OpenVINO - IE - Python API

[junxnone]

基于 OpenVINO 2019.R3 文档

• Python API 是调用的 C++ API, 使用 cython 编译为 so 文件
• 使用 cmake/make 组织编译

Reference

• [ ] Python API
• [ ] Build Python API
• [ ] Overview of Inference Engine Python* API
• [ ] Python Samples

Python Classes

• IECore

  2019.R2 引入，与IEPlugin 有部分重复

• IENetwork
• ExecutableNetwork
• IEPlugin

IECore

• 添加 extension , 如 avx2 avx512 等

  ie = IECore()
  ie.add_extension(extension_path="/some_dir/libcpu_extension_avx2.so", device_name="CPU")

• 获取 metrics

  ie = IECore()
  ie.get_metric(metric_name="SUPPORTED_METRICS", device_name="CPU")

• 载入 network, network 由 IENetwork 定义

  net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
  ie = IECore()
  exec_net = ie.load_network(network=net, device_name="CPU", num_requsts=2)

• 查询 plugins

  net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
  ie = IECore()
  layers_map = ie.query_network(network=net, device_name="HETERO:GPU,CPU")

• 注册 plugin

  ie = IECore()
  ie.register_plugin(plugin="MKLDNNPlugin", device_name="MY_NEW_PLUGIN")

ie = IECore()
ie.register_plugins("/localdisk/plugins/my_custom_cfg.xml")

• 注销 plugin

  ie = IECore()
  plugin = IEPlugin("GPU")
  ie.register_plugin(plugin=plugin, device_name="MY_NEW_GPU")
  ie.unregister_plugin(device_name="GPU")

  IENetwork

• 载入 IR 文件
  • 可以是文件路径，也可以是文件内容

    net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)

• 设置 batch_size

  net.batch_size = 4

• 更改输入的 shape

  net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
  input_layer = next(iter(net.inputs))
  n, c, h, w = net.inputs[input_layer]
  net.reshape({input_layer: (n, c, h*2, w*2)}]

• Serializes the network and stores it in files. ？ 存为 IR 文件

  net = IENetwork(model=path_to_model, weights=path_to_weights)
  net.serialize(path_to_xml, path_to_bin)

ExecutableNetwork

• 获取执行 graph 信息

  net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
  plugin = IEPlugin(device="CPU")
  exec_net = plugin.load(network=net, num_requsts=2)
  exec_graph = exec_net.get_exec_graph_info()

• 获取 runtime metric

  ie = IECore()
  net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
  exec_net = ie.load_network(net, "CPU")
  exec_net.get_metric("NETWORK_NAME")

• sync inference

  net = IENetwork(model=path_to_xml_file, weights=path_to_bin_file)
  plugin = IEPlugin(device="CPU")
  exec_net = plugin.load(network=net, num_requests=2)
  res = exec_net.infer({'data': img})
  res
  {'prob': array([[[[2.83426580e-08]],
              [[2.40166020e-08]],
              [[1.29469613e-09]],
              [[2.95946148e-08]]
              ......
             ]])}

• aync inference

  infer_request_handle = exec_net.start_async(request_id=0, inputs={input_blob: image})
  infer_status = infer_request_handle.wait()
  res = infer_request_handle.outputs[out_blob_name]

Python API & C++ API

Python Samples

[junxnone]

junxnone/AI#63