【人脸检测】复现Pytorch_Retinaface(Pytorch版本)

淩亂°似流年 2022-11-21 14:58 249阅读 0赞

[https://github.com/biubug6/Pytorch\_Retinaface][https_github.com_biubug6_Pytorch_Retinaface]

训练这个模型很耗GPU，mobilenet0.25还好（训练8G以内，测试1G以内），resnet50至少要43G。

## 0.环境 ##

ubuntu16.04
    python3.6
    cuda9.0
    torch==1.1.0
    ipython

## 1.准备数据与模型 ##

### （1）准备数据 ###

原始数据：train+val+test [http://shuoyang1213.me/WIDERFACE/WiderFace\_Results.html][http_shuoyang1213.me_WIDERFACE_WiderFace_Results.html]

标签数据：[https://pan.baidu.com/s/1Laby0EctfuJGgGMgRRgykA][https_pan.baidu.com_s_1Laby0EctfuJGgGMgRRgykA]

按照官方给的格式放：

./data/widerface/
        train/
          images/
          label.txt
        val/
          images/
          label.txt
        test/
          images/
          label.txt

由于val只有label.txt，所以我们要自己转一下（这一步一定要转一下，不然会测试时会出错的）：

# -*- coding: UTF-8 -*-
    '''
    @author: mengting gu
    @contact: 1065504814@qq.com
    @time: 2020/11/2 上午11:47
    @file: widerValFile.py
    @desc: 
    '''
    import os
    import argparse
    
    parser = argparse.ArgumentParser(description='Retinaface')
    parser.add_argument('--dataset_folder', default='./data/widerface/val/images/', type=str, help='dataset path')
    args = parser.parse_args()
    
    if __name__ == '__main__':
        # testing dataset
        testset_folder = args.dataset_folder
        testset_list = args.dataset_folder[:-7] + "label.txt"
    
        with open(testset_list, 'r') as fr:
            test_dataset = fr.read().split()
        num_images = len(test_dataset)
    
    
        for i, img_name in enumerate(test_dataset):
            print("line i :{}".format(i))
            if img_name.endswith('.jpg'):
                print("     img_name :{}".format(img_name))
                f = open(args.dataset_folder[:-7] + 'wider_val.txt', 'a')
                f.write(img_name+'\n')
        f.close()

最后将数据按照作者的要求，写成下面这种：

./data/widerface/
        train/
          images/
          label.txt
        val/
          images/
          label.txt
          wider_val.txt
        test/
          images/
          label.txt

### （2）准备模型 ###

[google cloud][] and [baidu cloud][] Password: fstq

## 2.训练 ##

CUDA_VISIBLE_DEVICES=0 python train.py --network mobile0.25

要想训练Resnet50的话，在train.py代码的139行后， 添加：

torch.cuda.empty_cache()

## 3.测试评估 ##

这份代码原来是测试与评估是分开的。

### （1）测试 ###

python test_widerface.py --trained_model weight_file --network mobile0.25 
    
    # eg
    python test_widerface.py --trained_model ./weights/mobilenet0.25_epoch_245.pth --network mobile0.25 
    
    # eg 测试结果保存 去掉test_widerface.py中参数设置parser（save_image） action="store_true"
    python test_widerface.py --trained_model ./weights/mobilenet0.25_epoch_245.pth --network mobile0.25 -s=True

检测结果jpg与txt分别在results与widerface\_evaluate/widerface\_txt目录下。

检测画框结果：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70][]

检测txt结果：

0_Parade_marchingband_1_20
    261
    542 357 36 42 0.9983047 
    29 404 27 33 0.9953412 
    254 371 29 33 0.98255116 
    629 394 19 21 0.96621066 
    464 355 27 29 0.9414734 
    81 391 22 25 0.9405682 
    386 368 17 20 0.9325298 
    730 346 17 19 0.88296247

### （2）评估 ###

cd ./widerface_evaluate
    python setup.py build_ext --inplace
    python evaluation.py

![20201102141827511.png][]

下面是作者自己跑出来的结果：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1][]

对比来看与第二行这项结果相差较小，基本能复现结果了。

## 参考 ##

[1.Pytorch\_Retinaface][https_github.com_biubug6_Pytorch_Retinaface]

[2.官方RetinaFace][2._RetinaFace]

[https_github.com_biubug6_Pytorch_Retinaface]: https://github.com/biubug6/Pytorch_Retinaface
[http_shuoyang1213.me_WIDERFACE_WiderFace_Results.html]: http://shuoyang1213.me/WIDERFACE/WiderFace_Results.html
[https_pan.baidu.com_s_1Laby0EctfuJGgGMgRRgykA]: https://pan.baidu.com/s/1Laby0EctfuJGgGMgRRgykA
[google cloud]: https://drive.google.com/open?id=1oZRSG0ZegbVkVwUd8wUIQx8W7yfZ_ki1
[baidu cloud]: https://pan.baidu.com/s/12h97Fy1RYuqMMIV-RpzdPg
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70]: /images/20221120/90aa64cc96b04f3f830ed759298bc7af.png
[20201102141827511.png]: /images/20221120/bca31411c9424ef6a9533da0c2de5721.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1]: /images/20221120/11e4f9faa6204262a1172b152956e173.png
[2._RetinaFace]: https://github.com/deepinsight/insightface/tree/master/RetinaFace