Guillaume Dalle

In some cases, the forward pass will compute a Jacobian along with the output. We should use it instead of duplicating the work in the conditions

Inspired by https://arxiv.org/abs/1909.01377 (section 3.1.3)

At the moment, autodiff of implicit functions relies on ChainRules, which means ForwardDiff won't work. Suggested fix: manually overload the callable on ForwardDiff.Dual

Change target for "How to install Julia and Pluto?"

3

The "Edit or run this notebook" button is one of my favorite features of Pluto, since it contributes a great deal to ease of use and reproducibility. At the moment,...

Hi everyone, Whenever I want to use PyCall.jl, I find myself scrolling around the README looking for the information I need. Would a classic web page built with Documenter.jl be...

AD failing with ComponentArrays

10

Hi there, and thanks for the great package Are ComponentArrays supposed to be AD-compatible? If so, did I do something wrong with the following code? ```julia julia> using ComponentArrays, Zygote...

Blitz tutorial takes too long to run (GD instead of SGD?)

2

Hi there! In the 60-minute blitz tutorial (https://fluxml.ai/tutorials/2020/09/15/deep-learning-flux.html), the part where we train a network on CIFAR10 takes longer than expected. Could it be because we actually go through every...

Gamma measure

6

Hey @cscherrer Here's the implementation of the Gamma measure based on the version from Distributions.jl, with both standard parametrizations from Wikipedia.

Categorical distribution

14

Partially solves #145. I did not add alternative (log-p) parametrizations yet. I think this one is a typical example of the limits of our approach, which was to default on...