@mlmodel attribute

dp.kinect3 dp.oak

signature

mlmodel PATH_STRING

examples

@mlmodel c:\mymodels\squeezenet.onnx
@mlmodel darknet-abc123.onnx

Enable machine learning and run inference with the given ONNX model on the colormap. Models supported include those that have a single input and that output numeric data (not strings).

Input into model

The color image from the sensor is the single input currently implemented.

1. Transcode into different formats using @mlinput.
2. Scale to the model’s tensor input size requirement without preserving aspect ratio. Use @mlcrop to override this and control the crop and resize.
3. Element type is checked/converted to the models’s tensor input type requirement, e.g. 8-bit unsigned int, 32-bit float, 16-bit float, etc. It prevents under/overflows. It does not normalize.
4. @mlmean is subtracted from the element values
5. @mlscale is multiplied to the element values

Output inference results

Model results can be a tensor, a sequence, or a map. Sparse tensor output is not supported.

These results are output on the 5th outlet as a Max dictionary. route and routepass are helpful to send it to other dict objects for post-processing.

This dictionary contains a named entry for each result tensor output by the model. Their names are defined by the model itself.

{
    "type_id": "dp-nd-tensor",    // n-dimensional tensor
    "sizes": [ 1, 1000, 1, 1 ],   // shape
    "dt": "f",                    // element type abbreviation
    "data": [ 1.49474326e-05, 2.23620050e-03, 1.65242789e-04,
        9.92372283e-04, 7.42915319e-04, 5.80585969e-04,
        ...
    ]
}

• type_id is always “dp-nd-tensor”
• sizes is the shape of the tensor with an entry for every rank and its value as the extent
• dt is the tensor element type abbreviation
• data contains the flattened element data of the tensor

Element type abbreviation

All except the last three element types are converted into the json dictionary.

Onnx Native Element Type	C++ Type	Abbreviation
ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT	float	f
ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8	uint8_t	u
ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8	int8_t	c
ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16	uint16_t	w
ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16	int16_t	s
ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32	int32_t	i
ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64	int64_t	j
ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING	std::string	t
ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL	bool	b
ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16	half	h
ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE	double	d
ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32	uint32_t	k
ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64	uint64_t	m
ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX64
ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX128
ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16