Multiple classes in ClassConfig

[natedailey7]

I'm trying to input 2 classes into my Raster Vision model, but I can't figure out how to pass in my class ID from my geojson. For example, my geojson file class ID is named 'id', with the classes 0 and 1.

class_config = ClassConfig(
    names=['0', '1'],
    colors=['lightgreen', 'darkred'],
    )

Is there a place I can specify the field name 'id', so that Raster Vision knows where to look for the class ID?

[AdeelH]

You will need to use the ClassInferenceTransformer and pass it a class_id_to_filter (example usage). In your case, the class_id_to_filter should be something like class_id_to_filter={0: ['==', 'id', 0], 1: ['==', 'id', 1]}. In case it is unclear where to pass the transformer, please share more of your code and I can help you with that.

[natedailey7]

Thanks for your response! That makes sense, adding class_id_to_filter to a GeoJSONVectorSource. I'm not sure how to apply this to the ObjectDetectionSlidingWindowGeoDataset.from_uris() function.

That is, I want to add class_id_to_filter={0: ['==', 'id', 0], 1: ['==', 'id', 1]} to the following code:

class_config = ClassConfig(
    names=['0', '1'],
    colors=['lightgreen', 'darkred'],
    )

train_image_uri = "path"
train_label_uri = "path"
val_image_uri   = "path"
val_label_uri   = "path"

data_augmentation_transform = A.Compose([
    A.Flip(p=0.5),
    A.HorizontalFlip(p=0.5),
    ], bbox_params=DEFAULT_BBOX_PARAMS
)

train_ds = ObjectDetectionSlidingWindowGeoDataset.from_uris(
    class_config=class_config,
    image_uri=train_image_uri,
    label_vector_uri=train_label_uri,
    transform=data_augmentation_transform,
    size=400,
    stride=400,
)

I notice the parameter transform for ObjectDetectionSlidingWindowGeoDataset.from_uris(), however it seems that this is just for data augmentation transforms?

[AdeelH]

You need to pass it to the label_source_kw argument (see docs):

train_ds = ObjectDetectionSlidingWindowGeoDataset.from_uris(
    class_config=class_config,
    image_uri=train_image_uri,
    label_vector_uri=train_label_uri,
    label_vector_source_kw=dict(
        transformers=ClassInferenceTransformerConfig(
            class_id_to_filter={0: ['==', 'id', 0], 1: ['==', 'id', 1]})),
    transform=data_augmentation_transform,
    size=400,
    stride=400,
)

ClassInferenceTransformerConfig also accepts a default_class_id which is the catch-all ID for features in the GeoJSON whose class ID cannot be inferred by other means. If you set it to None (the default) such features will be ignored.

[natedailey7]

I'm getting the following error. Do you have any recommendations?

ValidationError: 1 validation error for GeoJSONVectorSourceConfig
transformers
  value is not a valid list (type=type_error.list)

This is after adding

    label_vector_source_kw=dict(
        transformers=ClassInferenceTransformerConfig(
            class_id_to_filter={0: ['==', 'id', 0], 1: ['==', 'id', 1]})),

To my ObjectDetectionSlidingWindowGeoDataset.from_uris()

[AdeelH]

Oops, the value passed to transformers needs to be a list:

    label_vector_source_kw=dict(
        transformers=[ClassInferenceTransformerConfig(
            class_id_to_filter={0: ['==', 'id', 0], 1: ['==', 'id', 1]}))],

[natedailey7]

That worked! Thank you so much! 🙇