license: apache-2.0
YOLOv5: Target Detection
Yolov5 is a one-stage structure target detection network framework, in which the main structure consists of 4 parts, including the network backbone composed of modified CSPNet, the high-resolution feature fusion module composed of FPN (Feature Paramid Network), composed of SPP (Spatial Pyramid Pooling) constitutes a pooling module, and three different detection heads are used to detect targets of different sizes.
The YOLOv5 model can be found here
CONTENTS
Source Model
The steps followed the yolov5 tutorials to get the source model in ONNX format.
The source model YOLOv5s.onnx also can be found here.
Environment Preparation
git clone https://github.com/ultralytics/yolov5 # clone
cd yolov5
pip install -r requirements.txt # install
Export to ONNX
python export.py --weights yolov5s.pt --include torchscript onnx --opset 12
Performance
| Device | SoC | Runtime | Model | Size (pixels) | Inference Time (ms) | Precision | Compute Unit | Model Download |
|---|---|---|---|---|---|---|---|---|
| AidBox QCS6490 | QCS6490 | QNN | YOLOv5s(cutoff) | 640 | 6.7 | INT8 | NPU | model download |
| AidBox QCS6490 | QCS6490 | QNN | YOLOv5s(cutoff) | 640 | 15.2 | INT16 | NPU | model download |
| AidBox QCS6490 | QCS6490 | SNPE | YOLOv5s(cutoff) | 640 | 5.5 | INT8 | NPU | model download |
| AidBox QCS6490 | QCS6490 | SNPE | YOLOv5s(cutoff) | 640 | 13.4 | INT16 | NPU | model download |
| APLUX QCS8550 | QCS8550 | QNN | YOLOv5s(cutoff) | 640 | 4.1 | INT8 | NPU | model download |
| APLUX QCS8550 | QCS8550 | QNN | YOLOv5s(cutoff) | 640 | 13.4 | INT16 | NPU | model download |
| APLUX QCS8550 | QCS8550 | SNPE | YOLOv5s(cutoff) | 640 | 2.3 | INT8 | NPU | model download |
| APLUX QCS8550 | QCS8550 | SNPE | YOLOv5s(cutoff) | 640 | 5.8 | INT16 | NPU | model download |
| AidBox GS865 | QCS8250 | SNPE | YOLOv5s(cutoff) | 640 | 21 | INT8 | NPU | model download |
Model Conversion
Demo models converted from AIMO(AI Model Optimizier).
The demo model conversion step on AIMO can be found blow:
| Device | SoC | Runtime | Model | Size (pixels) | Precision | Compute Unit | AIMO Conversion Steps |
|---|---|---|---|---|---|---|---|
| AidBox QCS6490 | QCS6490 | QNN | YOLOv5s(cutoff) | 640 | INT8 | NPU | View Steps |
| AidBox QCS6490 | QCS6490 | QNN | YOLOv5s(cutoff) | 640 | INT16 | NPU | View Steps |
| AidBox QCS6490 | QCS6490 | SNPE | YOLOv5s(cutoff) | 640 | INT8 | NPU | View Steps |
| AidBox QCS6490 | QCS6490 | SNPE | YOLOv5s(cutoff) | 640 | INT16 | NPU | View Steps |
| APLUX QCS8550 | QCS8550 | QNN | YOLOv5s(cutoff) | 640 | INT8 | NPU | View Steps |
| APLUX QCS8550 | QCS8550 | QNN | YOLOv5s(cutoff) | 640 | INT16 | NPU | View Steps |
| APLUX QCS8550 | QCS8550 | SNPE | YOLOv5s(cutoff) | 640 | INT8 | NPU | View Steps |
| APLUX QCS8550 | QCS8550 | SNPE | YOLOv5s(cutoff) | 640 | INT16 | NPU | View Steps |
| AidBox GS865 | QCS8250 | SNPE | YOLOv5s(cutoff) | 640 | INT8 | NPU | View Steps |
Inference
Step1: convert model
a. Prepare source model in onnx format. The source model can be found here.
b. Login AIMO and convert source model to target format. The model conversion step can follow AIMO Conversion Step in Model Conversion Sheet.
c. After conversion task done, download target model file.
note: you can skip convert model step, and directly download converted model in Performance Sheet.
Step2: install AidLite SDK
The installation guide of AidLite SDK can be found here.