youshyee/CEP

This is the official implementation of paper Back to the Future: Cycle Encoding Prediction for Self-supervised Contrastive Video Representation Learning

CEP

Official implementation for paper "Back to the Future: Cycle Encoding Prediction for Self-supervised Video Representation Learning"

Getting Started

Install required packages

All dependencies can be installed using pip:

pip install -r requirements.txt

Datasets

Data for pre-training (Kinetics-400) and fine-tuning (UCF101, HMDB51).

After the Download, organise the dataset directory hierarchy is as follow:

├── data
    ├── kinetics400
    │   ├── train_video
    │   │   ├── answering_questions
    │   │   │   └── *.mp4
    │   │   └── ...
    │   └── val_video
    │       └── (same as train_video)
    ├── UCF101
    │   ├── ucfTrainTestlist
    │   │   ├── classInd.txt
    │   │   ├── testlist01.txt
    │   │   ├── trainlist01.txt
    │   │   └── ...
    │   └── UCF-101
    │       ├── ApplyEyeMakeup
    │       │   └── *.avi
    │       └── ...
    ├── hmdb51
    │   ├── metafile
    │   │   ├── brush_hair_test_split1.txt
    │   │   └── ...
    │   └── videos
    │       ├── brush_hair
    │       │   └── *.avi
    │       └── ...

Self-supervised pre-training

Modify the data path, workdir path in the config files and train model with,

# Arch: ResNet-18
sh dist_train.sh $GPU_NUM  configs/kinetics_r18_pretrain.py

# Arch: ResNet-34
sh dist_train.sh $GPU_NUM  configs/kinetics_r34_pretrain.py

# Arch: S3D-G
sh dist_train.sh $GPU_NUM  configs/kinetics_s3dg_pretrain.py

# Arch: Slowfast
sh dist_train.sh $GPU_NUM  configs/kinetics_slowfast_pretrain.py

# Arch: R(2+1)D
sh dist_train.sh $GPU_NUM  configs/kinetics_r2+1_pretrain.py

$GPU_NUM is the number of gpus used for distributed training

Training in Cluster

sbatch ./cep_slurm.sh