Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
1612: fix AdamW and improve decays docs r=DhairyaLGandhi a=CarloLucibello There is great disorder under the sky with optimizers. Since in chaining optimizers ``` opt = Optimizer(opt1, opt2) ``` the order generally matters (a lot!) we have to be very careful in documenting how to use decays. In fact, we were giving completely wrong indirections for `InvDecays` and `ExpDecays`. The correct ordering for standard use is ```julia Optimizer(WeightDecay(), ADAM()) # equivalent to L2 regularization Optimizer(ADAM(), InvDecay()) # learning rate scheduling Optimizer(ADAM(), ExpDecay()) # learning rate scheduling ``` Different orderings are to be typically considered as bugs in user code. This PR fixes examples and tries to clarify documentation in this regard. Also fixes AdamW, which was doing something totally wrong due to the aforementioned confusion. (see https://towardsdatascience.com/why-adamw-matters-736223f31b5d for how AdamW works). Related in model-zoo: FluxML/model-zoo#303 and FluxML/model-zoo#304 Co-authored-by: CarloLucibello <carlo.lucibello@gmail.com> Co-authored-by: Carlo Lucibello <carlo.lucibello@unibocconi.it>
- Loading branch information