quibble: neural network code reformatting

Quibble will use neural networks to learn code formatting style, using a rigorous parsing / pretty printing layer to ensure the NN doesn't change the AST meaning.

Prerequisites

1. Install Torch Make sure to install torch following the instructions from torch.ch, not from someplace else. The installer scripts should work. If you want GPU support, make sure to install CUDA beforehand, and make sure the torch installer picks it up.

2. Follow the rest of the char-rnn install instructions

3. For the Haskell side, do

    # Put quibble in the same directory as char-rnn
    git clone git@github.com:eddysystems/quibble.git
   
    # Run one of
    sudo apt-get install ghc cabal-install
    brew install ghc cabal-install
   
    # Build quibble
    cd quibble
    cabal update
    cabal install happy alex
    cabal install --only-dependencies
    cabal configure
    cabal build
   
    # Run
    ln -s dist/build/quibble/quibble
    ./quibble -m <model> <javascript>
   

Sampling a checkpoint file

If there is a checkpoint file, run (in the char-rnn directory)

th sample.lua checkpoint-file -temperature 0.6 -gpuid -1 > output.js

The temperature can usefully be between 0 and 1. If the checkpoint file has been computed on a GPU, it has to be sampled on a GPU (-gpuid >= 0). Pass in a seed with -primetext "...".

Training

Use the instructions for char-rnn for training, except: You can restart training from a checkpoint using -checkpoint.