Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,125 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: mit
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- ONNX
|
| 7 |
+
- ONNXRUNTIME
|
| 8 |
+
- AXEngine
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
## 简介
|
| 12 |
+
|
| 13 |
+
**PyAXEngine** 基于 cffi 模块实现了 Axera NPU Runtime 的 Python API,其 Python API 与 ONNXRuntime 高度兼(相)容(似),并同时支持开发板和M.2算力卡形态,方便开源社区开发者使用
|
| 14 |
+
Python 脚本快速构建 NPU 推理脚本
|
| 15 |
+
|
| 16 |
+
支持芯片
|
| 17 |
+
|
| 18 |
+
- AX650N
|
| 19 |
+
- AX630C
|
| 20 |
+
|
| 21 |
+
环境版本
|
| 22 |
+
|
| 23 |
+
- python >= 3.8
|
| 24 |
+
- cffi >= 1.0.0
|
| 25 |
+
- ml-dtypes >= 0.1.0
|
| 26 |
+
- numpy >= 1.22.0
|
| 27 |
+
|
| 28 |
+
*需要注意的是,如果您的开发环境是算力卡,那么更建议您优先考虑使用 [pyAXCL](https://github.com/AXERA-TECH/pyaxcl) 进行项目开发;pyAXCL 项目完整包含了算力卡形态的全部 API,更适合用于正式部署;PyAXEngine 项目更适合算法工程师进行快速原型验证,且用于计算卡环境时,PyAXEngine 不能调用编解码等模块(不是 PyAXEngine 的设计目标)。*
|
| 29 |
+
|
| 30 |
+
*AX650 SDK 2.18,AX620E SDK 3.12 以前的版本不支持 bf16,llm 模型会有返回 unknown 的 dtype问题,请注意升级*
|
| 31 |
+
|
| 32 |
+
*如果您评估认为不知道如何升级 SDK,也可以提交 issue 索要下载,不需要更新完整 SDK,只更新 libax_engine.so 即可*
|
| 33 |
+
|
| 34 |
+
## 快速上手
|
| 35 |
+
|
| 36 |
+
基于社区开发板 **爱芯派Pro(AX650N)** 进行展示
|
| 37 |
+
|
| 38 |
+
### 获取 wheel 包并安装
|
| 39 |
+
|
| 40 |
+
- [下载链接](https://github.com/AXERA-TECH/pyaxengine/releases/latest)
|
| 41 |
+
- 将 `axengine-x.x.x-py3-none-any.whl` 拷贝到开发板上,执行 `pip install axengine-x.x.x-py3-none-any.whl` 安装
|
| 42 |
+
|
| 43 |
+
### 简单示例
|
| 44 |
+
|
| 45 |
+
当前示例需要分别依赖 PIL 和 OpenCV,可以用 `pip install pillow opencv-python-headless` 安装。其中 `opencv-python-headless` 是 OpenCV 的 headless 版本,不依赖 GUI(非 headless 的版本需要依赖 OpenGL ES,运行环境中并没有)。
|
| 46 |
+
|
| 47 |
+
```python
|
| 48 |
+
将 [classification.py](https://github.com/AXERA-TECH/pyaxengine/blob/main/examples/classification.py) 拷贝到开发板上并执行。
|
| 49 |
+
|
| 50 |
+
```bash
|
| 51 |
+
root@ax650:~/samples# python3 classification.py -m /opt/data/npu/models/mobilenetv2.axmodel -i /opt/data/npu/images/cat.jpg
|
| 52 |
+
[INFO] Available providers: ['AXCLRTExecutionProvider', 'AxEngineExecutionProvider']
|
| 53 |
+
[INFO] Using provider: AxEngineExecutionProvider
|
| 54 |
+
[INFO] Chip type: ChipType.MC50
|
| 55 |
+
[INFO] VNPU type: VNPUType.DISABLED
|
| 56 |
+
[INFO] Engine version: 2.10.1s
|
| 57 |
+
[INFO] Model type: 0 (single core)
|
| 58 |
+
[INFO] Compiler version: 1.2-patch2 7e6b2b5f
|
| 59 |
+
------------------------------------------------------
|
| 60 |
+
Top 5 Predictions:
|
| 61 |
+
Class Index: 282, Score: 9.774
|
| 62 |
+
Class Index: 278, Score: 8.981
|
| 63 |
+
Class Index: 277, Score: 8.453
|
| 64 |
+
Class Index: 281, Score: 8.321
|
| 65 |
+
Class Index: 287, Score: 7.924
|
| 66 |
+
------------------------------------------------------
|
| 67 |
+
min = 0.890 ms max = 22.417 ms avg = 1.119 ms
|
| 68 |
+
------------------------------------------------------
|
| 69 |
+
```
|
| 70 |
+
|
| 71 |
+
示例也演示了如何选择计算设备:这意味着既可以在 **AX650/AX630C** 等开发板上运行,也可以在 AX650 M.2 算力卡上运行。
|
| 72 |
+
|
| 73 |
+
