Default arguments for xarray_filters.datasets.make_* functions

[PeterDSteinberg]

@gpfreitas This is related to issue #5 and #6 and tries to condense them into a TODO list.

Items to do related to the argument specs of make_* functions from xarray_filters.datasets:

• [x] Make MLDataset be the default return value rather than Dataset
• [x] Remove the requirement for the n_samples argument in this case: MLDataset(make_blobs(n_samples=2000, shape=(200,10))) where n_samples can be taken from shape
• [x] For functions that exist in dask_glm, e.g. make_classification, we should default to making a MLDataset as in the xarray_filters.datasets so far, but use dask_glm's funcs for a dask.array in each DataArray rather than sklearn.datasets numpy based approach.
  • [ ] Provide a use_dask_glm=True keyword to control whether the functions in dask_glm.datasets are used.
  • [ ] Change the sequence of acceptable strings for astype to the following (or equivalent way of specifying the data structures below as the output type): ( 'pandas.dataframe','dask.array', 'dask.dataframe', 'numpy.ndarray', ,'dataset', 'mldataset')
  • [x] xnames should be layers
  • [ ] docstring edits - See below: This is current docstring for make_blobs from xarray_filters - I think it needs more of the docs from the transformation part explained, e.g. that it typically outputs N-D DataArrays in an MLDataset or any differences between sklearn and xarray_filters like n_samples versus shape:

In [3]: ?make_blobs
Signature: make_blobs(n_samples=100, n_features=2, centers=3, cluster_std=1.0, center_box=(-10.0, 10.0), shuffle=True, random_state=None, *, astype='dataset', **kwargs)
Docstring:
Like sklearn.datasets.samples_generator.make_blobs, but with added functionality.

Parameters
---------------------
Same parameters/arguments as sklearn.datasets.samples_generator.make_blobs, in addition to the following
keyword-only arguments:

astype: str
    One of ('array', 'dataframe', 'dataset', 'mldataset') or None to return an NpXyTransformer. See documentation
    of NpXyTransformer.astype.

**kwargs: dict
    Optional arguments that depend on astype. See documentation of
    NpXyTransformer.astype.

Note - where I said dask_glm above - also look at dask-ml

[PeterDSteinberg]

Other TODOs I need to add:

• [x] Ensure that the named dims can be controlled, i.e. that dims like x,y,z,t can be named rather than dim_0 dim_1 by default.

[gpfreitas]

• MLDataset default: check
• no need for n_samples when shape passed: check (I chose to let shape overrides n_samples)
• layers instead of xnames: check

I think letting shape be a dict should be enough for letting the user customize dimension names.

So, what's left is the harder part:

• [ ] support dask data structures, see dask-glm
• [ ] change the sequence of acceptable strings for astype (already supported in master)

For astype, @PeterDSteinberg, we should leave the to_* methods intact, right? So, passing astype='numpy.ndarray' would call XyTransformer.to_array. Sounds good?

[gpfreitas]

Working on the dask-glm support.

[PeterDSteinberg]

Note the dask-ml / dask-glm related work is being addressed in a separate issue: https://github.com/ContinuumIO/xarray_filters/issues/36