search

nv-tlabs / ASE

Other
793 stars 128 forks source link

How to tune hyper parameters for GPU's with more memory? #56

Closed Robokan closed 1 year ago

Robokan commented 1 year ago

I have a dedicated RTX 4090 with 24 g of memory. When training ASE it only uses only half of the available memory. What changes to the hyper parameters can I make to use more of the GPU?

mjdxyr commented 1 year ago

Maybe you can look https://github.com/NVlabs/CALM/issues/4 for some help.

Robokan commented 1 year ago

Yes, I have been running CALM as well. I read that post a while back and used it to turn everything down. Hoping to turn everything up on ASE as it's less resource intensive.

xbpeng commented 1 year ago

You can increase memory usage by increasing the number of environments numEnvs in https://github.com/nv-tlabs/ASE/blob/main/ase/data/cfg/humanoid_ase.yaml, increase the number of samples collected per iteration horizon_length, and the minibatch size used for gradient calculation minibatch_size and amp_minibatch_size in https://github.com/nv-tlabs/ASE/blob/main/ase/data/cfg/train/rlg/ase_humanoid.yaml.

Robokan commented 1 year ago

Great thanks! Exactly what I was looking for. Did you ever try reducing the precision to say 16 to improve performance?