---

Classify-and-Extract for Android

Overview

This project demonstrates tag classification and Point of Interest (POI) extraction on Android devices, using optimized models for high-speed performance.

Getting Started

1. Download the Models:

   • The demo models are available for download here.

2. Setup:

   • After downloading, place the models into the assets folder.
   • Decompress the *.so zip file stored in the libs/arm64-v8a folder.

3. Model Information:

   • The demo model, named RexUniNLU, is converted from ModelScope and has undergone code optimizations for extreme execution speed.
   • The inputs and outputs of this demo model differ slightly from the original version.

4. ONNX Runtime Adaptation:

   • To better adapt to ONNX Runtime on Android, dynamic axes were not used during export. As a result, the exported ONNX model may not be optimal for x86_64 systems.
   • We plan to make the export method public in the future.

5. Performance:

   • A single inference takes about 42ms, and typically 3 rounds are required to complete classification and extraction tasks.

6. Limitations:

   • The REX model currently supports Chinese text exclusively.

Additional Resources

• Explore more projects: https://dakeqq.github.io/overview/

演示结果 Demo Results

---

标签分类和POI提取-安卓

概述

该项目在Android设备上展示了标签分类和兴趣点（POI）提取，使用经过优化的模型以实现高速度性能。

快速开始

1. 下载模型：

   • 演示模型已上传至云端硬盘：点击这里下载
   • 百度链接: 点击这里 提取码: dake

2. 设置：

   • 下载后，请将模型文件放入assets文件夹。
   • 解压存放在libs/arm64-v8a文件夹中的*.so压缩文件。

3. 模型信息：

   • 演示模型名为RexUniNLU，它是从ModelScope转换来的，并经过代码优化，以实现极致执行速度。
   • 因此，演示模型的输入输出与原始模型略有不同。

4. ONNX Runtime 适配：

   • 为了更好地适配ONNX Runtime-Android，导出时未使用动态轴。因此，导出的ONNX模型对x86_64可能不是最佳选择。
   • 我们计划在未来公开转换导出的方法。

5. 性能：

   • 一次推理大约需要42毫秒。通常需要3轮才能完成分类和提取任务。

6. 限制：

   • REX模型现在仅支持中文文本。

其他资源

• 看更多项目: https://dakeqq.github.io/overview/

---