Metriche repository
- Star
- (4725 star)
- Metriche merge PR
- (Merge medio 4h 27m) (2 PR mergiate in 30 g)
Descrizione
As things stand, only some layers in Flux have direct functional equivalents in NNlib - maxpool and meanpool do, for example, while the adaptive versions of those two layers don't. I know that they're only thin wrappers around maxpool and meanpool themselves, but would it be worth having functions that do the forward pass for these layers? While writing wrapper layers in Metalhead.jl it often feels weird to have entire Flux layers described inside of the forward pass of another layer. An example I'm working on right now is adaptive_avgmax_pool, currently written out as:
function adaptive_avgmax_pool(x, output_size = (1, 1))
x_avg = AdaptiveMeanPool(output_size)(x)
x_max = AdaptiveMaxPool(output_size)(x)
return 0.5 * (x_avg + x_max)
end
While this works perfectly fine, it feels kinda odd to have a layer inside of a function which is gonna be wrapped by a layer 😅 (this may just be a side-effect of having worked with torch quite a bit so I'm not sure if this request is justified or not)