RLE-Foundation / rllte

Long-Term Evolution Project of Reinforcement Learning
https://docs.rllte.dev/
MIT License
467 stars 86 forks source link

[Feature Request] Addition of Kaiming He Normalization Option for Layer Initialization #51

Open Zach-Attach opened 3 weeks ago

Zach-Attach commented 3 weeks ago

🚀 Feature

Add a third option for weight_init as 'kaiming he' (see https://pytorch.org/docs/stable/nn.init.html and https://arxiv.org/pdf/1803.08494).

Motivation

I have been using this method of initialization for a while as it performs better than default or orthogonal. I thought it would make sens to add this to the public repo.sense

Pitch

Already wrote the code changes here. https://github.com/RLE-Foundation/rllte/commit/5b08294dbfb49b8edaf005d04e364bd97de297a2

Alternatives

No response

Additional context

Tests do not currently exist for layer initialization, so I haven't added any for such a simple addition.

Checklist