Hi,
Thank you very much for providing the code! I am trying to reproduce the results provided in the paper for protocol 1. I used the processed dataset from the link provided in this repo.

I am using the following command for training for FSA_Net_capsule:
KERAS_BACKEND=tensorflow python FSANET_train.py --batch_size 16 --nb_epochs 90 --model_type 2 --db_name '300W_LP'..

I get MAE=5.687 for AFLW2000. I am expecting ~5.07 as shown in the paper. Am I using the wrong command or missing any hyperparameter setting?

Thank you!