https://thomasdougherty.github.io/pytorch-video-style-transfer/
PyTorch implementation of Element AI's Stabilizing neural style-transfer for video. Details can be found at my blog.
There are no extra compiled components in the repository and package dependencies are minimal, so the code is very simple to use. First, clone the repository locally:
git clone https://github.com/ThomasDougherty/pytorch-video-style-transfer.git
Next install the required dependencies:
pip install -r requirements.txt
Download and extract COCO 2014 train and val images with annotations from http://cocodataset.org. We expect the directory structure to be the following:
path/to/coco/
annotations/ # annotation json files
train2014/ # train images
val2014/ # val images
To train the model:
python main.py train --dataset path/to/coco/train2014/ --style-image path/to/style/style.png --save-model-dir ./ckpt/ --cuda 1
Video frames need to be extracted into a separate folder. Frames can be stylized without GPU by setting --cuda to 0.
python main.py eval --content-dir path/to/video_frames/ --model path/to/weight.model --output-dir save/styles/here --cuda 1