ECE 285 – MLIP – Project B Style Transfer
This is project B Style Transfer developed by team 996ICU composed of Qimin Chen, Yunfan Chen and Ke Xiao. Our group aim at implementing the style transfer by replicating the experiment proposed by Gatys et al. in 2015. We also implement the style transfer by using Cycle-GAN introduced by Jun Yan et al. Then we slightly modify the cycle consistency loss to force generator to focus more on the content structure in order to produce more realistic images.
Install package 'dominate' as: pip install dominate --user
Install package 'visdom' as: pip install visdom --user
If cuda runtime error (11) : invalid argument at /opt/conda/conda-bld/pytorch_1544174967633/work/aten/src/THC/THCGeneral.cpp:405
or any other cuda
issues happen, please restart the GPU cluster.
Neural Style Transfer -- Folder contains code for neural style transfer proposed by Gatys et al. demo.ipynb -- Run a demo of our code train.ipynb -- Run an example of our code utilis.py -- Module of model and dataset visu.py -- Module of visualize images and results Image-to-Image Translation using Cycle-GANs -- contains code for Cycle-GANs introduced by Jun Yan et al. data -- Folder contains dataset.py etc. model_checkpoints -- Folder contains trained cycle-gan model and trained improved cycle-gan model model -- Folder contains model.py and network.py etc. options -- Folder contains train and test configuration result -- Folder contains transferred results and recovered results scripts -- train and test scripts transferred_data_for_recover -- Folder contains transfered images for recovering util -- Folder contains utilities demo.ipynb -- Run a demo of our code test.py -- Test module train.py -- Train module utilis.py -- Utilities for visualization