English | 简体中文

Introduction

This repo provides a Weight Conversion Tool which can be used to export a Yolov5 model (e.g., yolov5s.pt) to a CoreML model (e.g., yolov5s.mlmodel) with a decoding layer and an non maximum suppression layer (NMS).

Advantages

Compared to other tools, it:

• Can be used simply and quickly

  After setting up the environment, you just need to input a Yolov5 model with its parameters in convert.py to get the CoreML model.

• Support quantization

  In addition to the default obtained Fp32 model, Fp16 and Int8 quantization is also supported in our tool. Note: Quantization is supported on both MacOS and Linux while we haven't tested it on windows yet.

• Can be previewed on Xcode without coding

  We added a decoding layer and an NMS layer in comparison to the official weight conversion tools so that the exported model works nice with Apple version framework and can be previewed on Xcode without any coding.

Before You Start

Yolov5 is required for this tool, so clone Yolov5 repo first and install requirements.txt in a Python >= 3.8.0 environment.

git clone https://github.com/ultralytics/yolov5 # clone
cd yolov5
git checkout v6.2                    # (recommend)switch to v6.2 tag
pip install -r requirements.txt      # install environment

Then, clone this repo and install coremltools == 6.0 .

git clone https://github.com/ClintRen/ yolov5_convert_weight_to_coreml.git  # clone
cd yolov5_convert_weight_to_coreml
pip install coremltools==6.0         # install
pip install numpy==1.23.1            # **IMPORTANT** New version of numpy may cause a attribute error

Note: We recommend using yolov5 == 6.2.0 and coremltools == 6.0 as we tested.

Usage

Some of the parameters should be changed in convert.py to meet you needs:

Parameters	Explanations
--yolov5-repo	path to your yolov5 repo
--weight	yolov5 weights path
--img-size	image imput size (pixels)
--conf-thres	confidence threshold
--iou-thres	NMS IoU threshold
--device	cuda device, i.e. 0 or 0,1,2,3 or cpu
--quantize	quantize model to FP16 and Int8
--description	model description
--author	model author
--version	model version
--license	model license

Example

This command exports a pretrained Yolov5s model to a CoreML model. The CoreML models (yolov5s.mlmodel, yolov5s_FP16.mlmodel and yolov5s_Int8.mlmodel) will be generated in the weight directory.

python convert.py --yolov5-repo /path/to/yolov5 --weight yolov5s.pt --img-size 640 --quantize

Tip : add --quantize to export models at Fp16 and Int8 precision for smaller file sizes.

Preview on Xcode

After exporting the CoreML models, you can preview them on Xcode without coding.

The details and preview examples are shown below, please give a star if it helps you, thanks.