BREAKING NEWS: we have a docker image, complete with pretrained models+sketch. Get it here:
https://hub.docker.com/r/ellisk/graphics/
If you want to run on drawing 29 (drawings/expert-29.png) with pretrained models, do:
python demo.py 29
which will parse the image, and use the learned synthesis policy to synthesize a program from the ground truth parse as well as the parse discovered by the neural network (which hopefully should be the same!); it will also produce extrapolations and export them, printing out the files which contain extrapolations.
To generate synthetic training data for the neural network:
python makeSyntheticData.py 100000
which will generate 100000 training examples in to the file syntheticTrainingData.tar
To train the neural networks:
python recognitionModel.py train --noisy --attention 16 # trains the proposal distribution
python recognitionModel.py train --noisy --distance # trains the distance metric
To run the neural network on all of the images in a directory called drawings/, with 1000 particles, do:
python recognitionModel.py test -t drawings -b 1000 -l 0 --proposalCoefficient 1 --parentCoefficient --distanceCoefficient 5 --distance --mistakePenalty 10 --attention 16 --noisy --quiet
To use the program synthesizer you will need sketch:
https://people.csail.mit.edu/asolar/
I used sketch 1.7.5.
To run the program synthesizer on the 38th drawing (found in drawings/expert-38.png), do:
python synthesizer.py -f 38 # to pass the entire problem all at once the sketch
python synthesizer.py -f 38 --incremental # to break the problem up into pieces and pass each piece to sketch
To synthesize all of the programs for all of the drawings in every way possible and distribute the work across twenty CPUs, do:
python synthesizer.py --makePolicyTrainingData --cores 20
To view some extrapolations, do:
python synthesizer.py -n policyTrainingData.p --view --extrapolate
which should place its outputs into extrapolations/.
Released under GPLv3.