Sean Moriarity

You can continue working in this PR! Let me know if you run into any problems and I'll be glad to help!

Hi @goodhamgupta I am closing this for now as stale, if you want to re-open and continue work feel free and I will do my best to guide you in...

Thinking about this a struct may or may not make sense, but it definitely makes sense to make it easier to access state of specific transformations. For example if I...

Hi @idabmat ! Thank you for your contributions! I think this is a good idea! Right now I don't have much bandwidth to maintain this library anymore, if you would...

Here's an example: ```elixir input = Axon.input("input") x1 = Axon.dense(input, 32) x2 = Axon.dense(input, 64) model = Axon.add(x1, x2) {init_fn, predict_fn} = Axon.build(model) init_fn.(Nx.template({1, 16}, :f32), %{}) ``` And the...

This was resolved earlier :)

@josevalim I think that's okay because you still wouldn't be able to deserialize the closure correctly in another program? In cases where you want more flexibility, you can just serialize...

There is a workaround for this which I just pushed, but as you can see it is failing because there are some inconsistencies in generating parameter shapes. Basically if you...

Sure, here's a failing inference: ```elixir model = Axon.input("data") |> Axon.dense(32) |> Axon.relu() |> Axon.dropout(rate: 1.0) {init_fn, predict_fn} = Axon.build(model, mode: :train) params = init_fn.(Nx.iota({1, 10}), %{}) predict_fn.(params, Nx.iota({1, 10}))...

@Ian-GL In most DL frameworks we scale the kept activations IOT preserve original input properties (mean, variance, etc.) In this case 1.0 is definitely an unrealistic value so it's not...