No foce_fp32 in 3.x branch?

[twmht]

In version 2.x, there is force_fp32 when calculating the loss (https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/dense_heads/fcos_head.py#L167).

But in 3.x branch, there is no force_fp32 (https://github.com/open-mmlab/mmdetection/blob/v3.0.0rc0/mmdet/models/dense_heads/fcos_head.py#L179), this may cause loss explosion, for example, in MESLoss, the maximum value is 65536, this is not enough in some cases. I have no idea how you handle such cases.

is 3.x branch ready to be used?

[twmht]

closed