Closed srodney closed 3 years ago
After reading and thinking on this, let's keep it simple:
each PNG that is input to the CNN will contain info from a triplet of images. So I'll suggest that we store these as RGB values:
Using the keras preprocessing image_dataset_from_directory should then be sufficient to capture all of the info cleanly into separate classes, without needing a new Python object Class defined within diffimageML.
Each PNG needs to be labeled with a particular class label, which the CNNs will train on. We want this to be extensible, allowing for the possibility of adding new classes in the future.
Probably we want a class definition for this, and it should follow an existing schema that works with, say, the TensorFlow Keras preprocessing library.
See
https://www.tensorflow.org/api_docs/python/tf/keras/preprocessing/image_dataset_from_directory