DeepStream SDK FAQ

Accelerated Computing Intelligent Video Analytics DeepStream SDK
52

42. How to convert NvDsInferTensorMeta output data to numpy ndarray

1.Get NvDsInferTensorMeta from user meta

tensor_meta = pyds.NvDsInferTensorMeta.cast(user_meta.user_meta_data)

# Boxes in the tensor meta should be in network resolution which is
# found in tensor_meta.network_info. Use this info to scale boxes to
# the input frame resolution.
layers_info = []

for i in range(tensor_meta.num_output_layers):
    layer = pyds.get_nvds_LayerInfo(tensor_meta, i)
    layers_info.append(layer)

2.Find the specified output layer

def layer_finder(output_layer_info, name):
    """ Return the layer contained in output_layer_info which corresponds
        to the given name.
    """
    for layer in output_layer_info:
        # dataType == 0 <=> dataType == FLOAT
        if layer.dataType == 0 and layer.layerName == name:
            return layer
    return None

3.Convert the tensor buffer of the output layer to numpy.ndarray

def layer_tensor_to_ndarray(layer: pyds.NvDsInferLayerInfo) -> np.ndarray:
    import ctypes
    if layer.dataType == pyds.NvDsInferDataType.FLOAT:
        # print(f"int_addr {type(pyds.get_ptr(layer.buffer))}") <class 'int'>
        addr = pyds.get_ptr(layer.buffer)
        if addr == 0:
            print("Buffer address is 0")
            return None
        # print(f"addr {type(addr)}") <class 'ctypes.c_int'>
        data_ptr = ctypes.cast(addr, ctypes.POINTER(ctypes.c_float))
        num_dims = layer.inferDims.numDims
        shape = []
        for i in range(num_dims):
            shape.append(layer.inferDims.d[i])
        # print(f"{shape}")
        layer_array = np.ctypeslib.as_array(data_ptr, shape=shape)
        # print(f"layer_array {type(layer_array)}")
        layer_ny = np.frombuffer(layer_array, dtype=np.float32)
        # print(f"boxes {type(layer_ny)}")
        return layer_ny
    return None