Closed alexander-soare closed 5 months ago
There's still a little more to do here. See https://github.com/huggingface/lerobot/pull/41
See https://github.com/huggingface/lerobot/issues/51. The results have been reproduced, but it was achieved by using the original repo's dataset/loaders instead of ours. We still need to fix ours.
@Cadene I'll add you to this as it will be a good test for your remove torchrl PR.
FYI, I am able to run a pretrained model from the notebook of diffusion policy. I didnt calculate the success rate, but I think it is higher than what I have been able to achieve by training from scratch with LeRobot :/
See this PR: https://github.com/Cadene/lerobot/pull/13/files
Also I tried different size of network, with and without prioritized sampler, but success rate is not converging to ~70%. I am training on 1 gpu only for 12 hours. I am not sure I am using the correct hyperparameters. The config yaml I am using is still not super clean and need to be sanitized ^^