emerald-ai / torchmetal

A library for Meta-Learning and Few-Shot Learning with PyTorch.
MIT License
1 stars 3 forks source link

Padding Speed / Pad Size #17

Open iron-beagle opened 2 years ago

iron-beagle commented 2 years ago

Currently the Padding for batch size is done using a torch collate function in ironbeagle-torchmetal BatchMetaCollatePad. Its call hard-codes the largest batch size, 500 currently. Also, the only tested image shape has been concrete through transforms in the helpers.

The desired behavior is to Pad according to the largest batch and largest image efficiently. Finding this across all tasks in an episode needs to be implemented and used in the BatchMetaCollatePad