the code for the ICML paper (Learning Neurosymbolic Generative Models via Program Synthesis) ]
Clone this directory. Make a directory called facades, make a directory inside their called data, and download the facades dataset to that directory.
To download the data for the facades dataset: Go to http://cmp.felk.cvut.cz/~tylecr1/facade/, extract the facade datasets and select only the .jpgs. Also download from the other links on that site.
To create a test-train split, run createtesttrain.py.
This will create a test/ folder and train/ folder in the facades directory.
To extract the full programs from all the files in the train dataset, first run gendiffmatstrainfull.py, and then run genprogtrainfull.py.
To extract the partially observed programs from all the files in the train dataset, first run gendiffmatstrainthird.py, and then run genprogtrainthird.py.
To learn a model from the partially observed programs to the full programs, run genprogtoprogtrain.py.
To render a partial extrapolated image using the generated generative model, run renderpartialprogtrain.py . This will generate a new folder in the facades folder called facadescycledata with two subdirectories: one called "trainA", and another called "trainB".
To extract the partially observed programs from all the files in the test dataset, first run gendiffmatstestthird.py, and then run genprogtestthird.py.
To use the learned model to render a partial extrapolated image for the test files, run renderpartialprogtest.py. This will create a new folder in the facadescycledata folder called testA.
Clone https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix, and cd into it. Copy the folder facades/facadescycledata into pytorch-CycleGAN-and-pix2pix/datasets.
In the pytorch-CycleGAN-and-pix2pix folder, create the CycleGAN model using the command python train.py --dataroot ./datasets/facadescycledata --name completion_cyclegan --model cycle_gan --display_id 0 --batch_size 2 --gpu_ids 0 --input_nc 3 --output_nc 3 --direction AtoB
Generate the final completions of the test data by, in the pytorch-CycleGAN-and-pix2pix folder, running python test.py --dataroot ./datasets/facadescycledata --name completion_cyclegan --model cycle_gan --input_nc 3 --output_nc 3