【行人重识别】fast-reid复现（20210111）

淩亂°似流年 2023-01-05 05:12 451阅读 0赞

参考代码：

[https://github.com/JDAI-CV/fast-reid][https_github.com_JDAI-CV_fast-reid]

# 0.环境 #

ubuntu16.04
    cuda9.0
    python3.6
    torch==1.1.0
    torchvision==0.3.0
    Cython
    yacs
    tensorboard
    future
    termcolor
    sklearn
    tqdm
    opencv-python==4.1.0.25
    matplotlib
    scikit-image
    numpy==1.16.4

安装apex（不要直接通过pip安装）：

git clone https://www.github.com/nvidia/apex
    cd apex
    # python setup.py install --cuda_ext --cpp_ext 
    pip install -v --no-cache-dir ./

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70][]

# 1.准备数据 #

参考[https://blog.csdn.net/qq\_35975447/article/details/106664593][https_blog.csdn.net_qq_35975447_article_details_106664593]

数据目录结构如下：

fast-reid
    	datasets
    		Market-1501-v15.09.15
    			bounding_box_train
    			bounding_box_test
    			query

# 2.修改 #

## 2.1 RandomErasing ##

我的torchvision是0.3.0版本的，所以需要修改一些地方。如果torchvision不是高版本的，需要

修改三处：

（1）fastreid/data/transforms/build.py line 66-77

# fastreid/data/transforms/build.py line 66-77
    
        # res.append(ToTensor()) # gmt modify
    
            if do_rea:
                res.append(RandomErasing(probability=rea_prob, mean=rea_value))  # gmt modify
            if do_rpt:
                res.append(RandomPatch(prob_happen=rpt_prob))
        else:
            size_test = cfg.INPUT.SIZE_TEST
            res.append(T.Resize(size_test, interpolation=3))
            # res.append(ToTensor()) # gmt modify
        res.append(ToTensor()) # gmt modify
        return T.Compose(res)

（2） fastreid/data/transforms/transforms.py line 7

# fastreid/data/transforms/transforms.py line 7
    # __all__ = ['ToTensor', 'RandomPatch', 'AugMix', ]   # gmt modify
    __all__ = ['ToTensor', 'RandomErasing', 'RandomPatch', 'AugMix', ]  # gmt modify

（3）fastreid/data/transforms/transforms.py line 44-86

# fastreid/data/transforms/transforms.py line 44-86
    
    class RandomErasing(object):
        """ Randomly selects a rectangle region in an image and erases its pixels.
            'Random Erasing Data Augmentation' by Zhong et al.
            See https://arxiv.org/pdf/1708.04896.pdf
        Args:
            probability: The probability that the Random Erasing operation will be performed.
            sl: Minimum proportion of erased area against input image.
            sh: Maximum proportion of erased area against input image.
            r1: Minimum aspect ratio of erased area.
            mean: Erasing value.
        """
    
        def __init__(self, probability=0.5, sl=0.02, sh=0.4, r1=0.3, mean=255 * (0.49735, 0.4822, 0.4465)):
            self.probability = probability
            self.mean = mean
            self.sl = sl
            self.sh = sh
            self.r1 = r1
    
        def __call__(self, img):
            img = np.asarray(img, dtype=np.float32).copy()
            if random.uniform(0, 1) > self.probability:
                return img
    
            for attempt in range(100):
                area = img.shape[0] * img.shape[1]
                target_area = random.uniform(self.sl, self.sh) * area
                aspect_ratio = random.uniform(self.r1, 1 / self.r1)
    
                h = int(round(math.sqrt(target_area * aspect_ratio)))
                w = int(round(math.sqrt(target_area / aspect_ratio)))
    
                if w < img.shape[1] and h < img.shape[0]:
                    x1 = random.randint(0, img.shape[0] - h)
                    y1 = random.randint(0, img.shape[1] - w)
                    if img.shape[2] == 3:
                        img[x1:x1 + h, y1:y1 + w, 0] = self.mean[0]
                        img[x1:x1 + h, y1:y1 + w, 1] = self.mean[1]
                        img[x1:x1 + h, y1:y1 + w, 2] = self.mean[2]
                    else:
                        img[x1:x1 + h, y1:y1 + w, 0] = self.mean[0]
                    return img
            return img

## 2.2 CUDA float类型 ##

# "./fastreid/layers/circle_softmax.py", line 38
    pred_class_logits = targets * s_p + (1.0 - targets) * s_n
    
    # 改为：
    pred_class_logits = targets.float() * s_p.float() + (1.0 - targets.float()) * s_n.float()

## 2.3 cross\_entroy\_loss ##

# fastreid/modeling/losses/cross_entroy_loss.py line 50
    
    non_zero_cnt = max(loss.nonzero().size(0), 1)  # gmt add
    # non_zero_cnt = max(loss.nonzero(as_tuple=False).size(0), 1)

## 2.4 data/commom.py(多数据训练的） ##

# line 26-27
    self.pids = sorted(list(pid_set))
    self.cams = sorted(list(cam_set))
    
    #改为：
    
    self.pids = sorted(list(pid_set), key=lambda x: str(x))
    self.cams = sorted(list(cam_set), key=lambda x: str(x))

## 2.5 config ##

由于sbs后面会调用bot的yaml文件，所以这里要注释掉bot中yaml的pretrain\_model（路径：configs/Base-bagtricks.yml）:

![20210115105251761.png][]

# 3.训练 #

CUDA_VISIBLE_DEVICES="0,1" python ./tools/train_net.py --config-file='./configs/Market1501/sbs_R101-ibn.yml'

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3][]

对比最后一项，可以看到mAP是可以复现对应结果的。

复现BOT R101-ibn（默认参数，步骤还是要按之前的来）:

CUDA_VISIBLE_DEVICES="0" python ./tools/train_net.py --config-file='./configs/Market1501/bagtricks_R101-ibn.yml'

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 4][]

[https_github.com_JDAI-CV_fast-reid]: https://github.com/JDAI-CV/fast-reid
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70]: /images/20221119/4513cdba58c44534a42df43384b68f7b.png
[https_blog.csdn.net_qq_35975447_article_details_106664593]: https://blog.csdn.net/qq_35975447/article/details/106664593
[20210115105251761.png]: /images/20221119/84413baa80c049689b69c0e1c3ce839b.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1]: /images/20221119/1c77553a340748bab289c4e0d17042c0.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2]: /images/20221119/7bb12f1fffd34c30acfd14316fe2dd8b.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/20210113093111626.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 4]: https://img-blog.csdnimg.cn/20210118105829425.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3,size_16,color_FFFFFF,t_70