This is the code for the RSL pretext, that is used for self-supervised visual representation learning from videos. The trained models for the Ablation experiments are available at RSL-Ablation. The trained models for the Main experiments are available at RSL-Models.
- The dataset (UCF-101 and HMDB-51) main folder should have two sub-folders named (Videos, Splits). The Videos sub-folder contains all the videos of the dataset, while the Splits sub-folder contains the files that define the videos for each split.
- Each experiment should have a name that is configured using the parse_args function. In addition, there should be a folder with the same experiment's name inside the experiments folder. Each experiment folder should have a sub-folder named Run.
We used a workstaion that has a conda environment with the following packages:
- tqdm 4.64.1
- pandas 1.5.2
- python 3.8.15
- pytorch 1.13.0
- torchaudio 0.13.0
- tensorboardx 2.5.1
- torchsummary 1.5.1
- torchvision 0.14.0
- scikit-learn 1.1.3
- scikit-video 1.1.11
- ffmpeg-python 0.2.0
- opencv-python 4.6.0.66