sigmoid_focal_loss

paddle.fluid.layers.sigmoid_focal_loss ( x, label, fg_num, gamma=2.0, alpha=0.25 ) [源代码]

Focal Loss 被提出用于解决计算机视觉任务中前景-背景不平衡的问题。该OP先计算输入x中每个元素的sigmoid值，然后计算sigmoid值与类别目标值label之间的Focal Loss。

Focal Loss的计算过程如下：

sigmoid_focal_loss - 图1

其中，已知：

sigmoid_focal_loss - 图2

参数：

  • x (Variable) – 维度为

    sigmoid_focal_loss - 图3

    的2-D Tensor，表示全部样本的分类预测值。其中，第一维N是批量内参与训练的样本数量，例如在目标检测中，样本为框级别，N为批量内所有图像的正负样本的数量总和；在图像分类中，样本为图像级别，N为批量内的图像数量总和。第二维:math:C 是类别数量（ 不包括背景类 ）。数据类型为float32或float64。

  • label (Variable) – 维度为

    sigmoid_focal_loss - 图4

    的2-D Tensor，表示全部样本的分类目标值。其中，第一维N是批量内参与训练的样本数量，第二维1表示每个样本只有一个类别目标值。正样本的目标类别值的取值范围是

    sigmoid_focal_loss - 图5

    , 负样本的目标类别值是0。数据类型为int32。

  • fg_num (Variable) – 维度为

    sigmoid_focal_loss - 图6

    的1-D Tensor，表示批量内正样本的数量，需在进入此OP前获取正样本的数量。数据类型为int32。

  • gamma (int|float) – 用于平衡易分样本和难分样本的超参数， 默认值设置为2.0。

  • alpha (int|float) – 用于平衡正样本和负样本的超参数，默认值设置为0.25。

返回： 输入x中每个元素的Focal loss，即维度为

sigmoid_focal_loss - 图7

的2-D Tensor。

返回类型： 变量（Variable），数据类型为float32或float64。

代码示例

  1. import numpy as np
  2. import paddle.fluid as fluid
  4. num_classes = 10  # exclude background
  5. image_width = 16
  6. image_height = 16
  7. batch_size = 32
  8. max_iter = 20
  11. def gen_train_data():
  12.     x_data = np.random.uniform(0, 255, (batch_size, 3, image_height,
  13.                                         image_width)).astype('float64')
  14.     label_data = np.random.randint(0, num_classes,
  15.                                    (batch_size, 1)).astype('int32')
  16.     return {"x": x_data, "label": label_data}
  19. def get_focal_loss(pred, label, fg_num, num_classes):
  20.     pred = fluid.layers.reshape(pred, [-1, num_classes])
  21.     label = fluid.layers.reshape(label, [-1, 1])
  22.     label.stop_gradient = True
  23.     loss = fluid.layers.sigmoid_focal_loss(
  24.         pred, label, fg_num, gamma=2.0, alpha=0.25)
  25.     loss = fluid.layers.reduce_sum(loss)
  26.     return loss
  29. def build_model(mode='train'):
  30.     x = fluid.data(name="x", shape=[-1, 3, -1, -1], dtype='float64')
  31.     output = fluid.layers.pool2d(input=x, pool_type='avg', global_pooling=True)
  32.     output = fluid.layers.fc(
  33.         input=output,
  34.         size=num_classes,
  35.         # Notice: size is set to be the number of target classes (excluding backgorund)
  36.         # because sigmoid activation will be done in the sigmoid_focal_loss op.
  37.         act=None)
  38.     if mode == 'train':
  39.         label = fluid.data(name="label", shape=[-1, 1], dtype='int32')
  40.         # Obtain the fg_num needed by the sigmoid_focal_loss op:
  41.         # 0 in label represents background, >=1 in label represents foreground,
  42.         # find the elements in label which are greater or equal than 1, then
  43.         # computed the numbers of these elements.
  44.         data = fluid.layers.fill_constant(shape=[1], value=1, dtype='int32')
  45.         fg_label = fluid.layers.greater_equal(label, data)
  46.         fg_label = fluid.layers.cast(fg_label, dtype='int32')
  47.         fg_num = fluid.layers.reduce_sum(fg_label)
  48.         fg_num.stop_gradient = True
  49.         avg_loss = get_focal_loss(output, label, fg_num, num_classes)
  50.         return avg_loss
  51.     else:
  52.         # During evaluating or testing phase,
  53.         # output of the final fc layer should be connected to a sigmoid layer.
  54.         pred = fluid.layers.sigmoid(output)
  55.         return pred
  58. loss = build_model('train')
  59. moment_optimizer = fluid.optimizer.MomentumOptimizer(
  60.     learning_rate=0.001, momentum=0.9)
  61. moment_optimizer.minimize(loss)
  62. place = fluid.CPUPlace()
  63. exe = fluid.Executor(place)
  64. exe.run(fluid.default_startup_program())
  65. for i in range(max_iter):
  66.     outs = exe.run(feed=gen_train_data(), fetch_list=[loss.name])
  67.     print(outs)