Resolve Runtime Error: CUDA error: device side assert triggered. When setting class_weight

[Laser-one]

Describe the bug

When setting class_weight=[1.0,9.0], a CUDA error: device side assert triggered appears because during data_preprocessor, seg_map that smaller crop_size will be pad to crop_size with 255. This leads to the absence of class_weight[255] when class_weight=[1.0, 9.0].

Bug fix
The simplest method is to set seg_ pad_val=-1 in data_preprocessor.
Then, set ignore_index=seg_ pad_val=-1 in decode_ head and/or auxiliary_head.
In addition, in order to avoid the loss of average coefficients and consider padding labels, set avg_non_ignore=True in loss_decode

[Laser-one]

当设置class_weight时，会出现device-side assert triggered，这是因为在图像预处理的时候，小于crop_size的seg_map会被填充为255。这就导致当class_weight=[1.0,9.0]时，不存在class_weight[255]。
一个最简单的方法是设置seg_pad_val=-1,
然后在decode_head和auxiliary_head也设置ignore_index=seg_pad_val=-1。
此外，为了避免损失平均系数考虑填充的label，在loss_decode中设置avg_non_ignore=True