Implement checks for models
there are now some additional properties on TorchVGSLModel objects that allow some basic sanity checks:
one_channel_mode [None, '1', 'L'] -> if the input channels of the model are 1 this field indicates if the model has been trained on binary ('1') or grayscale ('L') data. It is set automatically during training (after 1 epoch). Models trained before yesterday will be assumed to have been trained on binary data and will have '1' in this field.
None
indicates that the model expects more than one input channel.seg_type [None, 'bbox', 'baselines'] -> Encodes the type of segmentation the model has been trained on. None is again for legacy. Has to be set manually before training.
model_type [None, 'recognition', 'segmentation'] -> Indicates if the model is a segmentation or recognition model. Legacy models have a value of
None
. TheTorchSeqRecognizer
class will now raise a ValueError if trying to load segmentation model. Has to be set manually during training.