Is it possible to see what config JSON fields are being mapped to the model’s frozen graphs (.pb, meta, .index, …)?
This could be achived by e.g. looking into what “nvmidl.apps.train” does?
A use case would be to understand what nodes in covid19 classification model’s Densenet121 are being mapped to which values (e.g. NV_MODEL_INPUT = CT numpy array, NV_IS_TRAINING = True, …) so it’s possible to approximately reproduce the training process outside of Clara train.
I was successful in importing the frozen graph and assigning values for three placeholders “NV_MODEL_INPUT”, “NV_PLACEHOLDERDEFAULT_TRAINING”, and “NV_IS_TRAINING” but I’m still missing a few as I’m getting an error as a few other (input?) nodes are not initialized.