Suggestion: make it easier to initialize parameters

[AStupidBear]

Whenever I write a network with non-standard structures, I'm constantly bothered by the model initialization because it's even more tedious than writing the network itself. I have seen some efforts but that's still not mature and satisfactory. Do you have any good idea to mitigate this pain?

[denizyuret]

Do you have a specific suggestion?

…

On Tue, Jul 18, 2017, 15:18 Yao Lu ***@***.***> wrote: Whenever I write a network with non-standard structures, I'm constantly bothered by the model initialization because it's even more tedious than writing the network itself. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#147>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABvNpmjyKXrTE4iYSROMfwDW8EWmX3eHks5sPKKcgaJpZM4ObQvl> .

[AStupidBear]

I think PyTorch is quite good at this, but mimicking PyTorch's approach may require an overhaul of Knet and AutoGrad. Although the main focus of Knet now is transparent autograd on GPU, evolving a higher level API is still necessary for the long run.

[MikeInnes]

You may be interested in #144.

[jgbos]

I've also been playing with an toy project to address this

https://github.com/jgbos/KFuddles

[AStupidBear]

@jgbos Your project was mentioned in my hyperlink.

[denizyuret]

Please check out issue #347.