))
        for i in range(n_filters):
            ax = fig.add_subplot(1, n_filters, i + 1)
            ax.imshow(np.squeeze(layer[0, i].data.numpy()))
            ax.set_title('Output %s' % str(i + 1))
    
    
    # plt.imshow(gray_img, cmap='gray')
    
    fig = plt.figure(figsize=(12, 6))
    fig.subplots_adjust(left=0, right=1.5, bottom=0.8, top=1, hspace=0.05, wspace=0.05)
    for i in range(3):
        ax = fig.add_subplot(1, 3, i + 1, xticks=[], yticks=[])
        ax.imshow(filters[i], cmap='gray')
        ax.set_title('Filter %s' % str(i + 1))
    
    gray_img_tensor = torch.from_numpy(img_rgb1).unsqueeze(0)#.unsqueeze(1)
    gray_img_tensor=torch.tensor(gray_img_tensor,dtype=torch.float32)
    print('gray_img_tensor.shape',gray_img_tensor.shape)
    conv_x, activated_layer, pooled_layer = model(gray_img_tensor)

## 4.加入全连接层（未添加softmax），注意需要有一个"展平”的操作。 ##

import cv2
    import matplotlib.pyplot as plt
    import torchvision.models as models#就这一句，即可初始化卷积核
    #或者这个import torch.optim as optim
    
    
    img_path = '1.jpg'
    
    bgr_img = cv2.imread(img_path)
    
    # bgr_img = bgr_img.transpose(2,0,1)
    print(bgr_img.shape)
    b,g,r = cv2.split(bgr_img)
    img_rgb = cv2.merge([r,g,b])
    gray_img = cv2.cvtColor(img_rgb, cv2.COLOR_BGR2GRAY)
    img_rgb1 = img_rgb.transpose(2,0,1)#将通道数放在最前面的位置
    print(img_rgb.shape)
    
    
    plt.figure()
    plt.imshow(b)
    print(gray_img.shape)
    # Normalise
    gray_img = gray_img.astype("float32")/255
    # print(gray_img.shape)
    plt.figure()
    plt.imshow(img_rgb)
    plt.figure()
    plt.imshow(b)
    
    # plt.show()
    
    import numpy as np
    
    filter_vals = np.array([[-1, -1, 1, 1], [-1, -1, 1, 1], [-1, -1, 1, 1], [-1, -1, 1, 1]],dtype=np.float64)
    filter_vals=filter_vals
    print('Filter shape: ', filter_vals.shape)
    filter_1 = filter_vals
    filter_2 = -filter_1
    filter_3 = filter_1.T
    # filter_4 = -filter_3
    filters = np.array([filter_1, filter_2, filter_3])#, filter_4])
    import torch
    import torch.nn as nn
    import torch.nn.functional as F
    
    
    class Net(nn.Module):
    
        def __init__(self):
            super(Net, self).__init__()
            # initializes the weights of the convolutional layer to be the weights of the 4 defined filters
            # k_height, k_width = weight.shape[2:]
            # assumes there are 4 grayscale filters
            self.conv1 = nn.Conv2d(3,3,4, bias=False)
            # self.conv1 = torch.nn.Parameter(weight)
            # define a pooling layer
            self.pool1 = nn.MaxPool2d(2, 2)
            self.conv2 = nn.Conv2d(3,3,4, bias=False)
            # self.conv2.weight = torch.nn.Parameter(weight)
            # define a pooling layer
            self.pool2 = nn.MaxPool2d(2, 2)
            self.conv3 = nn.Conv2d(3,3,4, bias=False)
            # self.conv3.weight = torch.nn.Parameter(weight)
            # define a pooling layer
            self.pool3 = nn.MaxPool2d(2, 2)
            self.fc1 = nn.Linear(3*73*69,8)
    
    
        def forward(self, x):
            # calculates the output of a convolutional layer
            # pre- and post-activation
            conv_x = self.conv1(x)
            activated_x = F.relu(conv_x)
            print('activated_x-shape',activated_x.shape)
            # applies pooling layer
            pooled_x = self.pool1(activated_x)
            conv_x = self.conv2(pooled_x)
            activated_x = F.relu(conv_x)
            print('activated_x-shape',activated_x.shape)
            # applies pooling layer
            pooled_x = self.pool2(activated_x)
            conv_x = self.conv3(pooled_x)
            activated_x = F.relu(conv_x)
            print('activated_x-shape',activated_x.shape)
            # applies pooling layer
            pooled_x = self.pool3(activated_x)
            print("shape",pooled_x.shape)
            pooled_x = pooled_x.view(1, -1)
            fc_x = self.fc1(pooled_x)
    
            # returns all layers
            return conv_x, activated_x, fc_x
    
    
    # instantiate the model and set the weights
    # weight = torch.from_numpy(filters).unsqueeze(0)#.type(torch.FloatTensor)
    # print("weight shape",weight.shape)
    model = Net()
    
    # print out the layer in the network
    print(model)
    
    
    def viz_layer(layer, n_filters=1):
        fig = plt.figure(figsize=(20, 20))
        for i in range(n_filters):
            ax = fig.add_subplot(1, n_filters, i + 1)
            ax.imshow(np.squeeze(layer[0, i].data.numpy()))
            ax.set_title('Output %s' % str(i + 1))
    
    
    # plt.imshow(gray_img, cmap='gray')
    
    fig = plt.figure(figsize=(12, 6))
    fig.subplots_adjust(left=0, right=1.5, bottom=0.8, top=1, hspace=0.05, wspace=0.05)
    for i in range(3):
        ax = fig.add_subplot(1, 3, i + 1, xticks=[], yticks=[])
        ax.imshow(filters[i], cmap='gray')
        ax.set_title('Filter %s' % str(i + 1))
    
    img_tensor = torch.from_numpy(img_rgb1).unsqueeze(0)#.unsqueeze(1)
    img_tensor=torch.tensor(img_tensor, dtype=torch.float32)
    print('gray_img_tensor.shape', img_tensor.shape)
    conv_x, activated_layer, pooled_layer = model(img_tensor)