Awesome

label-distribution-learning-fer-tf

Official Tensorflow Implementation of "Uncertainty-aware Label Distribution Learning for Facial Expression Recognition" paper

Requirements

python>=3.7

pip install tensorflow-gpu==2.4.0
pip install pandas opencv-python scikit-learn matplotlib
pip install git+https://github.com/qubvel/classification_models.git

Datasets

Download the RAF-DB dataset and extract the aligned folder (contains aligned faces) into data/rafdb/aligned We provide the pseudo valence-arousal as well as the pre-built K-nearest-neighbors for each instance in train.csv. The annotation file should have the following columns

The preprocessed data annotations are available at Data Annotation

Trained model

Coming soon...

Training

Download our pretrained backbone on MS-Celeb1M and put the .h5 files into ./pretrained folder: Google Drive

Run the following command to train the model on datasets

# Training model with resnet50 backbone
python src/train.py --cfg=config_resnet50_raf --train_data_path=data/rafdb/raf_train.csv --train_image_dir=data/rafdb/aligned 

# Training model with resnet18 backbone
python src/train.py --cfg=config_resnet18_raf --train_data_path=data/rafdb/raf_train.csv --train_image_dir=data/rafdb/aligned 

#Resume training from the latest checkpoint
python train.py --train_data=/path/to/train_data.csv --resume

Evaluation

Run this command to evaluate the trained model on the test set. We use the classification accuracy as our evaluation metric.

python src/eval.py --cfg=config_resnet50_raf --trained_weights=trained_weights/trained_resnet50_raf --test_data_path=data/rafdb/raf_test.csv --test_image_dir=data/rafdb/aligned