def get_picture(picture_dir, transform):
    '''
    該算法實(shí)現(xiàn)了讀取圖片，并將其類型轉(zhuǎn)化為Tensor
    '''
    tmp = []
    img = skimage.io.imread(picture_dir)
    tmp.append(img)
    img = skimage.io.imread('./picture/4.jpg')
    tmp.append(img)
    img256 = [skimage.transform.resize(img, (256, 256)) for img in tmp]
    img256 = np.asarray(img256)
    img256 = img256.astype(np.float32)

    return transform(img256[0])

注意：神經(jīng)網(wǎng)絡(luò)的輸入是四維形式，我們返回的圖片是三維形式，需要使用unsqueeze()插入一個(gè)維度

神經(jīng)網(wǎng)絡(luò)的構(gòu)建

網(wǎng)絡(luò)的基于LeNet構(gòu)建，不過(guò)為了方便展示，將其中的參數(shù)按照2562563進(jìn)行的參數(shù)的修正

網(wǎng)絡(luò)構(gòu)建如下：

class LeNet(nn.Module):
    '''
    該類繼承了torch.nn.Modul類
    構(gòu)建LeNet神經(jīng)網(wǎng)絡(luò)模型
    '''
    def __init__(self):
        super(LeNet, self).__init__()

        # 第一層神經(jīng)網(wǎng)絡(luò)，包括卷積層、線性激活函數(shù)、池化層
        self.conv1 = nn.Sequential( 
            nn.Conv2d(3, 32, 5, 1, 2),   # input_size=(3*256*256)，padding=2
            nn.ReLU(),                  # input_size=(32*256*256)
            nn.MaxPool2d(kernel_size=2, stride=2),  # output_size=(32*128*128)
        )

        # 第二層神經(jīng)網(wǎng)絡(luò)，包括卷積層、線性激活函數(shù)、池化層
        self.conv2 = nn.Sequential(
            nn.Conv2d(32, 64, 5, 1, 2),  # input_size=(32*128*128)
            nn.ReLU(),            # input_size=(64*128*128)
            nn.MaxPool2d(2, 2)    # output_size=(64*64*64)
        )

        # 全連接層(將神經(jīng)網(wǎng)絡(luò)的神經(jīng)元的多維輸出轉(zhuǎn)化為一維)
        self.fc1 = nn.Sequential(
            nn.Linear(64 * 64 * 64, 128),  # 進(jìn)行線性變換
            nn.ReLU()                    # 進(jìn)行ReLu激活
        )

        # 輸出層(將全連接層的一維輸出進(jìn)行處理)
        self.fc2 = nn.Sequential(
            nn.Linear(128, 84),
            nn.ReLU()
        )

        # 將輸出層的數(shù)據(jù)進(jìn)行分類(輸出預(yù)測(cè)值)
        self.fc3 = nn.Linear(84, 62)

    # 定義前向傳播過(guò)程，輸入為x
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        # nn.Linear()的輸入輸出都是維度為一的值，所以要把多維度的tensor展平成一維
        x = x.view(x.size()[0], -1)
        x = self.fc1(x)
        x = self.fc2(x)
        x = self.fc3(x)
        return x

特征圖的提取

直接上代碼：

class FeatureExtractor(nn.Module):
    def __init__(self, submodule, extracted_layers):
        super(FeatureExtractor, self).__init__()
        self.submodule = submodule
        self.extracted_layers = extracted_layers
 
    def forward(self, x):
        outputs = []
        for name, module in self.submodule._modules.items():
        # 目前不展示全連接層
            if "fc" in name: 
                x = x.view(x.size(0), -1)
            print(module)
            x = module(x)
            print(name)
            if name in self.extracted_layers:
                outputs.append(x)
        return outputs

可視化展示

可視化展示使用matplotlib

代碼如下：

    # 特征輸出可視化
    for i in range(32):
        ax = plt.subplot(6, 6, i + 1)
        ax.set_title('Feature {}'.format(i))
        ax.axis('off')
        plt.imshow(x[0].data.numpy()[0,i,:,:],cmap='jet')
    plt.plot()

完整代碼

在此貼上完整代碼

import os
import torch
import torchvision as tv
import torchvision.transforms as transforms
import torch.nn as nn
import torch.optim as optim
import argparse
import skimage.data
import skimage.io
import skimage.transform
import numpy as np
import matplotlib.pyplot as plt

# 定義是否使用GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load training and testing datasets.
pic_dir = './picture/3.jpg'

# 定義數(shù)據(jù)預(yù)處理方式(將輸入的類似numpy中arrary形式的數(shù)據(jù)轉(zhuǎn)化為pytorch中的張量（tensor）)
transform = transforms.ToTensor()


def get_picture(picture_dir, transform):
    '''
    該算法實(shí)現(xiàn)了讀取圖片，并將其類型轉(zhuǎn)化為Tensor
    '''
    img = skimage.io.imread(picture_dir)
    img256 = skimage.transform.resize(img, (256, 256))
    img256 = np.asarray(img256)
    img256 = img256.astype(np.float32)

    return transform(img256)


def get_picture_rgb(picture_dir):
    '''
    該函數(shù)實(shí)現(xiàn)了顯示圖片的RGB三通道顏色
    '''
    img = skimage.io.imread(picture_dir)
    img256 = skimage.transform.resize(img, (256, 256))
    skimage.io.imsave('./picture/4.jpg',img256)

