Towards an Understanding of Our World by GANing Videos in the Wild
GitHub repository for "Towards an Understanding of Our World by GANing Videos in the Wild"
For more information please refer to our homepage.
Requirements
- Tensorflow 1.2.1
- Python 2.7
- ffmpeg
Data Format
Videos are stored as JPEGs of vertically stacked frames. Every frame needs to be at least 64x64 pixels; videos contain between 16 and 32 frames. For an example datasets see: http://carlvondrick.com/tinyvideo/#data
Training
python main_train.py
Important Parameters:
- mode: one of 'generate', 'predict', 'bw2rgb', 'inpaint' depending on weather you want to generate videos, predict future frames, colorize videos or do inpainting.
- batch_size: Recommended 64, for colorization use 32 for memory issues.
- root_dir: root directory of dataset
- index_file: must be in root_dir, containing a list of all training data clips; path relative to root_dir.
- experiment_name: name of experiment
- output_every: output loss to stdout and write to tensorboard summary every xx steps.
- sample_every: generate a visual sample every xx steps.
- save_model_very: save the model every xx steps.
- recover_model: if true recover model and continue training