[关闭]
@wanghuijiao 2021-01-22T13:51:04.000000Z 字数 4649 阅读 1161

解决数据不均衡：为FairMOT添加WeightedRandomSampler

实验报告

0 前言

1. 添加到FairMOT后的效果

  1. # ap_now:E15-1 epoch=4; ap_before:E14 epoch=10
  2. # 测试集：AVA_manfu_hik.val
  3. +------------+-------+-------+--------+-------+---------+
  4. | class      | gts   | dets  | recall | ap_now|ap_before|
  5. +------------+-------+-------+--------+-------+---------+
  6. | stand      | 18092 | 46136 | 0.804  | 0.276 | 0.259   |
  7. | sit        | 12715 | 32296 | 0.824  | 0.283 | 0.214   |
  8. | squat      | 238   | 1764  | 0.752  | 0.134 | 0.000   |
  9. | lie        | 989   | 12472 | 0.817  | 0.078 | 0.038   |
  10. | half_lie   | 1754  | 23154 | 0.808  | 0.081 | 0.071   |
  11. | other_pose | 43    | 33    | 0.000  | 0.000 | 0.000   |
  12. | person     | 33831 | 69055 | 0.838  | 0.356 | 0.356   |
  13. +------------+-------+-------+--------+-------+---------+
  14. | mAP        |       |       |        | 0.173 | 0.134   |
  15. +------------+-------+-------+--------+-------+---------+

2. 实验结论

3. WeightedRandomSampler 源码解析及使用示例

  1. class WeightedRandomSampler(Sampler[int]):
  2.     r"""Samples elements from ``[0,..,len(weights)-1]`` with given probabilities (weights).
  4.     Args:
  5.         weights (sequence)   : a sequence of weights, not necessary summing up to one
  6.         num_samples (int): number of samples to draw
  7.         replacement (bool): if ``True``, samples are drawn with replacement.
  8.             If not, they are drawn without replacement, which means that when a
  9.             sample index is drawn for a row, it cannot be drawn again for that row.
  10.         generator (Generator): Generator used in sampling.
  12.     Example:
  13.         >>> list(WeightedRandomSampler([0.1, 0.9, 0.4, 0.7, 3.0, 0.6], 5, replacement=True))
  14.         [4, 4, 1, 4, 5]
  15.         >>> list(WeightedRandomSampler([0.9, 0.4, 0.05, 0.2, 0.3, 0.1], 5, replacement=False))
  16.         [0, 1, 4, 3, 2]
  17.     """
  18.     weights: Tensor
  19.     num_samples: int
  20.     replacement: bool
  22.     def __init__(self, weights: Sequence[float], num_samples: int,
  23.                  replacement: bool = True, generator=None) -> None:
  24.         if not isinstance(num_samples, _int_classes) or isinstance(num_samples, bool) or \
  25.                 num_samples <= 0:
  26.             raise ValueError("num_samples should be a positive integer "
  27.                              "value, but got num_samples={}".format(num_samples))
  28.         if not isinstance(replacement, bool):
  29.             raise ValueError("replacement should be a boolean value, but got "
  30.                              "replacement={}".format(replacement))
  31.         self.weights = torch.as_tensor(weights, dtype=torch.double)
  32.         self.num_samples = num_samples
  33.         self.replacement = replacement
  34.         self.generator = generator
  35.     def __iter__(self):
  36.         rand_tensor = torch.multinomial(self.weights, self.num_samples, self.replacement, generator=self.generator)
  37.         return iter(rand_tensor.tolist())
  39.     def __len__(self):
  40.         return self.num_samples
  1. class WeightedRandomSampler(Sampler):
  2.     r"""Samples elements from ``[0,..,len(weights)-1]`` with given probabilities (weights)."""
  3.     def __init__(self, weights, num_samples, replacement=True):
  4.          # ...省略类型检查
  5.          # weights用于确定生成索引的权重
  6.         self.weights = torch.as_tensor(weights, dtype=torch.double)
  7.         self.num_samples = num_samples
  8.         # 用于控制是否对数据进行有放回采样
  9.         self.replacement = replacement
  10.     def __iter__(self):
  11.         # 按照加权返回随机索引值
  12.         return iter(torch.multinomial(self.weights, self.num_samples, self.replacement).tolist())

对于Weighted Random Sampler类的__init__()来说，replacement参数用于控制采样是否是有放回的(Ture为放回抽样)；num_sampler用于控制生成的个数；weights参数对应的是“样本”的权重而不是“类别的权重”。其中__iter__()方法返回的数值为随机数序列，只不过生成的随机数序列是按照weights指定的权重确定的，测试代码如下：

  1. # 位置[0]的权重为0，位置[1]的权重为10，其余位置权重均为1.1
  2. weights = torch.Tensor([0, 0.5, 0.1, 0.1, 0.1, 0.1, 0.1])
  3. wei_sampler = sampler.WeightedRandomSampler(weights, 6, True)
  4. # 下面是输出：
  5. index: 1
  6. index: 2
  7. index: 3
  8. index: 4
  9. index: 1
  10. index: 1

4. 加入FairMOT之前需要准备什么

5. 添加到FairMOT需要做什么

  1.     # 加载weights list
  2.     with open(opt.weights_dir,'r') as file:
  3.         line = file.readlines()
  4.         weights = line[0][1:-2].split(', ')           
  5.         weights = list(map(float, weights))
  7.     # list -> Tensor
  8.     weights = torch.Tensor(weights)
  9.     sampler = torch.utils.data.sampler.WeightedRandomSampler(
  10.         weights, len(weights), replacement=True)
  11.     # Get dataloader
  12.     train_loader = torch.utils.data.DataLoader(dataset,
  13.                                                batch_size=opt.batch_size,
  14.                                                shuffle=False,
  15.                                                num_workers=opt.num_workers,
  16.                                                pin_memory=True,
  17.                                                drop_last=True,
  18.                                                sampler=sampler)
添加新批注
在作者公开此批注前，只有你和作者可见。
回复批注