diff --git a/train_colab.py b/train_colab.py index b7243de..93a418e 100644 --- a/train_colab.py +++ b/train_colab.py @@ -487,7 +487,7 @@ def main(args): betas = (0.9, 0.999) epsilon = 1e-8 if args.amp: - epsilon = 1e-8 + epsilon = 2e-8 weight_decay = 0.01 if args.useadam8bit: