Benoît Roussel comments

Repositories
Issues
Comments

Results 2 comments of


                                            Benoît Roussel

训练过慢，且梯度为NaN

Hello, I also experimented with integrating CLEX into my model and observed similarly a significant reduction in training speed and `inf` gradient norms. Increasing the timestep from `0.01 `to `0.1`...

Encountering "Expected to have finished reduction in the prior iteration before starting a new one" Error During Training

If you encounter a frame with no GT, the denoising part of the model will not go through dn.class_embed() for one GPU of your DDP setup, thus raising an error...