/video_decaptioning

Primary LanguagePythonMIT LicenseMIT

Chalearn Inpainting challenge Track 2

1st place on final test phase (challenge ID 'SanghyunWoo') https://competitions.codalab.org/competitions/18421#results

Final model overview

architecture

Preparation:

  1. Install Python 3.6.4 version and Pytorch 0.3.1.post3
  2. Install Dependencies
  • visdom (training loss curve visualization)
  • ffmpeg (video to png)
  1. Install pretrained weight

Brief code instruction:

  1. Extract png files for each mp4 videos (use video_png.py)
  2. Set root path (modify --root_path flag in scripts)
  3. Run the code using scripts
  • scripts/train.sh (for training the final model)
    • we trained for 200 epochs (about 3days using 2 gpus, GTX 1080 ti)
  • scripts/test.sh (for testing the final model)
    • 1~2 sec per video
  • Note that we attached pretrained weight of the final model at google drive.(final_model.pth) Please properly modify the path of pretrained weight in test.sh file for testing.