Wancong (Kevin) Zhang comments

Repositories
Issues
Comments

Results 2 comments of


                                            Wancong (Kevin) Zhang

bug

has anyone found the solution to this bug? I also encountered it.

mismatch between paper and repo hyperparameters

Another difference is that clip_grad_norm and clip_model_grad_norm are set to -1 in [sgim_finetune.sh](https://github.com/mila-iqia/SGI/blob/master/scripts/experiments/sgim_finetune.sh), whereas in the SPR paper you clipped the gradients to 10