This is the GitHub repository for the code part of the Deep Learning final project on normalizing flows and their performance in generating novel anime faces.
Used dataset:
Used algorithms:
Density estimation using RealNVP by Dihn, L., Sohl-Dickstein, J. and Bengio, S. (2017) arXiv: 1605.08803
Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks by Radford, A., Metz, L. and Chintala, S. (2016) arXiv: 1511.06434
Code references and used repositories:
Glow implementation on Pytorch by Joost van Amersfoort (@y0ast) & Caleb Carr (@calebmcarr)
Removed due to consistent issues / concerns about implementation. However, the code was used during training of a Glow algorithm, and the results from that experiment were included in the final report - see glow_progression.jpg.
Implementation of selected normalizing flows on PyTorch by Ilya Kostrikov (@ikostrikov)
realNVP by Fangzhou Mu (@fmu2)
DCGAN by Nathan Inkawhich @inkawhich
Results are reported in the final report and the code demo. Samples from the training are available in the samples folder.
To reproduce project results, consult code demo file.
PATH refers to the dataset path - see an additional DATA_README for information on how to correctly import a dataset and about data prunning used. OUTPUT_PATH refers to the path of program outputs - by default, states are saved for all trained models as either realnvp_state{_optim}.pt or {generator,discriminator}_state{_optim}.pt.
For random-seed, 64 x 64 x 3 -> 4 x 4 x 48 training with 4 residual blocks/32 features per block:
python3 main.py --algo realnvp --epochs 25 --res-blocks 4 --base-dim 32 --datapath PATH --output-dir OUTPUT_PATH --nofixed
python3 main.py --algo realnvp --epochs 25 --res-blocks 4 --base-dim 32 --datapath PATH --output-dir OUTPUT_PATH --nofresh --saved-path OUTPUT_PATH/states --nofixed
python3 main.py --algo realnvp --epochs 25 --res-blocks 4 --base-dim 32 --datapath PATH --output-dir OUTPUT_PATH --nofresh --saved-path OUTPUT_PATH/states --nofixed
Note: due to the limitations of Google Colab free tier, the author has trained the models in several passes, which are reflected in the commands above.
For fixed seed, 64 x 64 x 3 -> 4 x 4 x 48 training with 4 residual blocks/32 features per block:
python3 main.py --algo realnvp --epochs 25 --res-blocks 4 --base-dim 32 --datapath PATH --output-dir OUTPUT_PATH --fixed-seed 409
python3 main.py --algo realnvp --epochs 25 --res-blocks 4 --base-dim 32 --datapath PATH --output-dir OUTPUT_PATH --nofresh --saved-path OUTPUT_PATH/states --fixed-seed 409
python3 main.py --algo realnvp --epochs 25 --res-blocks 4 --base-dim 32 --datapath PATH --output-dir OUTPUT_PATH --nofresh --saved-path OUTPUT_PATH/states --fixed-seed 409
For training 32 x 32 x 3 -> 16 x 16 x 6, 8 residual blocks/64 features per block, download files from commit bab504a, and use the following command:
python3 main.py --algo realnvp --epochs 50 --image-size 32 --datapath PATH --output-dir OUTPUT_PATH --nofixed
python3 main.py --algo realnvp --epochs 50 --image-size 32 --datapath PATH --output-dir OUTPUT_PATH --nofresh --saved-path OUTPUT_PATH/states --nofixed
python3 main.py --algo realnvp --epochs 50 --image-size 32 --datapath PATH --output-dir OUTPUT_PATH --nofresh --saved-path OUTPUT_PATH/states --nofixed
python3 main.py --algo realnvp --epochs 20 --image-size 32 --datapath PATH --output-dir OUTPUT_PATH --nofresh --saved-path OUTPUT_PATH/states --nofixed
!python3 dl-normalizing-flows/main.py --algo gan --epochs 500 --datapath PATH --output-dir OUTPUT_PATH --nofixed
For general usage, see -h or --help. Additional information is available in additional DATA_README file.