Adding metadata to ONNX model

The plugin allows you to load the meta parameters of the onnx model automatically. Predefined parameter types are designed for the simplicity of the user’s work. The written metadata is especially important for users who are not familiar with the technical details.

List of parameters parsed by plugin

Parameter	Type	Example	Description
model_type	str	'Segmentor'	Types of models available: Segmentor, Regressor, Detector.
class_names	dict	{0: 'background', 1: 'field'}	A dictionary that maps a class id to its name.
resolution	float	100	Real-world resolution of images (centimeters per pixel).
tiles_size	int	512	What size (in pixels) is the tile to crop.
tiles_overlap	int	40	How many percent of the image size overlap.
seg_thresh	float	0.5	Segmentor: class confidence threshold.
seg_small_segment	int	7	Segmentor: remove small occurrences of the class.
reg_output_scaling	float	1.0	Regressor: scaling factor for the model output.
det_conf	float	0.6	Detector: object confidence threshold.
det_iou_thresh	float	0.4	Detector: IOU threshold for NMS.
det_remove_overlap	bool	True	Detector: Whether overlapping detection should be removed

Example

The example below shows how to add string, float, and dictionary metadata into a model. Note that metadata is created while model.metadata_props.add() is called. Moreover, the metadata value has to be a byte type.

import json
import onnx

model = onnx.load('deeplabv3_landcover_4c.onnx')

class_names = {
    0: '_background',
    1: 'building',
    2: 'woodland',
    3: 'water',
    4: 'road',
}

m1 = model.metadata_props.add()
m1.key = 'model_type'
m1.value = json.dumps('Segmentor')

m2 = model.metadata_props.add()
m2.key = 'class_names'
m2.value = json.dumps(class_names)

m3 = model.metadata_props.add()
m3.key = 'resolution'
m3.value = json.dumps(50)

onnx.save(model, 'deeplabv3_landcover_4c.onnx')

You can also use a script setting the most important parameters in the plugin repository (file tools/add_model_metadata.py).