DNNGraph – A deep neural network model generation DSL in Haskell(github.com)
github.com
DNNGraph – A deep neural network model generation DSL in Haskell
https://github.com/ajtulloch/dnngraph
3 comments
Hi folks, author here - thanks for the interest in the project. It's literally ~1,000 lines of Haskell that I wrote a couple of weekends ago (and don't use at all in production), but LMK if you find it useful in any way (or have feature requests). Thanks!
This is really cool work. There is a lackluster theano feature which allows you to print a flowchart figure for the "computation graph" corresponding to the symbolic representation of your model.
@ajtulloch's library provides what I imagined the feature would be at first glance - a comprehensible, elegant graphical representation of your NN model. And on top of that, all in Haskell, with Haskell DSL for running torch - so cool.
@ajtulloch's library provides what I imagined the feature would be at first glance - a comprehensible, elegant graphical representation of your NN model. And on top of that, all in Haskell, with Haskell DSL for running torch - so cool.
Ok, how does this compare to machine learning tools in packages like R and the other api's out there.
I think the idea is it sets up Neural Networks that are then run in Torch, with some nice diagram generating tools. I don't know if that's something people will actually use, but it looks like a pretty concise way to generate a pretty complicated Neural Network, which could be a worthwhile idea considering how complex the more advanced ones are.
Ah I see. I guess Haskell wrapper for Torch with diagrams doesn't sound as impressive.
It's not exactly a wrapper, since you presumably still use the generated NN in LuaJIT. AI isn't my field, but it seems like a useful tool - actually setting up a complex neural network seems to be a lot of grunt work, comparatively speaking. There's probably a place for a tool that nicely abstracts over that part of the process.
This is a great service actually. Visualizing network architectures is good for teaching and discussing things. I use json in deeplearning4j and it's crazy hard to keep track of all the possible combinations of nets so it gets messy quick.
(Author here) - the interactive workflow at http://nbviewer.ipython.org/gist/anonymous/115cc761db1a0d290... might be interesting, FWIW.