Edan Toledo
Edan Toledo
Hey, i thought I'd also chime in here. I realised this difference and i simply made a wrapper to achieve the same auto-reset style as Gym API. My wrapper is...
I haven't read the hindsight relabeling paper so there might be context i am missing but this sounds achievable just with the trajectory buffer and no extra functionality. Correct me...
Aah I had a feeling this would happen soon. I'll try do this asap. But ideally we fix it so we can use the latest version. Is the error with...
You would have to specify the batch index as well as the starting time index but this is pretty easy functionality to add.
Thanks so much, I'll try review this and test it tomorrow on a GPU.
I just did a comparison, and it seems like sebulba on main is faster. Looking at all the timing statistics, its the pipeline that is slowing things down. Everything else...
> I just did a comparison, and it seems like sebulba on main is faster. Looking at all the timing statistics, its the pipeline that is slowing things down. Everything...
so i ended up completely refactoring the sebulba architectures taking a mix of inspiration from cleanba and instadeeps one. When timing the previous systems versus cleanba, cleanba was much faster...
I think its due to the import, however, its possible this issue is not correct. I'll try make a dummy script to check at some point.
Thanks for the response! So i'm having trouble getting any performance above zero (let alone matching the performance of the paper) for the door key 8x8 environment. Using these reproduction...