Closed BartekCupial closed 6 months ago
I'm getting this message when the script finishes, not sure yet why.
[W CudaIPCTypes.cpp:15] Producer process has been terminated before all shared CUDA tensors released. See Note [Sharing CUDA tensors]
@alex-petrenko can I ask for the review, please?
Attention: 318 lines
in your changes are missing coverage. Please review.
Comparison is base (
6379cf9
) 79.52% compared to head (a5fde39
) 76.32%.
:exclamation: Your organization needs to install the Codecov GitHub app to enable full functionality.
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
I'm getting this message when the script finishes, not sure yet why.
[W CudaIPCTypes.cpp:15] Producer process has been terminated before all shared CUDA tensors released. See Note [Sharing CUDA tensors]
Yes, this started to happen after some PyTorch update. I've been meaning to investigate, but I also found that it's totally harmless and in practice can be ignored.
Overview
This PR introduces a new script,
eval.py
, designed for faster evaluation using multiple environments and leveraging the efficiency of the Sample Factory sampler. Additionally, an example usage ofeval.py
is demonstrated in the addedeval_mujoco.py
script withinsf_examples
.Key Changes
New Script:
eval.py
enjoy.py
.Example Usage:
eval_mujoco.py
eval.py
script.Core Logic in
evaluation_sampling_api.py
simplified_sampling_api.py
.Additional features
checkpoint_override_defaults
. Has the same purpose asload_from_checkpoint(cfg)
, but keeps the option of overriding cfg with argv.EpisodeCounterWrapper
, which counts the episodes for each environment. Used duringeval
in order not to bias the results with shorter episodes.