ParameterSchedulers.jl
ParameterSchedulers.jl copied to clipboard
FluxML
Common hyperparameter scheduling for ML
### Motivation and description The generation of a sequence of finite-length schedules like, e.g., ```julia Sequence(1e-1 => 5, 5e-2 => 4, 3.4e-3 => 10) ``` would conceptually be a finite-length...
`ComposedSchedule` is only implicitly tested through the cyclic schedules. This probably leaves out edge cases.
This is to write down a thought which came from #34 and https://github.com/FluxML/Optimisers.jl/pull/89. Presently, we rely on mutably/immutably updating any objects which depend on the schedule value after each step....
### Motivation and description Currently it's pretty difficult to extract current parameters from a scheduler for two reasons: - You typically have to descend into a parameter tree structure to...
Hi! Thank you for developing and maintaining this library! I just noticed that the [docstring of `CosAnneal`](https://github.com/FluxML/ParameterSchedulers.jl/blob/22742e1e3c955c783fb512ea4e92a1a2e2065443/src/cyclic.jl#L189) may be incorrect? Shouldn't it be ``` t̂ = restart ? mod(t -...
This pull request changes the compat entry for the `InfiniteArrays` package from `0.10.4, 0.11, 0.12, 0.13, 0.14` to `0.10.4, 0.11, 0.12, 0.13, 0.14, 0.15`. This keeps the compat entries for...
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "github-actions[bot] avatar"){kind=avatar}
This docstring reads to me as if the the first argument will be `l0` (taking the same interpretation as in the kwarg constructor) https://github.com/FluxML/ParameterSchedulers.jl/blob/f6ca703973fdf93e69bc17932fdafe84a8ee9b37/src/cyclic.jl#L128 But the "arguments" list below reveals...
Just some typo fixes for the equations describing the Poly schedule to match the code implementation and the intuition for how this schedule should behave with step count. ### PR...
