There are two transforms with the Friendly Name "Term Transform"

[sfilipi]

In the list of entry points there are two identical transforms, one with the name field: "Transforms.TextToKeyConverter" and the other with the name field: "Transforms.Dictionarizer".
Their Friendly Name filed is the same: "Term Transform".

This will be confusing for systems interfacing with ml.net through the entry points; ml.net should not present the same entry point choice more than once.

[TomFinley]

What is the proposed final name? Are we happy with either "Dictionarizer" or "TextToKey"? The first seems a bit funky (at least to my eye), but the second is not descriptive since obviously this can be applied to more than just text. So I prefer the first but only because being basically meaningless is preferable to being flat-out misleading and wrong.

[justinormont]

Yeah, I was about to say, "TextToKeyConverter" is a bit wrong as it can take in about anything and convert to Key, but @TomFinley beat me to it.

I don't like either one, really. Do we want to expose the concept of keys? If so, "ToKey" is rather terse and hopefully descriptive. I'd say "AnythingToKey", but I don't know that every type, for instance the Picture type, can be converted to a Key type (and the future will bring more types).

To capture the primary use case, "LabelConverter" may be suitable. Scikit calls this a "LabelEncoder": http://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.LabelEncoder.html

I maybe wrong but Tensorflow seems to not have a specialized transform for this, and uses the categorical featurizers:

categorical_column_with_hash_bucket(...): Represents sparse feature where ids are set by hashing.
categorical_column_with_identity(...): A _CategoricalColumn that returns identity values.
categorical_column_with_vocabulary_file(...): A _CategoricalColumn with a vocabulary file.
categorical_column_with_vocabulary_list(...): A _CategoricalColumn with in-memory vocabulary

[TomFinley]

mmm. ToKey... I actually like that name. ToKey. Super simple, fairly descriptive in five characters. Of course, hash is also a "to key." But it's still better than the original name "term."

We sort of have to expose them. They're basically the same things as factors in R, and as far as I can tell you simply can't get around the fact that enumerations into sets is a fairly central concept of ML. Of course whether we actually wind up using "key" in the name, I don't know.

LabelEncoder might be fine, indeed term's first name was "auto-label", but I worry somewhat about what will happen when someone uses the text metatransform, inspects the pipeline, and one of the first things there is them "label-encoding" their feature inputs. :)

[jwood803]

@TomFinley, if it's ok I'd like to work on this. Was ToKey decided to be the name to change one of the Term Transform's to?