Pipeline Objects

The pipeline classes Pipeline and ModelPipeline are made up of base pipeline objects / pieces. These are based on the scikit learn concept of a Pipeline.

Across all base ModelPipeline pieces, e.g., Model or Scaler, there exists an obj param when initializing these objects. Beyond choice of obj parameter, pipeline objects share a number of other common parameters which allow for extra customization, as well as some objects which have unique parameters. Shared parameters include params, scope, extra_params> and cache_loc.

obj

The ‘obj’ parameter is the core parameter for any pipeline object. It can broadly refer to either a str, which indicates a valid pre-defined custom obj for that piece, or depending on the pieces, this parameter can be passed a custom object directly.

For example if we want to make an instance of RobustScaler from sklearn:

In [1]: import BPt as bp

In [2]: scaler = bp.Scaler('robust', scope='all')

In [3]: scaler
Out[3]: Scaler(obj='robust', scope='all')

# See what this object looks like internally
In [4]: scaler.build()
Out[4]: 
(('robust',
  ScopeTransformer(estimator=RobustScaler(quantile_range=(5, 95)), inds='all')),
 {})

We can do this because ‘robust’ exists as a default option already available in BPt (See Scalers). That said, if it wasn’t we could pass it as a custom object as well.

In [5]: from sklearn.preprocessing import RobustScaler

In [6]: scaler = bp.Scaler(RobustScaler())

In [7]: scaler
Out[7]: Scaler(obj=RobustScaler())

params

params are used to either specify or select from a default existing choice of associated fixed or distribution of hyper-parameter values for this object. For example, we can choose to associate an existing hyper-parameter distribution for the robust scaler from before with:

In [8]: scaler = bp.Scaler('robust', params="robust gs")

In [9]: scaler
Out[9]: Scaler(obj='robust', params='robust gs')

We could also set it to a custom distribution using ref:Parameter<api.dists>:

In [10]: quantile_range = bp.p.Choice([(5, 95), (10, 90), (15, 85)])

In [11]: scaler = bp.Scaler('robust', params={'quantile_range': quantile_range})

In [12]: scaler
Out[12]: 
Scaler(obj='robust',
       params={'quantile_range': Choice([(5, 95), (10, 90), (15, 85)])})

See Params for more information on how to set Parameters.

scope

Pipeline objects also have the argument Scope as an input argument. This argument allows for pipeline objects to work on just a subset of columns. Notably, if the scope in reference to a specific Dataset is empty, then the piopeline object in question will just be silently skipped. This is useful for defining generic pipelines to different types of datasets.