[DRAFT] 201-Export-onnx-model

[rojberr]

feat(): Export model to onnx and save in wandb as artifact

#201

[kapiblue]

We have new integration details after the meeting with Bartek:

1. We can implement our model as a Deepness Segmentor (see Fire Risk). The output should be a mask equal to the size of the input patch, where each pixel is associated with the class (one class for the whole patch)
2. Pay attention to ONNX metadata: https://qgis-plugin-deepness.readthedocs.io/en/latest/creators/creators_add_metadata_to_model.html
3. Deepness performs normalization by default.
4. QGIS provides 4 channels (near-infrared, red, green, blue). Channels for model inference can be selected in QGIS.
5. It is possible to perform inference on a selected polygon on the map (instead of the whole visible part). Our model does not currently support surfaces other than forests. We could incorporate class 'other' to the training set (houses, lakes, cars, etc.). But let's first deploy the model as it is.

Fire risk: