• 2025/03/13: Face Detection added in onnx_inference.py.
• 2025/03/13: ONNX Export and Inference has been added.
• 2025/01/03: We released the Face-Recognition training framework and pretrained model weights.

git clone https://github.com/yakhyo/face-recognition.git
cd face-recognition
pip install -r requirements.txt

Codebase supports DDP, to run using DDP please use below example command:

torchrun --nproc_per_node=2 train.py --root data/train/ms1m_112x112 --database MS1M --network mobilenetv1 --classifier MCP

If you have a single GPU then use the below example command:

python train.py --root data/train/ms1m_112x112 --database MS1M --network mobilenetv1 --classifier MCP

To evaluate, please modify model, weights, and validation data filenames in evaluate.py

python evaluate.py

The following pretrained model weights are available for download under the release v0.0.1:

1. Download the model weights from the links above.
2. Place the weights in the desired directory (e.g., weights/).
3. Update your training or inference script to load the appropriate model weights.

The following ONNX model weights are available for download under the release v0.0.1:

1. Download the ONNX model weights from the links above.
2. Use the ONNX models directly for inference with ONNXRuntime.
3. These models are optimized for production deployment and cross-platform compatibility.
4. Run python onnx_inference.py to test ONNX model inference.

Run following command to export to ONNX:

python onnx_export.py -w [path/to/weight/file] -n [network/architecture/name] --dynamic[Optional]

Run onnx_inference.py to use ONNXRuntime. This inference calculates the similarity between two face images.

python onnx_inference.py

Run inference.py for PyTorch model inference. This inference calculates the similarity between two face images.

You can download aligned and cropped (112x112) training and validation datasets from Kaggle.

• CASIA-WebFace 112x112 from opensphere
  • Identities: 10.6k
  • #Images: 491k
• VGGFace2 112x112 from opensphere
  • Identities: 8.6k
  • #Images: 3.1M
• MS1MV2 112x112 from insightface
  • Identities: 85.7k
  • #Images: 5.8M

Validation data contains AgeDB_30, CALFW, CPLFW, and LFW datasets.

• AgeDB_30, CALFW, CPLFW, LFW 112x112

data/
|-- train/
|   |-- ms1m_112x112/
|   |-- vggface2_112x112/
|   |-- webface_112x112/
|-- val/
|   |-- agedb_30_112x112/
|   |-- calfw_112x112/
|   |-- cplfw_112x112/
|   |-- lfw_112x112/
|   |-- agedb_30_ann.txt
|   |-- calfw_ann.txt
|   |-- cplfw_ann.txt
|   |-- lfw_ann.txt

This project is licensed under the MIT License. See the LICENSE file for details.