pytorch实现用Resnet提特征并保存为txt文件

接触pytorch一天，发现pytorch上手的确比TensorFlow更快。可以更方便地实现用预训练的网络提特征。

以下是提取一张jpg图像的特征的程序：

# -*- coding: utf-8 -*-

import os.path

import torch
import torch.nn as nn
from torchvision import models, transforms
from torch.autograd import Variable  

import numpy as np
from PIL import Image  

features_dir = './features'

img_path = "hymenoptera_data/train/ants/0013035.jpg"
file_name = img_path.split('/')[-1]
feature_path = os.path.join(features_dir, file_name + '.txt')


transform1 = transforms.Compose([
        transforms.Scale(256),
        transforms.CenterCrop(224),
        transforms.ToTensor()    ]
)

img = Image.open(img_path)
img1 = transform1(img)

#resnet18 = models.resnet18(pretrained = True)
resnet50_feature_extractor = models.resnet50(pretrained = True)
resnet50_feature_extractor.fc = nn.Linear(2048, 2048)
torch.nn.init.eye(resnet50_feature_extractor.fc.weight)

for param in resnet50_feature_extractor.parameters():
    param.requires_grad = False
#resnet152 = models.resnet152(pretrained = True)
#densenet201 = models.densenet201(pretrained = True) 
x = Variable(torch.unsqueeze(img1, dim=0).float(), requires_grad=False)
#y1 = resnet18(x)
y = resnet50_feature_extractor(x)
y = y.data.numpy()
np.savetxt(feature_path, y, delimiter=',')
#y3 = resnet152(x)
#y4 = densenet201(x)

y_ = np.loadtxt(feature_path, delimiter=',').reshape(1, 2048)

以下是提取一个文件夹下所有jpg、jpeg图像的程序：

# -*- coding: utf-8 -*-
import os, torch, glob
import numpy as np
from torch.autograd import Variable
from PIL import Image  
from torchvision import models, transforms
import torch.nn as nn
import shutil
data_dir = './hymenoptera_data'
features_dir = './features'
shutil.copytree(data_dir, os.path.join(features_dir, data_dir[2:]))


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


    x = Variable(torch.unsqueeze(img, dim=0).float(), requires_grad=False)
    if use_gpu:
        x = x.cuda()
        net = net.cuda()
    y = net(x).cpu()
    y = y.data.numpy()
    np.savetxt(saved_path, y, delimiter=',')
    
if __name__ == '__main__':
    extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']
        
    files_list = []
    sub_dirs = [x[0] for x in os.walk(data_dir) ]
    sub_dirs = sub_dirs[1:]
    for sub_dir in sub_dirs:
        for extention in extensions:
            file_glob = os.path.join(sub_dir, '*.' + extention)
            files_list.extend(glob.glob(file_glob))
        
    resnet50_feature_extractor = models.resnet50(pretrained = True)
    resnet50_feature_extractor.fc = nn.Linear(2048, 2048)
    torch.nn.init.eye(resnet50_feature_extractor.fc.weight)
    for param in resnet50_feature_extractor.parameters():
        param.requires_grad = False   
        
    use_gpu = torch.cuda.is_available()

    for x_path in files_list:
        print(x_path)
        fx_path = os.path.join(features_dir, x_path[2:] + '.txt')
        extractor(x_path, fx_path, resnet50_feature_extractor, use_gpu)
        

另外最近发现一个很简单的提取不含FC层的网络的方法：

        resnet = models.resnet152(pretrained=True)
        modules = list(resnet.children())[:-1]      # delete the last fc layer.
        convnet = nn.Sequential(*modules)

另一种更简单的方法：

resnet = models.resnet152(pretrained=True)
del resnet.fc