Data class refactor

[cgohil8]

Massive refactor of the Data class

Changes:

• This PR contains an API change for the Data.prepare method and the load_data wrapper in the config API.
• Standardisation no longer occurs by default when calling .prepare().
• Can now chain different preparation methods.
• Can easily add new (parallelised) methods for preparing the data.
• Added new methods for: downsampling, calculating amplitude envelopes, applying a sliding window, standardization, TDE and PCA.
• Can set the buffer size.
• Better naming of variables.
• Updated examples and tutorials for the new data preparation.
• We no longer support reading SPM files. OSL supports SPM to fif conversion, which can be read in osl-dynamics. (We can still read Matlab files.)

Preparation methods

We can now chain methods, e.g. for TDE-PCA data preparation we use:

data.tde_pca(n_embeddings=15, n_pca_components=80)
data.standardize()

# equvialently
data = data.tde_pca(n_embeddings=15, n_pca_components=80)
data = data.standardize()

# equivalently
data = data.tde_pca(n_embeddings=15, n_pca_components=80).standardize()

To prepare amplitude envelope data we use:

data.filter(low_freq=1, high_freq=45)
data.amplitude_envelope()
data.sliding_window(n_window=5)
data.standardize()

Tests confirm all new data preparation methods produce the same data as old code.

The Data.prepare method now takes a dict which contains methods to call in series. E.g.

methods = {
    "tde_pca": {"n_embeddings": 15, "n_pca_components": 80},
    "standardize": {},
}
data.prepare(methods)

(The order of the dict should be preserved from creation - added in Python 3.6+)

Saving and loading

Internal variables names were changed significantly. In particular, the Data.save method was changed significantly. However, the changes to saving and loading are backwards compatible, data saved with old code will still be readable.

Potential improvements

Rather than defining _apply functions in each method, we could see if we can use functools.partial or define a decorator for the functions in data.processing.

[evanr70]

To implement true chaining (data.method1(...).method2(...).method3(...)), we should return self from any methods for which it makes sense (like .standardize)

[evanr70]

I'll make a few operational changes to pipeline and rw, but excellent work overall