    # 取單一通道值顯示
    # for i in range(3):
    #     img = img256[:,:,i]
    #     ax = plt.subplot(1, 3, i + 1)
    #     ax.set_title('Feature {}'.format(i))
    #     ax.axis('off')
    #     plt.imshow(img)

    # r = img256.copy()
    # r[:,:,0:2]=0
    # ax = plt.subplot(1, 4, 1)
    # ax.set_title('B Channel')
    # # ax.axis('off')
    # plt.imshow(r)

    # g = img256.copy()
    # g[:,:,0]=0
    # g[:,:,2]=0
    # ax = plt.subplot(1, 4, 2)
    # ax.set_title('G Channel')
    # # ax.axis('off')
    # plt.imshow(g)

    # b = img256.copy()
    # b[:,:,1:3]=0
    # ax = plt.subplot(1, 4, 3)
    # ax.set_title('R Channel')
    # # ax.axis('off')
    # plt.imshow(b)

    # img = img256.copy()
    # ax = plt.subplot(1, 4, 4)
    # ax.set_title('image')
    # # ax.axis('off')
    # plt.imshow(img)

    img = img256.copy()
    ax = plt.subplot()
    ax.set_title('image')
    # ax.axis('off')
    plt.imshow(img)

    plt.show()


class LeNet(nn.Module):
    '''
    該類繼承了torch.nn.Modul類
    構(gòu)建LeNet神經(jīng)網(wǎng)絡(luò)模型
    '''
    def __init__(self):
        super(LeNet, self).__init__()

        # 第一層神經(jīng)網(wǎng)絡(luò)，包括卷積層、線性激活函數(shù)、池化層
        self.conv1 = nn.Sequential( 
            nn.Conv2d(3, 32, 5, 1, 2),   # input_size=(3*256*256)，padding=2
            nn.ReLU(),                  # input_size=(32*256*256)
            nn.MaxPool2d(kernel_size=2, stride=2),  # output_size=(32*128*128)
        )

        # 第二層神經(jīng)網(wǎng)絡(luò)，包括卷積層、線性激活函數(shù)、池化層
        self.conv2 = nn.Sequential(
            nn.Conv2d(32, 64, 5, 1, 2),  # input_size=(32*128*128)
            nn.ReLU(),            # input_size=(64*128*128)
            nn.MaxPool2d(2, 2)    # output_size=(64*64*64)
        )

        # 全連接層(將神經(jīng)網(wǎng)絡(luò)的神經(jīng)元的多維輸出轉(zhuǎn)化為一維)
        self.fc1 = nn.Sequential(
            nn.Linear(64 * 64 * 64, 128),  # 進(jìn)行線性變換
            nn.ReLU()                    # 進(jìn)行ReLu激活
        )

        # 輸出層(將全連接層的一維輸出進(jìn)行處理)
        self.fc2 = nn.Sequential(
            nn.Linear(128, 84),
            nn.ReLU()
        )

        # 將輸出層的數(shù)據(jù)進(jìn)行分類(輸出預(yù)測(cè)值)
        self.fc3 = nn.Linear(84, 62)

    # 定義前向傳播過(guò)程，輸入為x
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        # nn.Linear()的輸入輸出都是維度為一的值，所以要把多維度的tensor展平成一維
        x = x.view(x.size()[0], -1)
        x = self.fc1(x)
        x = self.fc2(x)
        x = self.fc3(x)
        return x

# 中間特征提取
class FeatureExtractor(nn.Module):
    def __init__(self, submodule, extracted_layers):
        super(FeatureExtractor, self).__init__()
        self.submodule = submodule
        self.extracted_layers = extracted_layers
 
    def forward(self, x):
        outputs = []
        print(self.submodule._modules.items())
        for name, module in self.submodule._modules.items():
            if "fc" in name: 
                print(name)
                x = x.view(x.size(0), -1)
            print(module)
            x = module(x)
            print(name)
            if name in self.extracted_layers:
                outputs.append(x)
        return outputs


def get_feature():
    # 輸入數(shù)據(jù)
    img = get_picture(pic_dir, transform)
    # 插入維度
    img = img.unsqueeze(0)

    img = img.to(device)

    # 特征輸出
    net = LeNet().to(device)
    # net.load_state_dict(torch.load('./model/net_050.pth'))
    exact_list = ["conv1"，"conv2"]
    myexactor = FeatureExtractor(net, exact_list)
    x = myexactor(img)

    # 特征輸出可視化
    for i in range(32):
        ax = plt.subplot(6, 6, i + 1)
        ax.set_title('Feature {}'.format(i))
        ax.axis('off')
        plt.imshow(x[0].data.numpy()[0,i,:,:],cmap='jet')

    plt.show()

# 訓(xùn)練
if __name__ == "__main__":
    get_picture_rgb(pic_dir)
    # get_feature()