This repository has been archived by the owner on Jul 1, 2024. It is now read-only.
Tags: facebookresearch/ClassyVision
Tags
Update mixup transform to be compatible with latest pytorch (#753) Summary: Pull Request resolved: #753 Pull Request resolved: fairinternal/ClassyVision#97 `torch.clip` is an alias for `torch.clamp`, see https://pytorch.org/docs/stable/generated/torch.clip.html?highlight=clip#torch.clip and `torch.clip` is no longer in the latest pytorch version Updated `fairscale` dependency to resolve cpu/gpu test circleci failures: https://app.circleci.com/pipelines/github/facebookresearch/ClassyVision/2201/workflows/e3b37369-677c-4a08-8b4b-5c5c217c5560/jobs/4322 Reviewed By: mannatsingh Differential Revision: D29520476 fbshipit-source-id: 584f079762bc4c6de644cc7595232109b364b4db
Profiler bug-fixes and improvements (#482) Summary: Pull Request resolved: #482 Made the following changes to the profiler code - - The parameter calculation skipped parameters defined in modules, and had unnecessary complexity - updated the logic - `AdaptiveAvgPool2d` handles a single number `output_size` as well - Added FLOPs for the `Identity` module - Added support to specify an `activations` function to fetch activations from (similar to the `flops` function) - Replaced the hacky list append logic with a class, `_ComplexityComputer` - Implemented test cases which verify that the fixes work Reviewed By: vreis Differential Revision: D21009734 fbshipit-source-id: 926d93164c13c6c98eb88f9131d295b61d6acda7
Make bn weight decay configurable (#65) Summary: Pull Request resolved: fairinternal/ClassyVision#65 Make the bn weight decay configurable, for some datasets it might be desirable to turn it off. Reviewed By: vreis Differential Revision: D20140487 fbshipit-source-id: 77debf2c4600a080081668565d70b7a3ddc788f4