Python Openvino IR model

Read IR model

• Method 1

Problem: Recognize first image need more time (initalize time)

net.setPreferableBackend 參數
cv2.dnn.DNN_BACKEND_DEFAULT
cv2.dnn.DNN_BACKEND_INFERENCE_ENGINE
cv2.dnn.DNN_BACKEND_OPENCV

net.setPreferableTarget 參數
cv2.dnn.DNN_TARGET_CPU
cv2.dnn.DNN_TARGET_MYRIAD
cv2.dnn.DNN_TARGET_GPU

# Use Opencv read Openvino IR model
net = cv2.dnn.readNetFromModelOptimizer(xml_path, bin_path)

net.setPreferableBackend(cv2.dnn.DNN_BACKEND_INFERENCE_ENGINE)

# Use VPU(NCS)
net.setPreferableTarget(cv2.dnn.DNN_TARGET_MYRIAD)

# Output node name
layerName = net.getLayerNames()

# Input Image (dims == 4)
net.setInput(blob_img)

# forward
net.forward(layerName)

• Method 2

ie.load_network –> device_name 參數
CPU
GPU
MYRIAD

from openvino.inference_engine import IECore

# Read model and use openvino inference_engine to start
ie = IECore()

net = ie.read_network(model=model_xml, weights=model_bin)

# if want reuse network and define running device
net = ie.load_network(network=net, num_requests=2, device_name='CPU')

# 取得 placeholder
# Ready to input space
inputs = next(iter(net.input_info))

# 平行運算
# async run
net.start_async(request_id=request_id, inputs={inputs:blob_img})
if net.requests[request_id].wait(-1) == 0:
  pass
  
# 進行推論
result = net.infer(inputs={inputs: blob_img})

Preprocessing

# input default type: BGR
# output: [n, c, h, w]
blob = cv2.dnn.blobFromImage(img, size=(w, h))

Check input shape

for i in net.input_info:
  print(net.input_info[i].input_data.shape)
  
  # nchw
  print(net.input_info[i].input_data.layout)
  
  # name
  print(net.input_info[i].input_data.name)
  
  # FP32
  print(net.input_info[i].input_data.precision)

Show information

def showinfo():
    ie = IECore()
    print("Available devices:")
    for device in ie.available_devices:
        print("\tDevice: {}".format(device))
        print("\tMetrics:")
        print('getversion:', ie.get_versions(device))



        for metric in ie.get_metric(device, "SUPPORTED_METRICS"):
            try:
              metric_val = ie.get_metric(device, metric)
              print("\t\t{}: {}".format(metric, param_to_string(metric_val)))
            except TypeError:
              print("\t\t{}: UNSUPPORTED TYPE".format(metric))

        print("\n\tDefault values for device configuration keys:")
        for cfg in ie.get_metric(device, "SUPPORTED_CONFIG_KEYS"):
            try:
              cfg_val = ie.get_config(device, cfg)
              print("\t\t{}: {}".format(cfg, param_to_string(cfg_val)))
            except TypeError:
              print("\t\t{}: UNSUPPORTED TYPE".format(cfg))