The dataset I used is Face Place, see face_db.md to get more info.
The results I get are to large to upload, see models.md to get more info.
This repo is my graduation project, Face Recognition. The whole code I used is in code folder. I use Keras to design net architecture(ptorch is not finished, and will not be append), trian and analyse, for convnet visualization, I use Keras-vis, it is convient and helps a lot.
I experiment SGD, RMSProp and Adam on Vanilla CNN, and found Adam > RMSProp > SGD.
I experiment Vanilla CNN, ResNet and DenseNet, and found DenseNet > ResNet > Vanilla CNN.
I reduced the params of ResNet and DenseNet, found with #params reduced, epoch to achive 90% acc is increased.
I visualize convnet by Saliency and Grad-CAM, results is in code/analyse.
code/src mainly are raw python code of my experiment.
code/experiment mainly are jupyter notebook files of experiment results.
code/analyses mainly are some summary and visualizion result.
You can explore the code folder to find more specific info.
If you have any questions about this, please email me or just throw issues.