AdamW

Source: optim/AdamW.luau

AdamW (Loshchilov & Hutter, ICLR 2019; arXiv:1711.05101). Decoupled weight decay. Bias-corrected m and v per step.

Constructor

luau

Class: new(params: {Tensor}, lr: number, opts: {beta1?, beta2?, eps?, weightDecay?})

API

step()
zeroGrad()
setLr(lr)
getLr()
stateDict()
loadStateDict(state)
paramCount()

Methods

`new(params: { any }, lr: number, opts: any?) -> any`

`step()`

`zeroGrad()`

`setLr(lr: number)`

`getLr() -> number`

`paramCount() -> number`

`stateDict() -> any`

`loadStateDict(st: any)`