Code for Noisy Injective Flows Paper
See requirements.txt for required pip packages, or use the following to download all dependencies
all dependencies:
pip install -r requirements.txtfashion-mnist, cifar10, and celebA were used in the paper. fashion-mnist and cifar10 are downloaded automatically using the dataloader found in datasets.py. CelebA must be downloaded seperately from https://drive.google.com/drive/folders/0B7EVK8r0v71pWEZsZE9oNnFzTm8
To train a model, first create a yaml file for the model architecture. All model's used in the paper have an existing yaml file. Once a yaml file exists, run the following:
python train_runner.py --name yamlfilename
To run experiments, run the following:
python plot_experiments.py --names model_name1 --experiment_root 'Experiments'
additional commandline arguments for running each experiment is as follows: --compare_baseline_samples --compare_vertical --compare_samples --compare_full_samples --reconstructions --vary_t --vary_s --compare_t --interpolations --best_s_for_nll --nll_comparison --save_embedding --plot_embeddings To run the plot_embeddings experiment, you must first run save_embedding seperately for the two model you wish to compare, for example:
python plot_experiments.py --name celeba_64n --experiment_root 'Experiments' --save_embedding
python plot_experiments.py --name celeba_glown --experiment_root 'Experiments' --save_embedding
To run plot_embeddings, on these two models, you would then run:
python plot_experiments.py --name celeba_64n celeba_glown --experiment_root 'Experiments' --plot_embeddings
To run the FID scores, you need a valid version of tensorflow (we used tensorflow-gpu version 1.15.2). This version is only supported in Python 3.7. There is a Dockerfile in the FID folder that can be used to create an environment to run the FID code.
To generate the FID scores, you must first run generate_images_for_fid.py on existing experiments. This will create a folder of images that will be fed into the FID code.
python generate_images_for_fid.py --names <experiment names>
Next, go into the FID folder and, using the correct environment, run evaluate_fid.py
python evaluate_fid.py --names <experiment names>