ilkerkesen/draw

Knet implementation of DRAW: A Recurrent Neural Network For Image Generation

Hi there 👋

I am a researcher at Koç University, KUIS AI Center. Previously, I was a PhD student at Koç University, under the supervision of Deniz Yuret and Aykut Erdem. My main research area is natural language processing (NLP). In particular, I'm currently working on language-vision problems and Turkish NLP. The easiest way to contact me is to send an email. Here are some helpful links about me,

• Twitter
• Google Scholar
• LinkedIn

draw's Issues

gradient problem

@ilkerkesen

1. I implemented a new utility function gcheck (you need to checkout master AutoGrad and explicitly write using AutoGrad: gcheck to use it). This works with regular models and loss functions, e.g. gcheck(nll, model, x, y) where any of the inputs have param components should work.

2. I checked this on new versions of models like lenet (from the tutorial) and it works.

3. As expected it failed on your draw: include("debug.jl"); gcheck(loss,model,x).

4. Then I fixed the unboxing problem replacing Knet/src/rnn.jl:130 value.(hidden) with hidden.

5. I tried gcheck again but it still fails.

I was expecting (5) to pass. Did you try (4) and confirm the gradients work?