YOLOv5 inference using ONNX, with no complicated installations setup and zero precession loss!

• Google-Colab

• FPS-Accuracy-CPU

• FPS-Accuracy-GPU

Tested with Linux based systems (Colab T4/P4/K80, AWS t2.micro (Intel Xeon), Ubuntu-GTX 1650)

$ git clone https://github.com/BlueMirrors/Yolov5-ONNX.git
$ cd Yolov5-ONNX
$ pip install -r requirements.txt

Now run inference on video or image file (with pretrained weights).

python detect.py --input $PATH_TO_INPUT_FILE --output $OUTPUT_FILE_NAME

You can also pass --weights to use your own custom onnx weight file.

For pretrained default weights (--weights yolov5s), scripts will download the weights file automatically.

If you want to run the inference for your custom weights, simply do the following:

• Train Yolov5 on your custom dataset
• Export Weights PyTorch weights to ONNX

Make sure you use the ---dynamic flag while exporting your custom weights.

python export.py --weights $PATH_TO_PYTORCH_WEIGHTS --dynamic --include onnx

In our tests, ONNX had identical outputs as original pytorch weights.

Based on 500-700 inference iterations after 50 iterations of warmups. Includes Image Preprocessing (letterboxing etc.), Model Inference and Output Postprocessing (NMS, Scale-Coords, etc.) time only.

In our tests, ONNX had identical outputs as original pytorch weights.

Based on 5000 inference iterations after 100 iterations of warmups. Includes Image Preprocessing (letterboxing etc.), Model Inference and Output Postprocessing (NMS, Scale-Coords, etc.) time only.

• Batch support will be added next week (after 15th August)

• Yolov5
• CVU