add pixel size validators for satellite imagery semantic segmentation

[QazyBi]

It is better(hypothesis, soon will check) to train models on images where same pixel size is set. For that regard we could add some dataset validators which will run before training so that people can get insights about their data.

Some things to include:

• show distribution of object sizes(to catch outliers)
• show distribution of pixel scales(to catch outliers)
• show crs distribution among images
• check if image divisor is too big(save sample images after augmentations)

[QazyBi]

one more validator is following:

class DataAbnormalityDetector(pl.Callback:
    def __init__(self, metrics_deviation_max_value: float = 0.5, action: str = 'stop_run'): pass

• calculate metrics on all training images
• if any image has greater score devation from average than self.metrics_deviation_max_value then stop run or notify

possible actions:

• stop_run
• log error, log warning
• wait for n epochs and if it is still the same then log or stop run