SemanticSegmentationConfig's predict_options vs chip_options?

[NoaMillsUSDA-ARS]

Could you please help me understand the difference between the predict_options field and the chip_options field of the SemanticSegmentationConfig object? I'd like to specify the chip size, stride length, padding, and pad direction of the chips in the validation and test sets. Within the SemanticSegmentationChipOptions object, I can specify the stride length, padding, pad direction, and "size" which appears to be different from chip size because it allows for the resizing of images. Within the SemanticSegmentationPredictOptions object, I can specify the stride length and the chip size, but not the padding or pad direction. Do I need to specify some parameters in one object and other parameters in another?

[NoaMillsUSDA-ARS]

To add some more context, here's what my SemanticSegmentationConfig object looked like before I added stride and padding:

return SemanticSegmentationConfig(
        root_uri=output_uri,
        dataset=scene_dataset,
        backend=backend,
        predict_options=SemanticSegmentationPredictOptions(chip_sz=chip_sz))

And here's what it looks like after. My code runs, but I'm wondering what the affect is of having both the chip_sz field in SemanticSegmentationPredictionOptions, and the size field in SemanticSegmentationChipOptions both set, and if I have the stride parameter in the appropriate place:

return SemanticSegmentationConfig(
    root_uri=output_uri,
    dataset=scene_dataset,
    backend=backend,
    predict_options=SemanticSegmentationPredictOptions(chip_sz=chip_sz),
    chip_options=SemanticSegmentationChipOptions(
        sampling=WindowSamplingConfig(
            padding=padding,
            pad_direction="both",
            stride=stride_length,
            size=chip_sz
        )
    )
)

[AdeelH]

chip_options applies to the chip stage (which extracts chips into image files before the train stage) of the pipeline. If you are using GeoDatasets (via GeoDataConfig), then you are not using the chip stage, so chip_options has no effect. See the effect of the nochip param in this example.

Similarly, predict_options applies to the predict stage. It currently does not allow specifying padding and pad_direction; however, as of #2190 (not released yet), the padding value used by default is calculated such that pixels around the edges of the scene do not get left out of the prediction, even if crop_sz is specified.

It might make sense, in the future, to let PredictOptions take a WindowSamplingConfig just like ChipOptions and issue a warning if the user passes a random sampling config (because that's probably not what you want to do when predicting).