Script to generate training data for ERRNet

[tathey1]

In order to evaluate SmartEM networks more efficiently, I think there is a need to make a script that automatically generates training data for an ERRNET. This thread is to discuss specifics of the script. I plan on following the following specs:

Inputs:

• Weights to a EM2MB model
• Dataset with short dwell time images, and ground truth membrane maps

Ouput: A dataset of membrane prediction, error mask pairs where the membrane predictions are from the EM2MB model on short dwell time images, and the error masks are generated from the get_error_map function.

[tathey1]

@pythonLoader Any comments?

[tathey1]

I noticed a couple classes in Dataset.py such as PatchAugmentDatasetError but from what I can tell, these do not automatically generate "ground truth" error maps. Maybe I still should use these somehow to get what I want - but I'm not sure what they are for?

[cfpark00]

I agree, this class is needed. But I think it should not be on the fly as generating error maps can be slow, especially the scikit-image's watershed is slow.

[pythonLoader]

I guess the proposed class's objective is to create training data for ERRNET. So, it would be in the offline folder where we create the class.

I think Tommy's specs are good enough for the class. If anyone hasn't done that already, I can create and test one. Relevant to this, I have a few questions regarding the output of the get_error_map function. I will ping here once I get to it.