[Refactor] CSL

[zytx121]

• Align test
• Align train
• UT

Experiments

rotated_retinanet_obb_csl_gaussian_r50_fpn_fp16_1x_dota_le90.py
report AP50: 69.51
AP50: 0.6804708695155224
AP75: 0.39561891292657414
mAP: 0.3920965305629459

[liuyanyi]

The shape of returns is wrong, i think the return here is (num_anchors * H * W).
In order to support batch input like (B, N, coding_len), the argmax dim can be -1.

In retinanet, csl directly assign the angle into rbox

# `rbbox` has shape (N, 5)
# `ang` has shape (N)
rbbox[..., 4] = ang

In some angle_branch head, such as Rotated FCOS and Rotated YOLOX, they use concat to get final bbox.

# `hbbox` has shape (N, 4)
# `ang` has shape (N, 1)
rbbox = torch.cat([hbbox, ang], dim=-1)

I think we can add a keepdim arg in decode method, and use like this:

angle_cls_inds = torch.argmax(angle_preds, dim=-1, keepdim=keepdim)

When keepdim is true, return (N, 1) or (B, N, 1), otherwise (N,) or (B, N)

[zytx121]

Done.