Flow Model

N-D Flow Model (RealNVP architecture)[]. This model was trained on a low-resolution (32 × 32) version of the CelebA-HQ dataset that has been quantized to 2 bits per color channel (due to lack of more powerful GPU).

Run

1. Install all dependencies listed in requirements.txt. Note that the model has only been tested in the versions shown in the text file.
2. Set following options:
   • name stands for name of the saved model (pt file)
   • lr stands for learning rate (default 5e-4)
   • epochs stands for total number of training epochs (100 is pretty good)
   • gpu stands for enabling CUDA (True/False)

cd src && python3 main.py --epochs 100 --gpu True

As far as the output, several plots will be saved in images directory (train plot + samples from the final trained model + interpolation images). Inetrpolation image consists of 5 rows of interpolations between real images in the test set (right and left in a row). 4 intermediate are interpolation ones.