golbstein / keras-face-recognition

Who is your doppelgänger and more with Keras face recognition

Face Recognition and more with KERAS

Keras implementation of the paper: FaceNet: A Unified Embedding for Face Recognition and Clusterin

• Dataset:

  • CelebA Dataset
  • Link for downloading face images: img_align_celeba

• Dependencies

  • Keras 2 (tensorflow backend)
  • open-cv
  • tqdm
  • pandas

• Model

  • Feature extractor model: Xception
  • Embedding model: FaceNet

  

• Loss

  Triplet Loss with cosine similarity

  

  Where f_i is the embedding vector of image i

• Recognize celebrities with trained FaceNet model
• Find out your Doppelgänger
• Beauty test

keras-face-recognition's Issues

Carrie

Can your Git guess Carrie's phone number?
;)

Training loss doesn't decrease

Hi, I tried to run your notebooks. For notebook 2, my training loss is still around 0.19 after 200 epochs so the resulting model is still pretty bad. The only thing I changed is "use_multiprocessing" to False because it keeps getting an error when it's True.

Do you know how to fix this issue?

cannot create group in read only mode when running "load_model("siamese_xception.h5")"

cannot create group in read only mode when running "load_model("siamese_xception.h5")"