EDiRobotics / GR1-Training

Reimplementation of GR-1, a generalized policy for robotics manipulation.
Apache License 2.0
91 stars 3 forks source link

About Training Costs #4

Open HUAFOR opened 1 month ago

HUAFOR commented 1 month ago

Hi, Thanks for you great work to reproduce the training code for GR-1! I wonder how long it takes to complete the training process for GR-1 from scratch?[ABC->D setting]

StarCycle commented 2 weeks ago

@HUAFOR

Sorry! My fault! I just saw your issue here...

It's not recommended to use this repo to train it from scratch. Some developers tried it but the performance is not as good as the original version, though I try my best to recover every training details they used.

By contrast, you can train from the pretrained checkpoint provided by GR-MG.

For faster response you can send me an email...sorry again...I am working on video generation model and my own MimicTest policy toolbox in these days

Best, Zhuoheng

StarCycle commented 2 weeks ago

For your original question, please refer to this issue

They use 32 V100 32GB. But no worry, in my experience you can achieve roughly the same speed with 8*4090 GPU. If you open torch compiling option in my repo, it can even be 50% faster!