5 implement mmd

[philswatton]

This PR implements the following:

• mmd function which takes two NP arrays, an embedding name, and a kernel name as input
  • internal dicts are used to transform strings to embedding functions and kernel functions
  • this is for compatability with using configs
• load_config to handle loading configs given a file path
  • atm it only loads a config for similarity metrics, but is set up for easy extension to e.g. dataset and experiment configs in the future
• tests for both of these, which use their own config file

This resolves #5 and contributes to #19.

I've also opened two new issues to further extend this initial implementation:

• MMD by class #20 suggests looking at implementing MMD by class (rather than MMD over the whole dataset) in the future, with arguments to handle this to be added to both config and mmd
• Embeddings #21 suggests looking at using feature embeddings as an additional hyperparameter to the MDD

EDIT: I also added some CIFAR bits to the .gitignore as suggested by @jack89roberts at some point somewhere

[jack89roberts]

Few initial comments whilst I wait for some things to run 🙂

[lannelin]

Looking good so far! Some fairly minor changes requested in comments.

[lannelin]

looks great, a couple of very minor comments then happy to merge without re-review.

Have run the following on my machine and returns a very small value for mmd (1.5625e-05) with rbf kernel. I imagine this is somewhat expected even with fairly large drop sizes?

from modsim2.data.loader import DMPair

metric_config = {'mmd_rbf': {'function': 'mmd', 'arguments': {'embedding_name': 'matrix', 'kernel_name': 'rbf'}}}

dmpair = DMPair(metric_config=metric_config, drop_percent_A=0.2, drop_percent_B=0.2)

dmpair.compute_similarity()

[jack89roberts]

Added a minor formatting comment but LGTM. Re James' last comment - maybe a sanity check/test that mmd with drop=0.8 is greater than with drop=0.2 (for example), would be helpful just to convince that it is working as expected.