Face Recognition

Face recognition in PyTorch.

Installation

Create an environment (Python 3.10):

conda create -n face-recognition python=3.10

Install dependencies:

• For CPU:

  pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cpu

• For AMD GPU (for Windows users):

  pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cpu
  pip3 install torch-directml

• For NVIDIA GPU:

  pip install torch torchvision --index-url https://download.pytorch.org/whl/cuxxx

pip install opencv-python rich

Clone this repository:

git clone https://github.com/MarkIvory2973/face-recognition.git

Usage

cd face-recognition
python src/main.py train -d E:/Datasets -c ./checkpoints -e 40 -r 0.02 -s 16 -g 0.94
python src/main.py register -i 0 -c ./checkpoints -n Guest01
python src/main.py verify -i 0 -d E:/Datasets -c ./checkpoints -n Guest01

Parameters

Train mode:

Parameter	Required	Default	Description
--datasets-root,-d	✓	-	Datasets root
--checkpoints-root,-c	-	./checkpoints	Checkpoints root
--total-epoch,-e	-	100	Total epoch
--learning-rate,-r	-	0.02	Learning rate
--batch-size,-s	-	16	Batch size
--gamma,-g	-	0.98	The gamma of ExponentialLR

Register mode:

Parameter	Required	Default	Description
--camera-id,-i	-	0	Camera ID
--checkpoints-path,-c	-	./checkpoints	Checkpoints path
--username,-n	✓	-	Username

Verify mode:

Parameter	Required	Default	Description
--camera-id,-i	-	0	Camera ID
--datasets-root,-d	✓	-	Datasets root
--checkpoints-path,-c	-	./checkpoints	Checkpoints path
--username,-n	✓	-	Username

Result

⚠ The metrics in checkpoint.N.pth are based on the output of each 16 batch, so their Accuracy may be not correct enough (lower than actual).

I tested again with batch_size=1. Here is the result of checkpoint.best.pth (checkpoint.100.pth):

	Accuracy (%)
LFW	94.03

References

1. FaceNet: A Unified Embedding for Face Recognition and Clustering

2. Deep Residual Learning for Image Recognition