# 4.3 Python编程详解

BMNNSDK通过SAIL库向用户提供Python编程接口。

SAIL (Sophon Artificial Intelligent Library) 是 Sophon Inference 中的核心模块。SAIL 对 BMNNSDK 中的 BMLib、BMDecoder、BMCV、BMRuntime 进行了封装，将 BMNNSDK 中原有的“加载 bmodel 并驱动 TPU 推理”、“驱动 TPU 做图像处理”、“驱动 VPU 做图像和视频解码”等功能抽象成更为简单 的 C++ 接口对外提供；并且使用 pybind11 再次封装，提供简洁易用的 python 接口。 

这个章节将会选取BMNNSDK2中的SSD检测算法作为示例， 来介绍python接口编程。

{% hint style="info" %}
**样例代码路径：**${BMNNSDK}/examples/SSD\_object/py\_ffmpeg\_bmcv\_sail
{% endhint %}

目前，SAIL 模块中所有的类、枚举、函数都在“sail”命名空间下，核心的类包括：

* Handle：BMNNSDK中BMLib的bm\_handle\_t的包装类，设备句柄，上下文信息，用来和内核驱动交互 信息。
* Tensor：BMNNSDK中BMLib的包装类，封装了对device memory的管理以及与system memory的同步。
* Engine：BMNNSDK中BMRuntime的包装类，可以加载bmodel并驱动TPU进行推理。一个Engine实例可以加载一个任意的bmodel，自动地管理输入张量与输出张量对应的内存。
* Decoder：使用VPU解码视频，JPU解码图像，均为硬件解码。
*  Bmcv：BMNNSDK中BMCV的包装类，封装了一系列的图像处理函数，可以驱动 TPU 进行图像处理。

其他关于接口的更相信的信息，请阅读${BMNNSDK}/documents/Sophon\_Inference\_zh.pdf。

本章主要介绍以下三点内容：

* 加载模型
* 预处理
* 推理

## 4.3.1 加载模型

```python
import sophon.sail as sail
engine = sail.Engine(tpu_id = 0)
engine.load(bmodel_path)
# engine = sail.Engine(bmodel_path,tpu_id,mode)
```

## 4.3.2 预处理

```python
class PreProcessor:
  def __init__(self, bmcv, scale):
    self.bmcv  = bmcv
    self.ab    = [x * scale for x in [1, -123, 1, -117, 1, -104]]

  def process(self, input, output):
    tmp = self.bmcv.vpp_resize(input, 300, 300)
    self.bmcv.convert_to(tmp, output, ((self.ab[0], self.ab[1]), (self.ab[2], self.ab[3]), (self.ab[4], self.ab[5])))

bmcv = sail.Bmcv(handle)   #图形处理加速模块
scale = engine.get_input_scale(graph_name, input_name)
pre_processor = PreProcessor(bmcv, scale)  #预处理初始化

img0 = decoder.read(handle)            #解码视频输出image
img1 = bmcv.tensor_to_bm_image(input)  #将推理的输入地址挂载到image

pre_processor.process(img0, img1)      #预处理
```

## 4.3.3 推理

```python
graph_name = engine.get_graph_names()[0]
engine.set_io_mode(graph_name, sail.IOMode.SYSO)

input_name   = engine.get_input_names(graph_name)[0]
output_name  = engine.get_output_names(graph_name)[0]

input_shape  = [1, 3, 300, 300]
output_shape = [1, 1, 200, 7]

handle = engine.get_handle()

input_dtype  = engine.get_input_dtype(graph_name, input_name)
output_dtype = engine.get_output_dtype(graph_name, output_name)

input  = sail.Tensor(handle, input_shape,  input_dtype,  False, True)
output = sail.Tensor(handle, output_shape, output_dtype, True,  True)

input_tensors  = { input_name:  input  }
output_tensors = { output_name: output }
...
#此处省略 解码，预处理 代码
...
engine.process(graph_name, input_tensors, output_tensors) #推理
out = output.asnumpy()

dets = post_processor.process(out, img0.width(), img0.height())  #后处理
...
```


---

# Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://sophgo-doc.gitbook.io/bmnnsdk2-bm1684/alg-optimize/python.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.