paddle复现pytorch踩坑(八）：smooth_l1的用法

fluid.layers.smooth_l1

• 作用：该layer计算变量 x 和 y 的smooth L1 loss，它以 x 和 y 的第一维大小作为批处理大小。对于每个实例，按元素计算smooth L1 loss，然后计算所有loss。输出变量的形状是[batch_size, 1]
• 链接：pp飞桨API说明

问题一：输入数据类型

# pytorch code
loss_bbox_x = F.smooth_l1_loss(bbox_x[active], bbox_x_tar[active], reduction='none')
    
# paddlepaddle code
loss_bbox_x = fluid.layers.smooth_l1(bbox_x[active], bbox_x_tar[active])

报错：Smooth_L1 只能用’float32’, 且输入shape与torch不能，rank>1
解决：

# paddlepaddle code
loss_bbox_x = fluid.layers.smooth_l1(fluid.layers.cast(fluid.layers.unsqueeze(fluid.layers.gather(bbox_x, index), 0), 'float32'),
                                     fluid.layers.cast(fluid.layers.unsqueeze(fluid.layers.gather(bbox_x_tar, index), 0), 'float32'))

问题二：注意输出的维度

输出loss shape 为 [1, 1]

# paddlepaddle code
loss_bbox_x = fluid.layers.smooth_l1(fluid.layers.reshape(fluid.layers.gather(bbox_x, index_active), [1, -1]),
                                     fluid.layers.reshape(fluid.layers.gather(bbox_x_tar, index_active), [1, -1]))

输出loss shape 为[N, 1]

# paddlepaddle code
loss_bbox_x = fluid.layers.smooth_l1(fluid.layers.reshape(fluid.layers.gather(bbox_x, index_active), [-1, 1]),
                                     fluid.layers.reshape(fluid.layers.gather(bbox_x_tar, index_active), [-1, 1]))