A simple facial recognition library in Python leveraging Keras
This library is built and modelled after MM-DFR, which is a model of multiple CNNs which feed it's activations into a stacked auto encoder for further dimensionality reduction. The output is a 512-length vector which can be seen as the encoded representation of a given face.
Typically, you will instantiate the API object from api.py and call the train(model_name) method.
This method will persist the model to disk given a certain name that can be used to later retrieve the model.
To retrieve the encoded face representations you have to call load_model(model_name) which loads the models from disk.
Then you can call get_face_vector(image) as many times as you want.
There will be functions that will return a similarity score given some face vectors. It is up to the user to determine they're threshold for what score determines which faces are the same.
If you want to train the model with your own dataset, you can call add_image(user_id, image) and it will be persisted to disk.
The only caveat is you have to retrain the model.
remove_images() removes all images that the user has added through add_image.
The requirements.txt show which python libs are required.
You'll need keras 1.0.3, theano 0.8.1, numpy, scikit-image and dlib.
Note: If you don't install the exact versions of keras and theano the library might break
because models are persisted differently between versions.
The library assumes you're running on AWS g2.8xlarge.
If you're not adjust the compile time constants in consts.py from 4 GPUs to the number of GPUs on your machine.
Also we make certain assumptions about the name of the GPUs so this should also be adjusted.
Note: There may be other things you need to adjust.
We use NVIDIA's CUDA Deep Neural Network library (cuDNN) v3 and CUDA v7. We use these to be able to run our code on the GPUs, without these our code only runs on the CPUs. Make sure this is set up so our models train and score fast.
This library is memory intensive so one should setup swap space. The following guide is useful.