Will Tebbutt

Updating gradients in-place

3

Add support for providing pre-allocated elements of the gradient tape. Currently viable as a hack of the following form: ``` using Nabla, NNlib t = Tape() W = Leaf(t, rand(5,5))...

Documentation

2

1. Find a decent git-hostable package for producing documentation. 2. Cover interface (how do I get gradients / how do I add my own sensitivities + reference to useful documents...

It would be useful to support stuff like this: https://github.com/denizyuret/Knet.jl/issues/144 We can pretty much already support this due to the design of the function `∇` in `core.jl`, specifically the method...

Currently admit Integer-valued quantities to be differentiated w.r.t.. This is problematic in that the derivative a function w.r.t. an input which is an Integer is not necessary itself an Integer...

Upon completion of #29 (and #30 for completeness) it should be possible to make existing packages differentiable automatically. This will be achieved by accessing the source code for the package...

Multi-output `broadcast` / `map`

1

Currently, when `broadcast` and `map` are used to apply a function `f` which accepts multiple inputs to the appropriate number of arrays of data, each input is treated individually as...

Support for Circulant matrices should, in principle, happen partly by default. There may, however, be certain situations in which hand-coding is necessary for efficiency.

Simple MLP demo

1

Code up a simple MLP classifier for the MNIST data set.

VAE Demo

1

Code up a demo of the Variational AutoEncoder for the MNIST data set.

Code up toy Real NVP for generating MNIST.