切换计算设备的方式是通过 `-p` 参数指定,如 `-p AxEngineExecutionProvider` 表示使用开发板上的 NPU 进行推理,而 `-p AXCLRTExecutionProvider` 表示使用 M.2 算力卡进行推理。
|
| 74 |
+
注意:在使用 M.2 算力卡进行推理时,需要将算力卡插入宿主机上,并且已经安装驱动,详见: [axcl](https://axcl-docs.readthedocs.io/zh-cn/latest/)。
|
| 75 |
+
|
| 76 |
+
```bash
|
| 77 |
+
root@ax650:~/samples# python3 classification.py -m /opt/data/npu/models/mobilenetv2.axmodel -i /opt/data/npu/images/cat.jpg -p AXCLRTExecutionProvider
|
| 78 |
+
[INFO] Available providers: ['AXCLRTExecutionProvider', 'AxEngineExecutionProvider']
|
| 79 |
+
[INFO] Using provider: AXCLRTExecutionProvider
|
| 80 |
+
[INFO] SOC Name: AX650N
|
| 81 |
+
[INFO] VNPU type: VNPUType.DISABLED
|
| 82 |
+
[INFO] Compiler version: 1.2-patch2 7e6b2b5f
|
| 83 |
+
------------------------------------------------------
|
| 84 |
+
Top 5 Predictions:
|
| 85 |
+
Class Index: 282, Score: 9.774
|
| 86 |
+
Class Index: 278, Score: 8.981
|
| 87 |
+
Class Index: 277, Score: 8.453
|
| 88 |
+
Class Index: 281, Score: 8.321
|
| 89 |
+
Class Index: 287, Score: 7.924
|
| 90 |
+
------------------------------------------------------
|
| 91 |
+
min = 1.587 ms max = 12.624 ms avg = 1.718 ms
|
| 92 |
+
------------------------------------------------------
|
| 93 |
+
root@ax650:~/samples# python3 classification.py -m /opt/data/npu/models/mobilenetv2.axmodel -i /opt/data/npu/images/cat.jpg -p AxEngineExecutionProvider
|
| 94 |
+
[INFO] Available providers: ['AXCLRTExecutionProvider', 'AxEngineExecutionProvider']
|
| 95 |
+
[INFO] Using provider: AxEngineExecutionProvider
|
| 96 |
+
[INFO] Chip type: ChipType.MC50
|
| 97 |
+
[INFO] VNPU type: VNPUType.DISABLED
|
| 98 |
+
[INFO] Engine version: 2.10.1s
|
| 99 |
+
[INFO] Model type: 0 (single core)
|
| 100 |
+
[INFO] Compiler version: 1.2-patch2 7e6b2b5f
|
| 101 |
+
------------------------------------------------------
|
| 102 |
+
Top 5 Predictions:
|
| 103 |
+
Class Index: 282, Score: 9.774
|
| 104 |
+
Class Index: 278, Score: 8.981
|
| 105 |
+
Class Index: 277, Score: 8.453
|
| 106 |
+
Class Index: 281, Score: 8.321
|
| 107 |
+
Class Index: 287, Score: 7.924
|
| 108 |
+
------------------------------------------------------
|
| 109 |
+
min = 0.897 ms max = 22.542 ms avg = 1.125 ms
|
| 110 |
+
------------------------------------------------------
|
| 111 |
+
```
|
| 112 |
+
|
| 113 |
+
## 社区贡献者
|
| 114 |
+
|
| 115 |
+
- [zylo117](https://github.com/zylo117): 提供了基于 cffi 的 AXCL Runtime Python API 实现
|
| 116 |
+
- [nnn](https://github.com/nnn112358),[HongJie Li](https://github.com/techshoww) 和 [Shinichi Tanaka](https://github.com/s1tnk) 报告 cffi 的使用问题,[Shinichi Tanaka](https://github.com/s1tnk) 提供了解决方案
|
| 117 |
+
|
| 118 |
+
|
| 119 |
+
## 关联项目
|
| 120 |
+
|
| 121 |
+
- [ax-samples](https://github.com/AXERA-TECH/ax-samples)
|
| 122 |
+
- [ax-llm](https://github.com/AXERA-TECH/ax-llm)
|
| 123 |
+
- [Pulsar2](https://pulsar2-docs.readthedocs.io/zh-cn/latest/)
|
| 124 |
+
- [AXCL](https://axcl-docs.readthedocs.io/zh-cn/latest/)
|
| 125 |
+
- [pyAXCL](https://github.com/AXERA-TECH/pyaxcl)
|