A TensorFlow implementation of ArcFace for face recognition.

Video Demo: TensorFlow Face Recognition Demo (Bilibili)

• Build with TensorFlow 2.4 Keras API
• Advanced model architecture: HRNet v2
• Build in softmax pre-training
• Fully customizable training loop from scratch
• Automatically restore from previous checkpoint, even epoch not accomplished.
• Handy utilities to convert and shuffle the training datasets.

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes.

For training

Additional packages for evaluation

Additional packages for inference with video

# From your favorite development directory
git clone --recursive https://github.com/yinguobing/arcface

You can use any dataset as long as they can be converted to TensorFlow Record files. If you do not have any dataset, please download one from the ArcFace official dataset list.

This project provides a demo file showing how to convert the downloaded MXNet dataset into TFRecord files. You can run it like this:

# From the project root directory
python3 -m utils.mx_2_tf

Most face recognition datasets contains millions of training samples. It is better to fully shuffle the data in the record file.

cd utils
python3 shard_n_shuffle.py

Do not forget setting a correct dataset file path.

Deep neural network training can be complicated as you have to make sure everything is ready like datasets, checkpoints, logs, etc. But do not worry. Following these steps you should be fine.

In the module train.py, setup your model's name.

# What is the model's name?
name = "hrnetv2"

These files do not change frequently so set them in the source code.

# Where is the training file?
train_files = "/path/to/train.record"

# How many identities do you have in the training dataset?
num_ids = 85742

# How many examples do you have in the training dataset?
num_examples = 5822653

These values varies with your dataset. Better double check before training.

# What is the shape of the input image?
input_shape = (112, 112, 3)

# What is the size of the embeddings that represent the faces?
embedding_size = 512

Set the hyper parameters in the command line. For softmax pre-training set --softmax=True. Otherwise the ArcLoss will be used.

# Softmax pre-training
python3 train.py --epochs=2 --batch_size=192 --softmax=True

# Train with ArcLoss
python3 train.py --epochs=4 --batch_size=192

Training checkpoints can be found in directory checkpoints. There is also another directory model_scout containing the best(max accuracy) model checkpoint. You get this feature for free.

Once the training was interrupted, you can resume it with the exact same command used for staring. The build in TrainingSupervisor will handle this situation automatically, and load the previous training status from the latest checkpoint.

# The same command used for starting training.
python3 train.py --epochs=4 --batch_size=192

However, if you want more customization, you can also manually override the training schedule with --override=True to set the global step, the epoch and the monitor value.

python3 train.py --epochs=4 --batch_size=192 --override=True

Use TensorBoard. The log and profiling files are in directory logs

tensorboard --logdir /path/to/arcface/logs

Even though the model wights are saved in the checkpoint, it is better to save the entire model so you won't need the source code to restore it. This is useful for inference and model optimization later.

Exported model will be saved in saved_model format in directory exported. You can restore the model with Keras directly.

python3 train.py --export_only=True

Once the model is exported, you can run an evaluation of the model with test datasets like LFW, etc.

python3 evaluate.py

After training for 84k steps(ignore the softmax pre-training steps) the prediction accuracy on LFW dataset is 0.9845 ± 0.0022. Following is the ROC curve figure.

Check the module code before running. It should not be difficult.

Once the model is exported, you can use predict.py to recognize faces. Please prepare some sample face images and set the paths in the python file. Then run

python3 predict.py --video /path/to/video.mp4

The most similar faces will be marked in the video frames. You can also use threshold to filter the results.

Yin Guobing (尹国冰) - yinguobing

• ArcFace: the official implementation in MXNet.
• InsightFace-tensorflow: Tensoflow implementation of InsightFace.
• Keras_insightface: Insightface Keras implementation