How to save predictions as geojson?

[lcoandrade]

I have predictions obtained like this:

from rastervision.core.data import SemanticSegmentationLabels

predictions = predictlearner.predict_dataset(
    dataset=prediction_dataset,
    raw_out=False,# raw_out False because I want class IDs
    numpy_out=True,
    predict_kw=dict(out_shape=IMG_SIZE),
    progress_bar=True,
)

pred_labels = SemanticSegmentationLabels.from_predictions(
    windows=prediction_dataset.windows,
    predictions=predictions,
    smooth=False, # trying both outputs to see what happens
    extent=prediction_dataset.scene.extent,
    num_classes=len(class_config),
)

And I'm saving them like this:

pred_labels.save(
    uri=os.path.join(OUTPUT_DIR, 'predictions'),
    crs_transformer=prediction_dataset.scene.raster_source.crs_transformer,
    class_config=class_config,
)

How can I save my predictions as geojson?
I trying to use SemanticSegmentationLabelStore with VectorOutputConfig for each class (I have 13 classes) with this:

from rastervision.core.data import (
    SemanticSegmentationLabelStore,
    VectorOutputConfig
)

vector_configs = [VectorOutputConfig(class_id=i) for i in range(len(class_config))]    

label_store = SemanticSegmentationLabelStore(
    uri=os.path.join(OUTPUT_DIR, 'predictions'),
    extent=prediction_dataset.scene.extent,
    crs_transformer=prediction_dataset.scene.raster_source.crs_transformer,
    class_config=class_config,
    discrete_output=True,
    smooth_output=False,
    vector_outputs=vector_configs,
)

label_store.write_vector_outputs(pred_labels)

But this raises this error:

---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
Cell In[22], line 20
      7 print(vector_configs)
     10 label_store = SemanticSegmentationLabelStore(
     11     uri=os.path.join(OUTPUT_DIR, 'predictions'),
     12     extent=prediction_dataset.scene.extent,
   (...)
     17     vector_outputs=vector_configs,
     18 )
---> 20 label_store.write_vector_outputs(pred_labels)

File /opt/conda/lib/python3.10/site-packages/rastervision/core/data/label_store/semantic_segmentation_label_store.py:297, in SemanticSegmentationLabelStore.write_vector_outputs(self, labels)
    292 with tqdm(self.vector_outputs, desc='Vectorizing predictions') as bar:
    293     for i, vo in enumerate(bar):
    294         bar.set_postfix(
    295             dict(
    296                 class_id=vo.class_id,
--> 297                 mode=vo.get_mode(),
    298                 denoise_radius=vo.denoise))
    300         if vo.uri is None:
    301             log.info(f'Skipping VectorOutputConfig at index {i} '
    302                      'due to missing uri.')

File /opt/conda/lib/python3.10/site-packages/rastervision/core/data/label_store/semantic_segmentation_label_store_config.py:46, in VectorOutputConfig.get_mode(self)
     45 def get_mode(self) -> str:
---> 46     raise NotImplementedError()

NotImplementedError: 

[AdeelH]

You need to use PolygonVectorOutputConfig instead of VectorOutputConfig. You will also need to specify URIs. Here's a snippet: #1741 (comment).

[lcoandrade]

It worked! I just needed to remove some parameters to stay like this:

predictions_path = os.path.join(OUTPUT_DIR, 'predictions')
pred_labels.save(
    uri=predictions_path,
    crs_transformer=prediction_dataset.scene.raster_source.crs_transformer,
    class_config=class_config,
    vector_outputs=[
        PolygonVectorOutputConfig(class_id=i, uri=os.path.join(predictions_path, f'class-{i}.json')) 
        for i in range(len(class_config))
    ],
)

[lcoandrade]

One more question about the vector outputs.

What is the reason to use one PolygonVectorOutputConfig for each class? Why not use the ClassConfig and generate one geojson file with an attribute for the class?

[AdeelH]

I agree it's a little convoluted. I believe the motivation was to allow fine-grained control over how each class is vectorized and also to allow vectorizing only certain classes (e.g. it might not always make sense to create polygons for the background class).