A Keras implementation of FECNet, which proposed in "A Compact Embedding for Facial Expression Similarity"

The original paper is "A Compact Embedding for Facial Expression Similarity"

I implemented the original structure with Keras 2.2.0

Using Tensorflow1.3.0 backend

Training device: GTX 1060 6G

Due to the limitation of device, I didn't train the model with whole FEC dataset.

Dataset link: FEC

Download the FEC dataset to the data/

Run the image_extract.py frist, then run the export_train_label (If you are in China, a VPN is necessary. I strongly suggest you rent a oversea server to run the code and download those images, this will save you a lot of time.)

Run the FEC.py to train a model, or you can use the create_model method in you own training code. Like in Classifi2.py.

If your device is limite (like me), I suggest you set a very small learning rate (less than 0.0005) because the model will be very easy to overfit.

Update I'm not sure why I still couldn't get good result as the original paper. Maybe because the device limitation so that I could only set a small batch size. If anyone find the bug in my code, plz create a pull request.

I think there may have some problem in triplet loss part or read triplet images part, if you find it, please tell me, thanks.

If you like this project, offer me a star!