This is my personal exercises that use fer2013 datasets with Tensorflow to classify face emotion. I used vgg and resnet respectively,but all perform poor.In the test datasets,two model achieve accuracy of 54% and 48%.

Except the reasons I do incorrectly, there are several problems with this datasets:

1. Pixel is too low
2. Have a front face, side face, cartoon face and so on. Some kinds of faces that I can't tell what.
3. Data is not a face.

The pre-trained model be put into model/ folder.

Vgg:

Loss

Accuracy

Resnet:

Loss

Accuracy