it occurs 'nan' in training_stage_1

[lwmming]

When we run 'train_stage_1.py' on CelebV-HQ dataset, we find that the training loss will be 'nan'.
Then we try to locate this issue and find that after running line76 in train_stage_1.py, i.e., 'pose_fea = self.pose_guider(pose_cond_tensor, ref_pose_tensor)' , 'nan' will occur in pose_fea[1].
Have you encountered similar problems and how did you solve them? Thank you~

[Zejun-Yang]

We recommend you to use fp16.