test_estimator not valid for vectorizers

[ysig]

We are currently developing a transformer compatible with sk-learn that behaves
as a vectorizer for graph type object. The test estimator method injects arrays of
n by m which are not valid to our current input.Tfidf vectorizer supports also a kind
o input that is not an array of n by m features, but rather a vector of strings.
Can the check_estimator constraints be softened for a vectorizer type transformer?

project-template/skltemplate/tests/test_common.py

Line 7 in ac1f099

return check_estimator(TemplateEstimator)

[agramfort]

send us a PR

[ysig]

Ok, the personal repo is the link that follows. The library is not yet mature, but the essential part is there and concerns the input form of the graph_kernel file. It is just an iterable (a vector) of objects (in our case a list of lists) where inside each sub-list we expect as elements a graph type input and optionally graph labels. Init function is used to set the kernel function applied in "transform".
https://github.com/ysig/GraKeL
https://github.com/ysig/GraKeL/blob/master/grakel/graph_kernels.py

[glemaitre]

Actually this should be reported to scikit-learn since that we want the check_estimator to be flexible enough.