Discriminating Against Unrealistic Interpolations in Generative Adversarial Networks

- Implemented add-on for ProGAN ("Progressive Growing of GANs for Improved Quality, Stability, and Variation" by Karras et al. (2017))

REPRODUCTION COMPABILITY SPECS:

• Ubunutu 16.04
• NVIDIA Tesla V100 32 GB GPUs with driver 465.19.01
• Conda local enviroment as per environment.yml (in essence; Cuda Toolkit version 9.0, and Tensorflow with GPU version 1.12)

USAGE:

This implementation may be used to find geodesics between user-selected points in the latent space, and corresponding images or videos from the output space. It can also be used to estimate statisics of the discriminator.

ENVIRONMENT SETUP:

We have set up a local python environment using conda on a computational machine, thus simplifying the installation of combatible packages, e.g., cudatoolkit and tensoflow-gpu. One only needs Nvidia drivers pre-installed. To set up this environment using conda, take these steps:

1. Clone this repository to your machine.
2. Install conda using apt-get or similar. We use version 4.9.2 here.
3. In terminal, from the local clone directory, create environment using the included environment.yml :

$ conda env create -f environment.yml

4. activate your enviroment by typing

$ conda activate progan_geodesics

RUNNING THE CODE:

The main file should be run from the repo folder with the command:

$ python main.py

Alternatively one can set up and run bash scripts from the same location with

$ bash bashScripts/name_of_bash_script

CONFIGURATIONS:

Current default values for configurations can be checked and adjusted in the beginning of configs.py above "CONFIGURATIONS END".

All default values can be overwritten from the command line by adding the correspoding flag. For example, to change the experiment_id and add an optional_run_id, use the command:

$ python main.py --experiment_id="my_new_experiment" --optional_run_id="test_1"

This will save the results in the folder results/my_experiment and all file names have a suffix of test_1

By default, the code calculates discriminator statistics, trains geodesics and creates a video. Each of these steps can be skipped by using the flags:

COMPUTE_STATISTICS_OFF
TRAIN_GEODESICS_OFF 
VIDEO_GENERATION_OFF 

For example, if you only want to train the geodesics, then use the command

$ python main.py --COMPUTE_STATISTICS_OFF --VIDEO_GENERATION_OFF

If you do not want to run all the default methods, then use the command

$ python main.py --methods "method_1" "method_2" "method_3"

REUSE_OF_DISCRIMINATOR_STATISTICS:

For each experiment_id, running the critic statistics saves the max and min value of the critic. If the same experiment_id is used again, the code loads these values in the following run during configuration, so these values are used if no new critic statistics are calculated.

This suggests the following usage: - Never use the same experiment_id for different models - Turn off critic statistics if the same model and experiment_id have been used before.

For example, if you want to run different start and endpoints for one model, use the commands

$ python main.py --start=1 --end=2 --experiment_id="my_experiment" --optional_run_id="seeds_1_2"
$ python main.py --start=3 --end=4 --experiment_id="my_experiment" --optional_run_id="seeds_3_4" --COMPUTE_STATISTICS_OFF 
$ python main.py --start=5 --end=6 --experiment_id="my_experiment" --optional_run_id="seeds_5_6" --COMPUTE_STATISTICS_OFF

RESULTS:

Computing critic statistics returns under results/experiment_id/statistics/ A sample grid of real images with critic values A sample grid of generated images with critic values A sample grid of generated images with the lowest found critic values A sample grid of generated images with the highest found critic values A histogramm comparing critic values of real and generated samples A numpy file (.npy) containing the estimated max and min value over all discriminator values

Training geodesics returns Images under results/experiment_id/images An image containing the geodesics for each method An image containing the critic values along the geodesics for each method An image containing the square differences along the geodesics for each method

Coefficients of the geodesics are saved under results/experiment_id/coefficient

Video generation returns A video showing all geodesics under results/experiment_id/videos An image of squared differences along the geodesic under results/experiment_id/images An image of critic values along the geodesic under results/experiment_id/images