fast-reid（HAA）之loss = getattr(Loss, cfg.MODEL.LOSSES.NAME[0])(cfg)(..., ..., ...)解析

怼烎@ 2022-12-20 05:36 105阅读 0赞

参考：[fast-reid（HAA）][fast-reid_HAA]

目的是理解如下代码：

from fastreid.modeling import losses as Loss
    loss = getattr(Loss, cfg.MODEL.LOSSES.NAME[0])(cfg)(pred_class_logits, global_features, gt_classes)

## 1.getattr内部参数 ##

Python getattr() 函数：getattr() 函数用于返回一个对象属性值。

解析Loss与对应的cfg文件中：

from fastreid.modeling import losses as Loss

涉及到\_\_init\_\_导入模型：

（参考：[https://github.com/xbq1994/HAA/blob/497abc5745/fastreid/modeling/losses/\_\_init\_\_.py][https_github.com_xbq1994_HAA_blob_497abc5745_fastreid_modeling_losses_init_.py]与[https://www.cnblogs.com/lands-ljk/p/5880483.html][https_www.cnblogs.com_lands-ljk_p_5880483.html]）

# encoding: utf-8
    """
    @author:  l1aoxingyu
    @contact: sherlockliao01@gmail.com
    """
    
    from .build_losses import reid_losses
    from .cross_entroy_loss import CrossEntropyLoss
    from .focal_loss import FocalLoss
    from .metric_loss import *

[https://github.com/xbq1994/HAA/blob/497abc5745/projects/Black\_reid/configs/Base-HAA.yml][https_github.com_xbq1994_HAA_blob_497abc5745_projects_Black_reid_configs_Base-HAA.yml]

MODEL:
      LOSSES:
        NAME: ("CrossEntropyLoss", "TripletLoss")

解析后：

loss = getattr(Loss, cfg.MODEL.LOSSES.NAME[0])(cfg)(pred_class_logits, global_features, gt_classes)
    
    为：
    loss = getattr(Loss, “CrossEntropyLoss”)(cfg)(pred_class_logits, global_features, gt_classes)

## 2.最外层getattr ##

对应的实例（[https://www.runoob.com/python/python-func-getattr.html][https_www.runoob.com_python_python-func-getattr.html]）：

>>>class A(object):
    ...     bar = 1
    ... 
    >>> a = A()
    >>> getattr(a, 'bar')        # 获取属性 bar 值
    1

解析后：

loss = getattr(Loss, “CrossEntropyLoss”)(cfg)(pred_class_logits, global_features, gt_classes)
    
    为：
    getattr()函数替换：
    loss = CrossEntropyLoss(cfg)(pred_class_logits, global_features, gt_classes)

## 3.后面第一个括号 ##

为什么有第一个括号里的参数？

用于实例化（[https://github.com/xbq1994/HAA/blob/497abc5745/fastreid/modeling/losses/cross\_entroy\_loss.py][https_github.com_xbq1994_HAA_blob_497abc5745_fastreid_modeling_losses_cross_entroy_loss.py]），\_\_init\_\_中所带参数实例化时就需要。

class CrossEntropyLoss(object):
        """
        A class that stores information and compute losses about outputs of a Baseline head.
        """
    
        def __init__(self, cfg):
            self._num_classes = cfg.MODEL.HEADS.NUM_CLASSES
            self._eps = cfg.MODEL.LOSSES.CE.EPSILON
            self._alpha = cfg.MODEL.LOSSES.CE.ALPHA
            self._scale = cfg.MODEL.LOSSES.CE.SCALE
    
            self._topk = (1,)

loss = (CrossEntropyLoss(cfg))(pred_class_logits, global_features, gt_classes)

## 4.后面第二个括号 ##

此处直接通过调用实例就执行了\_\_call\_\_函数（[https://blog.csdn.net/lmz\_lmz/article/details/89136970][https_blog.csdn.net_lmz_lmz_article_details_89136970]）。

调用\_\_call\_\_函数时需要三个参数（主要是要兼容不同的损失函数，才有三个参数的）：

def __call__(self, pred_class_logits, _, gt_classes):
            """
            Compute the softmax cross entropy loss for box classification.
            Returns:
                scalar Tensor
            """
            self._log_accuracy(pred_class_logits, gt_classes)
            if self._eps >= 0:
                smooth_param = self._eps
            else:
                # adaptive lsr
                soft_label = F.softmax(pred_class_logits, dim=1)
                smooth_param = self._alpha * soft_label[torch.arange(soft_label.size(0)), gt_classes].unsqueeze(1)
    
            log_probs = F.log_softmax(pred_class_logits, dim=1)
            with torch.no_grad():
                targets = torch.ones_like(log_probs)
                targets *= smooth_param / (self._num_classes - 1)
                targets.scatter_(1, gt_classes.data.unsqueeze(1), (1 - smooth_param))
    
            loss = (-targets * log_probs).mean(0).sum()
            return {
                "loss_cls": loss * self._scale,
            }

loss = (CrossEntropyLoss(cfg))(pred_class_logits, global_features, gt_classes)
    
    
    最后：
    loss = {
                "loss_cls": loss * self._scale,
            }

最后返回是一个损失函数字典。

## 5.总结 ##

一层层分析下来，也就不是那么难了，python中一些不需要调用直接可以使用的方法，还是需要慢慢去了解学习的。

### 参考 ###

[1. 解析代码来源][fast-reid_HAA]

[2. 1中参考\_\_init\_\_方法][https_github.com_xbq1994_HAA_blob_497abc5745_fastreid_modeling_losses_init_.py]

[3. 1中\_\_init\_\_解析][https_www.cnblogs.com_lands-ljk_p_5880483.html]

[4. 1中cfg文件][https_github.com_xbq1994_HAA_blob_497abc5745_projects_Black_reid_configs_Base-HAA.yml]

[5. Python getattr() 函数][https_www.runoob.com_python_python-func-getattr.html]

[6. 3中cross\_entroy\_loss.py][https_github.com_xbq1994_HAA_blob_497abc5745_fastreid_modeling_losses_cross_entroy_loss.py]

[7. 4中\_\_call\_\_方法][https_blog.csdn.net_lmz_lmz_article_details_89136970]

[fast-reid_HAA]: https://github.com/xbq1994/HAA/blob/497abc574584691529d0d0b49ec6a6ee8d2d0072/projects/Black_reid/build_losses.py#L5
[https_github.com_xbq1994_HAA_blob_497abc5745_fastreid_modeling_losses_init_.py]: https://github.com/xbq1994/HAA/blob/497abc5745/fastreid/modeling/losses/__init__.py
[https_www.cnblogs.com_lands-ljk_p_5880483.html]: https://www.cnblogs.com/lands-ljk/p/5880483.html
[https_github.com_xbq1994_HAA_blob_497abc5745_projects_Black_reid_configs_Base-HAA.yml]: https://github.com/xbq1994/HAA/blob/497abc5745/projects/Black_reid/configs/Base-HAA.yml
[https_www.runoob.com_python_python-func-getattr.html]: https://www.runoob.com/python/python-func-getattr.html
[https_github.com_xbq1994_HAA_blob_497abc5745_fastreid_modeling_losses_cross_entroy_loss.py]: https://github.com/xbq1994/HAA/blob/497abc5745/fastreid/modeling/losses/cross_entroy_loss.py
[https_blog.csdn.net_lmz_lmz_article_details_89136970]: https://blog.csdn.net/lmz_lmz/article/details/89136970