Normalize flag in WindowSamplingConfig

[bdevoghel]

🚀 Feature

When creating a SemanticSegmentation**GeoDataConfig** (maybe other use-cases than segmentaion too ? my experience is limited to this one), the scene_to_dataset creates a GeoDataset based on options from the sampling config. As WindowSamplingConfig has no normalize flag, the created dataset can only be created with normalize=True; which is undesired.

The requested feature is to create the normalize flag in the WindowSamplingConfig class and passing it down where required.

Motivation

When creating a SemanticSegmentation**GeoDataConfig**, I want my dataset to be sampled without built-in normalization.

Pitch

The class WindowSamplingConfig should have the additional normalize pydantic field.

Alternatives

Creating my own custom class similar to SemanticSegmentationGeoDataConfig but with the enforces normalize flag set to false.

[AdeelH]

Thanks for the issue and the PR! Can you say a bit more about your use case? Do you want to, say, feed in uint16 values type-casted directly to floats to the model? One existing way to do that is to attach a CastTransformer to your RasterSource. E.g.:

from rastervision.core.data import CastTransformerConfig

raster_source = RasterioSourceConfig(..., transformers=[CastTransformerConfig(to_dtype='float32')])

[bdevoghel]

For my use case I indeed use transformers for my RasterSource, however the SemanticSegmentationGeoDataConfig (used in the SemanticSegmentationLearnerConfig) creates (through scene_to_dataset()) a dataset SemanticSegmentation(Sliding|Random)WindowGeoDataset with normalize argument set to the default True; which is undesired behavior in my case.

I have custom transformers for my specific data and I don't want the data to be "normalized" (this happens in AlbumentationsDataset's __getitem__()

[AdeelH]

Right, I understand. What I meant was that that normalization only happens if the data is an unsigned integer type, so if the data is already type-casted to a float type (via the CastTransformer), it would not be affected even if normalize=True.

That said, I think exposing normalize via a Config makes sense. But I don't think WindowSamplingConfig is the best place for it, since that is supposed to be concerned only with the sampling of windows and not the transformation of data. What I would suggest is to:

• Add it DataConfig. That way it will be available in both GeoDataConfig and ImageDataConfig, both of which are based on AlbumentationsDataset. This will also require modifying dir_to_dataset() methods in *ImageDataConfigs, but feel free to skip that in this PR.
• Rename the field to normalize_uint (only in the Configs and not the Datasets) to better express what it is for.

Please also remember to sign the CLA.