Highlights:

(1) This model is very light, only 4 convolutional layers and 2 FC layers (including the FC for softmax loss).

(2) Recognition accuracy (pretrained on AffectNet database): 73.36 at FER2013 (test set), 88.91 at FERPlus (test set), 88.92 at RAF-DB.

(3) Very speed (9.0 s per epoch for training on RAF-DB) and lowest resoure requirments.

Note that:

Hyperparams using the default values in main.py

Requires setup the package for Gabor Convolutional networks https://github.com/jxgu1016/Gabor_CNN_PyTorch

The RAF-DB can be found at http://www.whdeng.cn/RAF/model1.html#dataset

FER2013 and FERPlus can be found at Link：https://pan.baidu.com/s/1265rT59qoUW7AQkaV9DobQ password：1111