fp16 report nan on mask rcnn

I use apex in the mask_rcnn_r50_fpn_1x.py mmdetection, and use pytorch roi layer.

1. when I choose opt_level: O1, report error:

  from torchvision.ops import roi_align as tv_roi_align
  File "/opt/anaconda3/lib/python3.7/site-packages/torchvision/__init__.py", line 1, in <module>
    from torchvision import models
  File "/opt/anaconda3/lib/python3.7/site-packages/torchvision/models/__init__.py", line 12, in <module>
    from . import detection
  File "/opt/anaconda3/lib/python3.7/site-packages/torchvision/models/detection/__init__.py", line 1, in <module>
    from .faster_rcnn import *
  File "/opt/anaconda3/lib/python3.7/site-packages/torchvision/models/detection/faster_rcnn.py", line 13, in <module>
    from .rpn import AnchorGenerator, RPNHead, RegionProposalNetwork
  File "/opt/anaconda3/lib/python3.7/site-packages/torchvision/models/detection/rpn.py", line 8, in <module>
    from . import _utils as det_utils
  File "/opt/anaconda3/lib/python3.7/site-packages/torchvision/models/detection/_utils.py", line 74, in <module>
    @torch.jit.script
  File "/opt/anaconda3/lib/python3.7/site-packages/torch/jit/__init__.py", line 1181, in script
    return _compile_function(fn=obj, qualified_name=qualified_name, _frames_up=_frames_up + 1, _rcb=_rcb)
  File "/opt/anaconda3/lib/python3.7/site-packages/torch/jit/__init__.py", line 1077, in _compile_function
    script_fn = torch._C._jit_script_compile(qualified_name, ast, _rcb, get_default_args(fn))
  File "/opt/anaconda3/lib/python3.7/site-packages/torch/jit/__init__.py", line 1009, in _try_compile_fn
    qualified_name = _qualified_name(fn)
  File "/opt/anaconda3/lib/python3.7/site-packages/torch/_jit_internal.py", line 404, in _qualified_name
    "__module__ can't be None.".format(name))
RuntimeError: Could not get qualified name for class 'log': __module__ can't be None.

2. when I choose opt_level: O2, report error:

ValueError: nan or inf found in loss.

I set loss_scale to 1.0, the program also report this error.

Reproduce the two configurations using mask_rcnn_r50_fpn_1x.py, Apex opt_level O1 and O2, and the torchvision.ops roi_align path mentioned in the traceback. Start by recording the PyTorch, torchvision, Apex, and mmdetection versions, then isolate whether the import failure and the NaN loss have separate causes. Done means identifying a reproducible cause and a verified fix or clear compatibility guidance.