Fix coreml quantization

[taka-wang]

Instead of importing platform, you can check it by sys.platform. By the way, prefix is a missing value in this script.

[AlexeyAB]

@taka-wang Thanks! Have you tried conversion and output using this fix?

[taka-wang]

Sure, it works on my MacOS 12.5 with python3.9.

[AlexeyAB]

@taka-wang Great! Did you use the same logic as in YOLOv5?

• https://github.com/ultralytics/yolov5/blob/master/models/common.py#L405-L408
• https://github.com/ultralytics/yolov5/blob/master/models/common.py#L484-L495

[AlexeyAB]

What latency and CPU/ANE ratio do you get for the yolov7.pt model and did you try int8 quantization? https://takawang.medium.com/execute-yolov7-model-on-ios-devices-bd9713834ad0

[taka-wang]

What latency and CPU/ANE ratio do you get for the yolov7.pt model and did you try int8 quantization? https://takawang.medium.com/execute-yolov7-model-on-ios-devices-bd9713834ad0

@AlexeyAB I follow the following two articles and update anchors accordingly.

• MobileNetV2 + SSDLite with Core ML
• yolov5 - CoreML Tools

• yolov7/cfg/baseline/r50-csp.yaml

  Line 7 in 9a00c17

  anchors:

---

Performance reports on these two devices, let me know if you need any further information, thanks.

• iPad Pro 5th gen 12.9 (M1)
• iPad Mini 6th gen (A15)
  
  

[AlexeyAB]

@taka-wang Thanks! Could you also show Latency and Compute Unit Mapping for YOLOv7-640-FP16?

[taka-wang]

@AlexeyAB

Here is model reports with NMS layers for v7_640_fp16 (around 3~4ms for nms from my reports).

For Compute Unit Mapping, maybe the upper diagram is more meaningful, thanks.

[AlexeyAB]

@taka-wang
So it's pretty fast, 20 FPS for YOLOv7-fp16 640x640, and mostly uses ANE (302 of 305 without NMS, and 335 of 364 with NMS).
If you would share the code or App in the AppStore with YOLOv7-fp16, I would post the link in this repository.

[taka-wang]

Ok, I will let you known when I have submit the app, thanks.

[taka-wang]

@taka-wang how did you get export to add the "preview" tab? Adding --include-nms does not seem to add preview or NMS for me.

@captncollins As far as I know, --include-nms option is only for onnx or tensorrt inference. You may refer to this article, MobileNetV2 + SSDLite with Core ML, to add your own nms function. My Xcode version is Version 14.0 beta 4.