resnet提取图片特征

下载模型

wget https://digix-algo-challenge.obs.cn-east-2.myhuaweicloud.com/2020/cv/6rKDTsB6sX8A1O2DA2IAq7TgHPdSPxJF/train_data.zip -o train_data.zip

ResNet提取图片特征向量

from torch.autograd import Variable
import torch
import torch.nn as nn
import torchvision.transforms as transforms
import pretrainedmodels
from PIL import Image

import ssl
ssl._create_default_https_context = ssl._create_unverified_context

TARGET_IMG_SIZE = 224
img_to_tensor = transforms.ToTensor()
def get_seresnet50():
    encoder = pretrainedmodels.se_resnet50()
    model = nn.Sequential(encoder.layer0,
                          encoder.layer1,
                          encoder.layer2,
                          encoder.layer3,
                          encoder.layer4,
                          encoder.avg_pool   # 平均池化，张成一个[batchSize，2048]的特征向量
                          )
    for param in model.parameters():
        param.requires_grad = False
    # model.cuda()   # 使用GPU，CPU版去掉
    model.eval()
    return model

# 特征提取
def extract_feature(model, imgpath):
    img = Image.open(imgpath)  # 读取图片
    img = img.resize((TARGET_IMG_SIZE, TARGET_IMG_SIZE))

    tensor = img_to_tensor(img)  # 将图片矩阵转化成tensor
    # tensor = tensor.cuda()  # GPU 使用GPU放开此注释
    tensor = torch.unsqueeze(tensor, 0)
    result = model(Variable(tensor))
    result_npy = result.data.cpu().numpy()[0].ravel().tolist()
    return result_npy

if __name__ == '__main__':
    model = get_seresnet50()
    feature = extract_feature(model, "1.jpeg")
    print(len(feature))
    print(feature)
    # 下载模型可能会报错，将上面下载的模型放在下载的目录

参考

https://www.codenong.com/cs107121229/

参考图片特征提取