Inference agent fix, improvements when using NCCL backend, and other minor improvements.

Lucaweihs commented 2 years ago

This PR makes the following changes:

Fixes some problems with the inference agent that occurred when resetting the agent.
Adds a read timeout to calls to vector sampled tasks so that the training process will exit correctly if the vector sampled tasks fails to respond for some amount of time (currently 1 minute but this may be too low).
Adds experimental code to restart vector sampled tasks if an environment ever happens to throw an uncaught error. This functionality is disabled by default.
Adds a new habitat experiment and HM3D support.
Improves the vision sensor to ensure that it copies images before modifing them (necessary for the new version of ai2thor)
A few other minor fixes.

lgtm-com[bot] commented 2 years ago

This pull request introduces 1 alert when merging 21b663a6007c31122fa84b1b8542b24e9e7878ef into 474fb84789fcdd3917fbc006365653971994e93f - view on LGTM.com

new alerts:

1 for Unused import

lgtm-com[bot] commented 2 years ago

This pull request introduces 1 alert when merging 481c44980fcf6e59ca2693646805e99258aaf2d1 into 474fb84789fcdd3917fbc006365653971994e93f - view on LGTM.com

new alerts:

1 for Unused import

lgtm-com[bot] commented 2 years ago

This pull request introduces 1 alert when merging 1d3c64ae1f4d9dc44763d7e9c449e01add0464af into 474fb84789fcdd3917fbc006365653971994e93f - view on LGTM.com

new alerts:

1 for Unused import

allenai / allenact

Inference agent fix, improvements when using NCCL backend, and other minor improvements. #345