tomasmikeska / face-identification

One-shot face identification using deep learning

Face Identification

This is a Keras implementation of face identification model using ArcFace loss and Center loss.

Requirements

• Python 3.x
• pip

Installation and setup

Install pip packages using

$ pip install -r requirements.txt

Add .env file to project root with environmental variables

COMET_PROJECTNAME={comet_project_name}
COMET_WORKSPACE={comet_workspace}
COMET_API_KEY={comet_api_key}

Download VGGFace2 dataset for training to data/ folder here (images and bb_landmarks) or CASIA-WebFace dataset here, LFW dataset for validation here (funneled images and pairs.txt).

[optional]

There is a Docker image included that was used for training in cloud. You can build it from local Dockerfile with

docker build -t ml-box .

or get it from Docker Hub

docker pull tomikeska/ml-box

Usage

Train model using command

$ python src/train.py

After training the weights are saved to model/ folder by default. These weights contain all training layers (2 inputs, 2 outputs - softmax and centerloss) so in order to convert them to production model (single input and output) run command

$ python src/convert_model.py --weights model/densenet121_arcface_weights.h5 --nclasses 5386 -o model/densenet121_arcface_prod.h5

Evaluate on LFW using command

$ python src/lfw_validate.py --model model/densenet121_arcface_prod.h5

Training

Training took ~30 hours on NVIDIA P5000 on subset of VGGFace2 (~1M images, ~5K identities) depending on chosen base model.