Currently in the config's pipeline section (a list of dictionaries indicating training or prediction actions), the actions must have the keys:
train, or
predict
We should add a 3rd one for transform to cover the cases of models that do not have a predict method but rather a transform and/or fit_transform. Included in this work would be ensuring the transform actions:
can be specified in config, perhaps in the sample_pipeline section of a pipeline step
can serialize / deserialize the outputs of transform for transforms like PCA
Currently in the config's
pipeline
section (a list of dictionaries indicating training or prediction actions), the actions must have the keys:train
, orpredict
We should add a 3rd one for
transform
to cover the cases of models that do not have apredict
method but rather atransform
and/orfit_transform
. Included in this work would be ensuring thetransform
actions:sample_pipeline
section of apipeline
steptransform
for transforms like PCA