PyTorch框架下分別使用Vgg、Resnet、Densenet提取圖像集特徵

這裏主要是使用預訓練好的模型進行圖片特徵的提取，分別使用三個模型進行抽取。

而特徵提取是提取神經網絡模型的倒數第二層，這裏自己對模型的進行了微調，話不多說，直接上代碼。

使用resnet

import torch
import torch.nn as nn
from torch.autograd import Variable
from torchvision import models, transforms
from PIL import Image
import numpy as np
import os, glob

data_dir = './test'   # train
features_dir = './Resnet_features_test'  # Resnet_features_train


//這裏自己修改網絡
class net(nn.Module):
    def __init__(self):
        super(net, self).__init__()
        self.net = models.resnet50(pretrained=True)

    def forward(self, input):
        output = self.net.conv1(input)
        output = self.net.bn1(output)
        output = self.net.relu(output)
        output = self.net.maxpool(output)
        output = self.net.layer1(output)
        output = self.net.layer2(output)
        output = self.net.layer3(output)
        output = self.net.layer4(output)
        output = self.net.avgpool(output)
        return output


model = net()
//加載cuda
model = model.cuda()


def extractor(img_path, saved_path, net, use_gpu):
    transform = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor()]
    )

    img = Image.open(img_path)
    img = transform(img)
    print(img.shape)

    x = Variable(torch.unsqueeze(img, dim=0).float(), requires_grad=False)
    print(x.shape)

    if use_gpu:
        x = x.cuda()
        net = net.cuda()
    y = net(x).cpu()
    y = torch.squeeze(y)
    y = y.data.numpy()
    print(y.shape)
    np.savetxt(saved_path, y, delimiter=',')


if __name__ == '__main__':
    extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']

    files_list = []
    x = os.walk(data_dir)
    for path,d,filelist in x:
        for filename in filelist:
            file_glob = os.path.join(path, filename)
            files_list.extend(glob.glob(file_glob))

    print(files_list)

    use_gpu = torch.cuda.is_available()

    for x_path in files_list:
        print("x_path" + x_path)
        file_name = x_path.split('/')[-1]
        fx_path = os.path.join(features_dir, file_name + '.txt')
        print(fx_path)
        extractor(x_path, fx_path, model, use_gpu)

使用vgg

import torch
import torch.nn as nn
from torch.autograd import Variable
from torchvision import models, transforms
from PIL import Image
import numpy as np
import os, glob

data_dir = './test'  # train
features_dir = './Vgg_features_test'  # Vgg_features_train


class Encoder(nn.Module):
    def __init__(self):
        super(Encoder, self).__init__()
        VGG = models.vgg16(pretrained=True)
        self.feature = VGG.features
        self.classifier = nn.Sequential(*list(VGG.classifier.children())[:-3])
        pretrained_dict = VGG.state_dict()
        model_dict = self.classifier.state_dict()
        pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
        model_dict.update(pretrained_dict)
        self.classifier.load_state_dict(model_dict)

    def forward(self, x):
        output = self.feature(x)
        output = output.view(output.size(0), -1)
        output = self.classifier(output)
        return output


model = Encoder()
model = model.cuda()


def extractor(img_path, saved_path, net, use_gpu):
    transform = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor()]
    )

    img = Image.open(img_path)
    img = transform(img)
    print(img.shape)

    x = Variable(torch.unsqueeze(img, dim=0).float(), requires_grad=False)
    print(x.shape)

    if use_gpu:
        x = x.cuda()
        net = net.cuda()
    y = net(x).cpu()
    y = torch.squeeze(y)
    y = y.data.numpy()
    print(y.shape)
    np.savetxt(saved_path, y, delimiter=',')


if __name__ == '__main__':
    extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']

    files_list = []
    x = os.walk(data_dir)
    for path, d, filelist in x:
        for filename in filelist:
            file_glob = os.path.join(path, filename)
            files_list.extend(glob.glob(file_glob))

    print(files_list)

    use_gpu = torch.cuda.is_available()

    for x_path in files_list:
        print("x_path" + x_path)
        file_name = x_path.split('/')[-1]
        fx_path = os.path.join(features_dir, file_name + '.txt')
        print(fx_path)
        extractor(x_path, fx_path, model, use_gpu)

使用Densnet

import torch
import torch.nn as nn
from torch.autograd import Variable
from torchvision import models, transforms
from PIL import Image
import numpy as np
import os, glob

data_dir = './test'  # train
features_dir = './DenseNet_features_test'  # DenseNet_features_train


class Encoder(nn.Module):
    def __init__(self):
        super(Encoder, self).__init__()
        densnet = models.densenet121(pretrained=True)
        self.feature = densnet.features
        self.classifier = nn.Sequential(*list(densnet.classifier.children())[:-1])
        pretrained_dict = densnet.state_dict()
        model_dict = self.classifier.state_dict()
        pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
        model_dict.update(pretrained_dict)
        self.classifier.load_state_dict(model_dict)

    def forward(self, x):
        output = self.feature(x)
        avg = nn.AvgPool2d(7, stride=1)
        output = avg(output)
        return output


model = Encoder()
model = model.cuda()


def extractor(img_path, saved_path, net, use_gpu):
    transform = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor()]
    )

    img = Image.open(img_path)
    img = transform(img)
    print(img.shape)

    x = Variable(torch.unsqueeze(img, dim=0).float(), requires_grad=False)
    print(x.shape)

    if use_gpu:
        x = x.cuda()
        net = net.cuda()
    y = net(x).cpu()
    y = torch.squeeze(y)
    y = y.data.numpy()
    print(y.shape)
    //保存爲txt
    np.savetxt(saved_path, y, delimiter=',')


if __name__ == '__main__':
    extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']

    files_list = []
    x = os.walk(data_dir)
    for path, d, filelist in x:
        for filename in filelist:
            file_glob = os.path.join(path, filename)
            files_list.extend(glob.glob(file_glob))

    print(files_list)

    use_gpu = torch.cuda.is_available()

 
    for x_path in files_list:
        # print("x_path" + x_path)
        file_name = x_path.split('/')[-1]
        fx_path = os.path.join(features_dir, file_name + '.txt')
        # print(fx_path)
        extractor(x_path, fx_path, model, use_gpu)

以上是使用三個模型的特徵提取，可以根據不同的需求進行選擇。