This repo contains our solution for the 5th ABAW Challenge. We find that ContraWarping could extract informative expression features in this in-th-wild large-scale video database.

Install

This project is based on MMClassification, please refer to their repo for installation and dataset preparation.

Usage

First, download pre-trained weights from ContraWarping repo.

To fine-tune a Res-50 model with two GPUs, run:

PYTHONPATH=$(pwd):$PYTHONPATH \
python -m torch.distributed.launch --nproc_per_node=2 \
    tools/train.py configs/abaw5/ir50.py --launcher pytorch \
    --cfg-options load_from=weights/pretrained_weight.pth # pre-trained weight