google / brax

Massively parallel rigidbody physics simulation on accelerator hardware.
Apache License 2.0
2.14k stars 234 forks source link

Updates Needed in Datasets #379

Open bebark opened 11 months ago

bebark commented 11 months ago

The grad_updates_per_step hyperparameter for SAC changed at some point in the last year or two from being a decimal value to being an integer. I believe the old values are just the inversions of the new (e.g. 8 is 0.125). Is that correct? Either way the brax/datasets readme and data are out of sync with the rest of the repo since they follow the old decimal convention.

LabChameleon commented 6 months ago

I had a look at the old training code and I don't think the number of grad_updates_per_step is the inverted decimal number. Considering https://github.com/google/brax/blob/46720002c49cfc287dc45e196227990f0be1e3f9/brax/training/sac.py#L450 the number of steps is num_envs * grad_updates_per_step.