face_sdk

Face SDK

A simple yet fully functional face SDK written in python, which acts as a demo to help the users learn the usage of each module and develop the further versions.

Requirements

• python >= 3.7.1
• pytorch >= 1.1.0
• torchvision >= 0.3.0
• yaml
• itertools
• skimage

Results by the SDK

Non-masked faces

Masked faces

Getting Started

Run the examples

python api_usage/face_detect.py  # Get bounding box of a face 
python api_usage/face_alignment.py  # Get 106 landmark of a face 
python api_usage/face_crop.py  # Get croped face from a image
python api_usage/face_feature.py  # Get features of a face
python api_usage/face_pipline.py # Run face recognition pipeline

The results will be saved at api_usage/temp

Update the models

• Step1: Train or finetune the new model and save it by the function torch.save.
• Step2: Create the configuration file model_meta.json according to the existing model.
• Step3: Put the model and the configuration file to the folder models/face_xxx/face_xxx_version. Here xxx refers to the corresponding function, e.g., detection, alignment and recognition.
• Step4: Modify the model name according to your own model in model_conf.yaml.
• Step5: Run the corresponding example.

Further development

To add a model trained by other method which is different with our sdk, or to add some modules which are not included in our sdk, you can implement with the following step.

• Step1: Train a model by your own method.
• Step2: Create the configuration file model_meta.json which should include the common configurations as mentioned in BaseModelLoader.py and the specific configurations.
• Step3: Put the model and the configuration file to the folder models/face_xxx/face_xxx_version. Here xxx refers to the corresponding function, e.g., detection, alignment and recognition.
• Step4: Create the network definition file in network_def.
• Step5: Create a model loader in model_loader, which should be inherited from the BaseModelLoader.
• Step6: Create a model handler in model_handler, which should be inherited from the BaseModelHandler.
• Step7: Write you own example in api_usage to show how to call the provided api.
• Step8: Run the new example.

About the logs

The logs in this sdk are configured by logging.conf.

• In the core module, we set logger = logging.getLogger('sdk'), which will print the detailed logs in logs/sdk.log for debugging.
• In the api_usage module, we set logger = logging.getLogger('api'), which will print the simple logs in the console.

Reference

This project is mainly inspired by Pytorch_Retinaface, chainer-ssd, pfld_106_face_landmarks, InsightFace, InsightFace_Pytorch