CVML / ABAW2-FPNMAA

Submission to ABAW2 Competition: Feature Pyramid Network for Multi-task Affective Analysis

ABAW2-FPNMAA

We participated in ICCV 2021: 2nd Workshop and Competition on Affective Behavior Analysis in-the-wild (ABAW). And more details can be found in our paper.

Requirement

We build the model on Pytorch 1.7.1 and use Pytorch Lightning framework.

conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 -c pytorch
pip install pytorch-lightning seaborn pretty_errors pandas PyYAML scikit-learn

Testing

download the dataset and create annotations

Please refer to the official website of ABAW for Aff-wild2 dataset. And this link for AffectNet, this link for ExpW. The final directory tree will be like this.

dataset/
├── AffectNet
│   ├── Manually_Annotated_file_lists
│   └── Manually_Annotated_Images
├── ExpW
│   ├── label.lst
│   ├── origin
│   └── readme.txt
├── Aff-Wild
│   ├── annotations
│   ├── cropped_aligned

1. open create_annotations.py files in separate dataset folders under create_annotation/single/ and change the path to dataset there.
2. run each create_annotations.py to get annotation for each dataset.
3. open create_annotation_file_Mixed_*.py files in separate task folders under create_annotation/mix/ and change the path to dataset there.
4. run each create_annotation_file_Mixed_*.py to get annotation for each task.

download the model and generate single task model

Please download the multi-task model from google drive.

1. fill in the model name in the ckpt_process.py.
2. put it in the folder where the model is.
3. run python ckpt_process.py.
4. you will get multi_va.ckpt, multi_expr.ckpt and multi_au.ckpt.

run test script

1. edit the dataset_dir in configuration files in configs/.
2. run test scripts. take va prediction for example.

python mono_fit.py --gpus 1 --config configs/train_va.yml --checkpoint /path/to/multi_va.ckpt 

Training

train single task model

take va training for example.

python mono_fit.py --gpus 1 --config configs/train_va.yml --train --max_epochs 20 --limit_train_batch 0.25 

generate soft label

take va label generation for example.

python gen_label.py  --gpus 1 --config configs/train_va.yml --checkpoint /path/to/single_va.ckpt

train multi task model

python multi_fit.py --gpus 1 --config configs/train_multi.yml --train --max_epochs 20 

Citation

If your work or research benefits from this repo, please cite the paper below.

@misc{he2021feature,
      title={Feature Pyramid Network for Multi-task Affective Analysis}, 
      author={Ruian He and Zhen Xing and Bo Yan},
      year={2021},
      eprint={2107.03670},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}