Learning Representational Invariances for Data-Efficient Action Recognition

Official PyTorch implementation for Learning Representational Invariances for Data-Efficient Action Recognition. We follow the code structure of MMAction2.

See the project page for more details.

Installation

We use PyTorch-1.6.0 with CUDA-10.2 and Torchvision-0.7.0.

Please refer to install.md for installation.

Data Preparation

First, please download human detection results and put them in the corresponding folder under data: UCF-101, HMDB-51, Kinetics-100.

Second, please refer to data_preparation.md to prepare raw frames of UCF-101 and HMDB-51. (Instructions of extracting frames from Kinetics-100 will be available soon.)

(Optional) You can download the pre-extracted ImageNet scores: UCF-101, HMDB-51.

Training

We use 8 RTX2080 Ti GPUs to run our experiments. You would need to adjust your training schedule accordingly if you have less GPUs. Please refer to here.

Supervised learning

PORT=${PORT:-29500}

python -m torch.distributed.launch \
--nproc_per_node=8 \
--master_port=$PORT \
tools/train.py \
$CONFIG \
--launcher pytorch ${@:3} \
--validate

You need to replace $CONFIG with the actual config file:

• For supervised baseline, please use config files in configs/recognition/r2plus1d.
• For strongly-augmented supervised learning, please use config files in configs/supervised_aug.

Semi-supervised learning

PORT=${PORT:-29500}

python -m torch.distributed.launch \
--nproc_per_node=8 \
--master_port=$PORT \
tools/train_semi.py \
$CONFIG \
--launcher pytorch ${@:3} \
--validate

You need to replace $CONFIG with the actual config file:

• For single dataset semi-supervised learning, please use config files in configs/semi.
• For cross-dataset semi-supervised learning, please use config files in configs/semi_both.

Testing

# Multi-GPU testing
./tools/dist_test.sh $CONFIG ${path_to_your_ckpt} ${num_of_gpus} --eval top_k_accuracy

# Single-GPU testing
python tools/test.py $CONFIG ${path_to_your_ckpt} --eval top_k_accuracy

NOTE: Do not use multi-GPU testing if you are currently using multi-GPU training.

Other details

Please see getting_started.md for the basic usage of MMAction2.

Acknowledgement

Codes are built upon MMAction2.