Fork of deepcam-cn/yolov5-face

Differences between original repository and fork:

• Compatibility with PyTorch >=2.1. (🔥)
• Original pretrained models and converted ONNX models from GitHub releases page. (🔥)
• Installation with requirements.txt file.
• The wider_val.txt file for WIDERFace evaluation.
• The following deprecations and errors has been fixed:
  • UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument.
  • DeprecationWarning: 'np.float' is a deprecated alias for builtin 'float'.
  • FutureWarning: Cython directive 'language_level' not set.
  • Cython Warning: Using deprecated NumPy API.
  • AttributeError: module 'numpy' has no attribute 'int'.
  • RuntimeError: result type Float can't be cast to the desired output type long int.
  • Fixed face bounding box drawing problem in the TensorRT example.
  • NameError: name 'warnings' is not defined.

Installation

pip install -r requirements.txt

Pretrained models

Name	Easy	Medium	Hard	FLOPs(G)	Params(M)	Link
yolov5n-0.5	90.76	88.12	73.82	0.571	0.447	PyTorch, ONNX
yolov5n	93.61	91.52	80.53	2.111	1.726	PyTorch, ONNX
yolov5s	94.33	92.61	83.15	5.751	7.075	PyTorch, ONNX
yolov5m	95.30	93.76	85.28	18.146	21.063	PyTorch, ONNX
yolov5l	95.78	94.30	86.13	41.607	46.627
yolov5l (non-original)	95.63	94.06	85.49	41.607	46.627	PyTorch, ONNX

Note: Unable to download the YOLOv5l-Face model from the original repository due to an expired Google Drive link and a lack of access to Baidu Netdisk. So, the decision has been made to train the YOLOv5l-Face model from scratch. If someone have the original YOLOv5l-Face model, please create an issue with a new download link.

YOLOv5l-Face model training took about 10.57 hours using NVIDIA RTX 4090. Results can be found in the yolov5l-face.txt file

Inference

python detect_face.py --weights weights/yolov5s-face.pt --source data/images/bus.jpg --save-img

WIDERFace evaluation

• Download WIDERFace validation dataset.
• Move dataset to data/widerface/val directory.

python test_widerface.py --weights weights/yolov5s-face.pt --dataset_folder data/widerface/val/images

cd widerface_evaluate

python setup.py build_ext --inplace

python evaluation.py

Export to ONNX format

pip install onnx onnxruntime

python export.py --weights weights/yolov5s-face.pt

Export to TensorRT format

pip install tensorrt pycuda

python export.py --weights weights/yolov5s-face.pt --onnx2trt

TensorRT Inference

python torch2trt/main.py --trt_path weights/yolov5s-face.trt --img_path data/images/bus.jpg

PyTorch vs TensorRT speed comparison

python torch2trt/speed.py --torch_path weights/yolov5s-face.pt --trt_path weights/yolov5s-face.trt

Training

• Download WIDERFace training dataset.
• Download WIDERFace validation dataset.
• Download annotation files.
• Move WIDERFace training images WIDER_train/images to data/widerface/tmp/train/images.
• Move WIDERFace validation images WIDER_val/images to data/widerface/tmp/val/images.
• Move training annotation file train/label.txt to data/widerface/tmp/train/label.txt.
• Move validation annotation file val/label.txt to data/widerface/tmp/val/label.txt.

python data/train2yolo.py data/widerface/tmp/train data/widerface/train

python data/val2yolo.py data/widerface/tmp data/widerface/val

pip install tensorboard

• Start training:

python train.py --data data/widerface.yaml --cfg models/yolov5n-0.5.yaml

python train.py --data data/widerface.yaml --cfg models/yolov5l.yaml --weights weights/yolov5l.pt

• Resume training:

python train.py --data data/widerface.yaml --cfg models/yolov5n-0.5.yaml --resume

python train.py --data data/widerface.yaml --cfg models/yolov5l.yaml --resume