Use errors='coerce' when converting to numerical or datetime in TableVectorizer's _apply_cast

[LeoGrin]

Fix #631

Questions:

• do we want to warn the user if there is an issue with one of the inferred types?
• what we would like to do with the new entry of the wrong type would be to ignore it. If we convert it to nan, some model may want to use what it learnt about missing values during training on this entry, though it wouldn't actually be missing.

[GaelVaroquaux]

Answers to your question:

• I think that we should warn, possibly with a dedicated error to make it easy to catch this error
• We should attempt the "best conversion possible", whatever that means. This may not be easy.

I'll review tomorrow, it's getting too late

[GaelVaroquaux]

One comment: I think that we should code a mechanism to raise a warning. We can discuss this to gauge how important it is.

[GaelVaroquaux]

With the current usage of "coerce" it is hard to do a warning if an invalid value is encountered.

If we feel that it is important to warn (I am leaning in this direction but not 100% sure), one way to do it is to do a try/except with 'errors="raise"', if an error (the "except" block), capture it, pop up a warning (ideally informative) and then do the same call but with 'error="coerce"'.

[GaelVaroquaux]

❤️

[GaelVaroquaux]

Thanks!! Merged