- This page guides you in training YOLO-series models using PyTorch or Darknet and converting them into fully quantized TFLite models for easy deployment on MCU/MPU devices. It also includes support for the Vela compiler for ARM NPU devices.
- There are three models available for deploying your customized object detection models on Nuvoton MCUs.
| Model | Training Framwork | Int8 Full Quantization TFLite | Folder | Description |
|---|---|---|---|---|
| Yolo Fastest v1.1 | Darknet | ✔️ | yolo_fastest_v1.1 | |
| YoloX-nano | PyTorch | ✔️ | yolox_ti_lite_tflite_int8 | Some model updates have been made to improve the accuracy of quantized models. Please check the folder link for more details. |
| Yolov8-nano | PyTorch | ✔️ | yolov8_ultralytics | Some model updates have been made to enhance the performance of quantized models. Please check the folder link for more details. |
- Users can select models based on their application usage scenarios.
- The performance below is generated using the ARM Vela Compiler v3.10 and includes only model inference on the NPU. The mAP values are based on the COCO val2017 dataset and validation scripts for the TFLite INT8 model.
-
-
-
 vs mAP50.png)
 vs mAP50.png)
-
Yolo_fastest_v1.1
-
YoloX-nano/ Yolov8-nano
- The ML_SampleCode repositories are private. Please contact Nuvoton to request access to these sample codes. Link
- MCU: ML_M55M1_SampleCode (private repo)
- ObjectDetection_FreeRTOS_yoloxn/
- ObjectDetection_YOLOv8n/