search

arrayfire / arrayfire-ml

ArrayFire's Machine Learning Library.
BSD 3-Clause "New" or "Revised" License
102 stars 23 forks source link

Conv2D #37

Closed plavin closed 7 years ago

plavin commented 7 years ago

The autograd part still needs to be written.

I also need to be able to join the outputs of multiple inputs, but I'm having trouble getting join to work. This is needed to do batched processing.

Currently arrayfire can join up to 10 arrays at a time. This may or may not be an acceptable batch size. A better solution would be to implement 3d and 4d matmul in arrayfire.

plavin commented 7 years ago

@pavanky Added the rest of the Conv2D layer - although it still isn't batched. I need to find a way to confirm it is working properly.

I added some more gradients of af functions.

pavanky commented 7 years ago

@plavin I want to figure out a way we can pass stuff that's not a variable to the backward pass.

Also wait until after https://github.com/arrayfire/arrayfire-ml/pull/36 is merged.

plavin commented 7 years ago

@pavanky This pr is kind of a mess now. I think I'll abandon this one and move changes to separate commits. One with extra functions and one with conv2d layer. And a third with the alexnet/example.cpp

pavanky commented 7 years ago

@plavin sounds good.