Pipeline - Parameterize transforms of MLDataset or xarray.Dataset

[PeterDSteinberg]

This issue is about creating Pipeline to control parameterization of transformations of MLDataset or Dataset objects. When this issue is done, we should be able to do the following:

from xarray_filters.pipeline import Pipeline
from xarary_filters.steps import Generic, Serialize
def step_1(dset, **kw):
    return kw['a'] * dset.mean(dim=('x', 'y')) ** kw['b']

def step_2(dset, **kw):
    return kw['a'] + dset * kw['b']
    
steps = (('s1', Generic(step_1)),
         ('s2', Generic(step_2)),
         ('s3', Serialize('two_step_pipeline_out.nc')))
pipe = Pipeline(steps=steps)
pipe.set_params(s1__a=2,
                s1__b=3,
                s2__a=0,
                s2__b=0,
                s3__fname='file_with_zeros.nc')
pipe.fit_transform(X)

See this example notebook and also look at the documentation on custom estimators in sklearn.

[gbrener]

Sample implementation:

https://github.com/ContinuumIO/xarray_filters/blob/param-mlds/notebooks/Parameterize-MLDataset.ipynb

Notes:

"Pipeline compatibility" doc

Illustrates how to write custom Estimators/Transformers to pass into a sklearn Pipeline: http://scikit-learn.org/stable/developers/contributing.html#pipeline-compatibility

Reusing sklearn.pipeline.Pipeline

In order to reuse as much of the sklearn.pipeline.Pipeline class as possible, the sample implementation (above) sets the final estimator to None; this has the effect of the "identity" function. Here is the relevant sklearn code that treats None as identity: https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/pipeline.py#L165

• @PeterDSteinberg commented:

  ...in xarray_filters, we will always be returning a MLDataset/Dataset (running the last step as well which will be a func taking a MLDataset and returning one.

Pipelining example from sklearn

Real-world example of Pipeline being used with GridSearchCV: http://scikit-learn.org/stable/auto_examples/plot_digits_pipe.html#sphx-glr-auto-examples-plot-digits-pipe-py

• We could interchange the xarray_filters Pipeline here instead of the sklearn Pipeline, and pass it into dask-searchcv (instead of GridSearchCV), for example

sklearn.pipeline.FeatureUnion

Similar to Pipeline, but parallelizes the computations over each feature, allowing for optional weighting: http://scikit-learn.org/stable/modules/generated/sklearn.pipeline.FeatureUnion.html
Example of where this is useful: unifying heterogeneous data sources (http://scikit-learn.org/stable/auto_examples/hetero_feature_union.html#sphx-glr-auto-examples-hetero-feature-union-py)

[PeterDSteinberg]

A couple points:

• Note also @gbrener there is some dask-related work with FeatureUnion here in dask_searchcv - we may want to take advantage of those ideas.
• I'm working on a separate experiment with subclassing the parent (base) class of GridSearchCV and hope to have something to show later this week.
• Your point under reusing Pipeline from sklearn and its None final identity transformer sounds good.
• We should have tests in place that we support the clone operation described in the pipeline compatibility doc here

[gbrener]

Thanks for the notes @PeterDSteinberg . I haven't looked much into FeatureUnion yet but I intend to soon. I added a unit test to confirm that clone works. Also, the build is now passing for Python 3.5 and Python 3.6, including the deploy section. Python 2.7 is still failing due to #12 - I believe @gpfreitas may be working on this.

I'm now adding builtin transformers from elm.sample_util (now located at earthio.filters) - hopefully this shouldn't be too involved now that the hard part is done. After that I'll see how we might incorporate dask into the subclasses which inherit from Step; at the moment my plan is to use dask.delayed.

[PeterDSteinberg]

PR looks good - adding Step based classes at this point sounds good - thanks @gbrener !

[PeterDSteinberg]

@gbrener I moved the FeatureUnion to a separate issue #22. I'm closing this one. If there's anything missing else on this issue not handled yet, please make other separate issues.