sforaidl / jeta Goto Github PK

max has a name collision with the in-built max. Don't know if could be a big problem but certainly, we could change max to something else. Ideas? (metax, metajax, ...) ?

Add:

• Wandb logger
• Tensorboard logger
• Local logger

Paper: https://arxiv.org/abs/1807.01613

Paper Link: https://arxiv.org/abs/1909.09157

This can also be added as an extension to MAML

Inherit form ABC

• codecov.yml
• .coveragerc.
• .isort.cfg
• requirements.txt
• tox.ini
• Also, .gitignore

Currently, there are no tests, and hence no code coverage.

We should test the algorithms (at least 1 or 2 epochs with a small fas) to check if the forward pass and parameter updates are working properly.

This can be tackled after #26 is merged.

cc: @veds12

Relevant:

[1] https://github.com/marketplace/actions/automatic-changelog-generator
[2] https://docs.github.com/en/actions/learn-github-actions/understanding-github-actions

[1] (FO-MAML) https://arxiv.org/abs/1803.02999
[2] (i-MAML) https://papers.nips.cc/paper/2019/file/072b030ba126b2f4b2374f342be9ed44-Paper.pdf

The above lines cause an error while running MAML using OptiTrainer

OptiTrainer calls the function maml_adapt and it's parameters are passed in OptiTrainer itself.

Note: maml_adapt is passed as a parameter to OptiTrainer as adapt_fn

jeta/jeta/opti_trainer.py

Line 145 in e55bf83

theta = adapt_fn(params, apply_fn, loss_fn, support_set)

An error is being thrown because maml_adapt expects the above 2 arguments which are not being passed.

fas can be included into OptiTrainer as this parameter is common across many Optimization based approaches, but maml_lr can't be included as it is specific to only MAML.
Hence a different approach must be used to allow it as a parameter.

A development guide will help new contributors set up the environment and thus create a smoother experience.
Maybe we can have a section for the same in the readme.md file or a separate contributing.md file.

Thanks for suggesting @AnirudhM1 and @Devanshu42 !

• We can follow the same structure as learn2learn's TaskDataset.
• Number of ways and shots can be passed as parameters to the TaskDataset

Flax optimizers have been deprecated in favour of optax optimizers. We need to make the shift in the existing implementation(s)

https://arxiv.org/abs/1707.09835