Use meta learning to estimate initial hyperparameters

[sumanthratna]

for PathFlow 2.0:

• better rotation detection
• better shape matching on the mixmatch side
• predict the model specification and hyperparameters to use based on the properties of the dataset

This way, we can suggest sane defaults for hyperparameters such as sigma, pyramid, regularisation_weight, etc.

[jlevy44]

One alternative to meta-learning could be a hyperparameter optimization method (eg. https://github.com/AIworx-Labs/chocolate https://github.com/scikit-optimize/scikit-optimize)