Example 12: Retrain GTA V classifiers for all supported reiders#

ℹ️ Source code and input file(s) -> {pyppbox repo}/examples

Install pyppbox, pyppbox-data-gta5, and the FaceNet, Torchreid, ArcFace, and AdaFace module data wheels using 🚀 Getting Started. Running all five configurations requires Python 3.10+ and one face runtime requirements file: requirements/faces.txt for CPU or requirements/faces_gpu.txt for NVIDIA GPU. Each model needs its own classifier output; they are not interchangeable.

The Torchreid configurations use device: cuda; set both to cpu if needed. ArcFace and AdaFace default to cpu. For CUDA, install a compatible ONNX Runtime GPU build and change their device values. FaceNet uses TensorFlow’s automatic device selection; gpu_mem controls its GPU memory allocation. See Devices for runtime requirements.

Run from the repository’s examples/ directory:

python example_12_retain_gta5_reid_classifiers.py

The script overwrites five classifier .pkl files and their companion class-name .txt files under the installed pyppbox package. Change classifier_pkl to a writable output path with an existing parent directory to keep the supplied classifiers. Remove entries from the training loop to run only the reiders you have installed. Each backend is closed after training, including on failure.

Torchreid uses the GTA V body images; FaceNet, ArcFace, and AdaFace use the face images. Only the identity classifiers are trained; neural-network weights remain frozen. See Example 18: Recognize faces and train an ArcFace or AdaFace classifier for training and recognizing a face with an ArcFace/AdaFace JSON/YAML configuration.

Example source#

#################################################################################
# Example of retraining all supported ReID backends' GTA V identity classifiers
#################################################################################

import os
import pyppbox
from pyppbox.standalone import close, trainReIDClassifier


pyppbox_root = os.path.dirname(pyppbox.__file__)

gta5_osnet_ain_config = {
    'ri_name': 'Torchreid', 
    'classifier_pkl': f'{pyppbox_root}/data/modules/torchreid/classifier/gta5_osnet_ain_ms_d_c.pkl', 
    'train_data': f'{pyppbox_root}/data/datasets/GTA_V_DATASET/body_128x256', 
    'model_name': 'osnet_ain_x1_0', 
    'model_path': f'{pyppbox_root}/data/modules/torchreid/models/torchreid/osnet_ain_ms_d_c.pth.tar', 
    'min_confidence': 0.35, 
    'device': 'cuda'
}

gta5_osnet_config = {
    'ri_name': 'Torchreid', 
    'classifier_pkl': f'{pyppbox_root}/data/modules/torchreid/classifier/gta5_osnet_ms_d_c.pkl', 
    'train_data': f'{pyppbox_root}/data/datasets/GTA_V_DATASET/body_128x256', 
    'model_name': 'osnet_x1_0', 
    'model_path': f'{pyppbox_root}/data/modules/torchreid/models/torchreid/osnet_ms_d_c.pth.tar', 
    'min_confidence': 0.35, 
    'device': 'cuda'
}

gta5_facenet_config = {
    'ri_name': 'FaceNet', 
    'gpu_mem': 0.585, 
    'model_det': f'{pyppbox_root}/data/modules/facenet/models/det', 
    'model_file': f'{pyppbox_root}/data/modules/facenet/models/20180402-114759/20180402-114759.pb', 
    'classifier_pkl': f'{pyppbox_root}/data/modules/facenet/classifier/gta5.pkl', 
    'train_data': f'{pyppbox_root}/data/datasets/GTA_V_DATASET/face_182x182', 
    'batch_size': 1000, 
    'min_confidence': 0.75, 
    'yl_h_calibration': [-125, 75], 
    'yl_w_calibration': [-55, 55]
}

# ArcFace/AdaFace use their default preprocessing for the packaged ONNX models.
# For NVIDIA GPU inference, install requirements/faces_gpu.txt and use 'cuda'.
gta5_arcface_config = {
    'ri_name': 'ArcFace',
    'model_det': f'{pyppbox_root}/data/modules/arcface/models/det_10g.onnx',
    'model_file': f'{pyppbox_root}/data/modules/arcface/models/w600k_r50.onnx',
    'classifier_pkl': f'{pyppbox_root}/data/modules/arcface/classifier/gta5.pkl',
    'train_data': f'{pyppbox_root}/data/datasets/GTA_V_DATASET/face_182x182',
    'device': 'cpu'
}

gta5_adaface_config = {
    'ri_name': 'AdaFace',
    'model_det': f'{pyppbox_root}/data/modules/adaface/models/det_10g.onnx',
    'model_file': f'{pyppbox_root}/data/modules/adaface/models/adaface_ir18_vgg2.onnx',
    'classifier_pkl': f'{pyppbox_root}/data/modules/adaface/classifier/gta5.pkl',
    'train_data': f'{pyppbox_root}/data/datasets/GTA_V_DATASET/face_182x182',
    'device': 'cpu'
}


if __name__ == '__main__':
    # Remove entries to train only selected classifiers. Each output is overwritten;
    # pretrained neural-network weights remain unchanged.
    for reider in (gta5_osnet_ain_config, gta5_osnet_config, gta5_facenet_config,
                   gta5_arcface_config, gta5_adaface_config):
        try:
            trainReIDClassifier(reider=reider)
        finally:
            close()
On this page