/unofficial-FVC

Unofficial PyTorch implementation of "FVC: A New Framework towards Deep Video Compression in Feature Space"

Primary LanguagePythonBSD 3-Clause Clear LicenseBSD-3-Clause-Clear

Unofficial PyTorch implementation of FVC: A New Framework towards Deep Video Compression in Feature Space

This repository is built upon CompressAI platform.

Please note that only the feature-sapce DCN part is implemented.

TODO

  • implement the multi-frame fusion part
  • set GOP size in args
  • modify the visualization code of DCN offsets

Setup

conda create --name <env> --file requirements.txt

Run

Train

Run the command in the project root directory.

python train_video.py -d ${DATA_PATH} --epochs 100 --batch-size 4 --lambda 256 -lr 5e-5 --cuda --save

Evaluation

Run the command in the project root directory.

python eval_video.py checkpoint ${DATA_PATH} ${OUTPUT_DIR} -a fvc -p ${MODEL_PATH} --keep_binaries -v