YOLOX-small model trained on COCO

YOLOX-small is the small version of YOLOX model trained on COCO object detection (118k annotated images) at resolution 640x640. It was introduced in the paper YOLOX: Exceeding YOLO Series in 2021 by Zheng Ge et al. and first released in this repository.

We develop a modified version that could be supported by AMD Ryzen AI.

Model description

Based on YOLO detector, the YOLOX model adopts anchor-free head and conducts other advanced detection techniques including decoupled head and the leading label assignment strategy SimOTA to achieve state-of-the-art results across a large scale range of models. The series of models were developed by Megvii Inc. and won the 1st Place on Streaming Perception Challenge (WAD at CVPR 2021).

Intended uses & limitations

You can use the raw model for object detection. See the model hub to look for all available YOLOX models.

How to use

Installation

Follow Ryzen AI Installation to prepare the environment for Ryzen AI. Run the following script to install pre-requisites for this model.

pip install -r requirements.txt 

Data Preparation (optional: for accuracy evaluation)

The dataset MSCOCO2017 contains 118287 images for training and 5000 images for validation.

Download the validation set of COCO dataset (val2017.zip and annotations_trainval2017.zip). Then unzip the files and move them to the following directories (or create soft links):

└── data
     └── COCO
           ├── annotations
           |   ├── instances_val2017.json
           |   └── ...
           └── val2017
               ├── 000000000139.jpg
               ├── 000000000285.jpg
               └── ...

Test & Evaluation

• Code snippet from infer_onnx.py on how to use

args = make_parser().parse_args()
input_shape = tuple(map(int, args.input_shape.split(',')))
origin_img = cv2.imread(args.image_path)
img, ratio = preprocess(origin_img, input_shape)
if args.ipu:
    providers = ["VitisAIExecutionProvider"]
    provider_options = [{"config_file": args.provider_config}]
else:
    providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
    provider_options = None
session = ort.InferenceSession(args.model, providers=providers, provider_options=provider_options)
# NCHW format
# ort_inputs = {session.get_inputs()[0].name: img[None, :, :, :]}
# NHWC format
ort_inputs = {session.get_inputs()[0].name: np.transpose(img[None, :, :, :], (0, 2 ,3, 1))}
outputs = session.run(None, ort_inputs)
outputs = [np.transpose(out, (0, 3, 1, 2)) for out in outputs] # for NHWC format
dets = postprocess(outputs, input_shape, ratio)
if dets is not None:
    final_boxes, final_scores, final_cls_inds = dets[:, :4], dets[:, 4], dets[:, 5]
    origin_img = vis(origin_img, final_boxes, final_scores, final_cls_inds,
                     conf=args.score_thr, class_names=COCO_CLASSES)
mkdir(args.output_dir)
output_path = os.path.join(args.output_dir, os.path.basename(args.image_path))
cv2.imwrite(output_path, origin_img)

• Run inference for a single image

python infer_onnx.py -m yolox-s-int8.onnx -i Path\To\Your\Image --ipu --provider_config Path\To\vaip_config.json

Note: vaip_config.json is located at the setup package of Ryzen AI (refer to Installation)

• Test accuracy of the quantized model

python eval_onnx.py -m yolox-s-int8.onnx --ipu --provider_config Path\To\vaip_config.json

Performance

Metric	Accuracy on IPU
[email protected]:0.95	0.370

 @article{yolox2021,
  title={YOLOX: Exceeding YOLO Series in 2021},
  author={Ge, Zheng and Liu, Songtao and Wang, Feng and Li, Zeming and Sun, Jian},
  journal={arXiv preprint arXiv:2107.08430},
  year={2021}
}