This is the code I wrote for my Master's thesis. This is a public branch, meaning it's stripped of all the bigger files (datasets, models, summaries, outputs). Thesis abstract describing the experiments is in the pdf file.
What it does: Code for training diffrent versions of GANs with few different architecture for Generator and Discriminator The code is built on a fork of this repository. Inspiration for the code style/structure is taken from here and here. Code is written in Python3 and has no python2 support at the moment.
Install python dependencies (virtual environment recommended):
pip install tqdm colorama sklearn scipy Pillow matplotlib
TensorFlow installation assumed. This code was tested on TF 1.4 and 1.5
Most of the training parameters are set in the flags.py file.
The codebase is designed to train many models in a row, the hiperparameter set can be tinkered with in the file mix_and_match.py
The architectures for G and D are defined in file DCGANS.py
To start training all the models with hyperparameters specified (one by one, in sequence), run:
python3 mix_and_match.py
It trains many architectures in a row by spawning separate processes to avoid TensorFlow's complaints about variables being on the same graph.
Some files are obsolete, as the main goal of the codebase changed with time.
======================================
- Python, NumPy, TensorFlow, SciPy, Matplotlib
- A recent NVIDIA GPU
- Datasets in ./data folder. They must be in tfrecords format, resolutions specified in flags.py
- Specify desired hyperparameters in mix_and_match.py. Archietures can be changed in DCGANS.py
- Wait for training to finish.
- Run score_avg_test.py to calculate mean offsets (real_fake_half.npz).
- Run half_real_fake_stats_proc.py
- Run real_images_test.py
- Run mix_and_match_test.py to see which discriminators were fooled
- Generate figures
- Generate LIME explanations
Final check before generating all stat figures: Need 4 .npz files:
- evaluation_stats.npz (by mix_and_match_test.py)
- mean_offsets.npz (by half_real_fake_stats_proc.py)
- real_evaluation_stats.npz (by real_images_test.py)
- real_fake_half.npz (by score_avg_test.py)
fid.py in minst test is the most up to date - fixed inefficient loading.
same with mix_and_match_test.py -- fixed memory hogging.
mnist test has MMD hanging problem fixed.