I just download robotwin2 open_lap task to finetune, use 8 cards H20, it will take 500hours to finish 1000000 steps, it is right?

the diff_loss and loss go to what level, 0.0001 or 0.00001? then i can stop train?

Training 1M, which means 1000k steps, is about 10 days or more on my side, maybe h20 will be slower.

Diffusion type vla does not make much sense to look at the loss, and we can focus on the mse and l2 indicators in eval.

Navigation Menu