Flux2Lux.FluxLayer — Type
FluxLayer(layer)

Serves as a compatibility layer between Flux and Lux. This uses Optimisers.destructure API internally.

Warning

Lux was written to overcome the limitations of destructure + Flux. It is recommended to rewrite your l in Lux instead of using this layer.

Warning

Introducing this Layer in your model will lead to type instabilities, given the way Optimisers.destructure works.

Arguments

  • layer: Flux layer

Parameters

  • p: Flattened parameters of the layer
Flux2Lux.transform — Method
transform(l; preserve_ps_st::Bool=false, force_preserve::Bool=false)

Convert a Flux Model to Lux Model.

Warning

transform always ingores the active field of some of the Flux layers. This is almost never going to be supported on Flux2Lux.

Arguments

  • l: Flux l or any generic Julia function / object.

Keyword Arguments

  • preserve_ps_st: Set to true to preserve the states and parameters of the l. This attempts the best possible way to preserve the original model. But it might fail. If you need to override possible failures, set force_preserve to true.

  • force_preserve: Some of the transformations with state and parameters preservation haven't been implemented yet, in these cases, if force_transform is false a warning will be printed and a core Lux layer will be returned. Else, it will create a FluxLayer.

Examples

using Flux2Lux, Lux, Metalhead, Random

m = ResNet(18)
m2 = transform(m.layers)

x = randn(Float32, 224, 224, 3, 1);

ps, st = Lux.setup(Random.default_rng(), m2);

m2(x, ps, st)