You can download the fer2013 dataset from Kaggle. Each picture is 48x48 pixel grayscale images of faces.
After download, decompress it to directory datasets so we have ./datasets/fer2013/fer2013.csv
Since we are going to turn Keras model to TF Estimator, so use my modified version of keras_vggface model. We are importing keras from tensorflow package.
git clone https://github.com/Tony607/keras_vggface
cd keras_vggface
python3 ./setup.py install
Run the jupyter notebook as usual.
Post your comments and questions in my blog and stay tuned for future practiacl deep learning experiences. Happy coding!