Warn on changing vocab for trained models

[juhoinkinen]

When several projects share a vocabulary, it is easy to unintentionally change the vocab for all them, when one means to change the vocab for only one project. Especially when some of the projects are already trained, this can be a problem, because changing the vocab then messes the models (is this always the case?). For example:

$ annif loadvoc tfidf-fi ~/annif-projects/Annif-corpora/vocab/yso-fi.tsv
$ annif train tfidf-fi ~/annif-projects/Annif-corpora/training/2019/yso-cicero-finna-fi-100000-lines.tsv
$ echo kissa | annif suggest tfidf-fi
<http:https://www.yso.fi/onto/yso/p19378>	kissa	0.8595056533813477
<http:https://www.yso.fi/onto/yso/p17959>	kasvianatomia	0.32491984963417053
<http:https://www.yso.fi/onto/yso/p20613>	eläinanatomia	0.31712543964385986
<http:https://www.yso.fi/onto/yso/p18313>	eläinfysiologia	0.2782534062862396
<http:https://www.yso.fi/onto/yso/p20292>	biomekaniikka	0.25385648012161255
<http:https://www.yso.fi/onto/yso/p18481>	eläinten käyttäytyminen	0.2513505518436432
<http:https://www.yso.fi/onto/yso/p10562>	kasvifysiologia	0.2394426316022873
<http:https://www.yso.fi/onto/yso/p11669>	nimipäivät	0.18711934983730316
<http:https://www.yso.fi/onto/yso/p675>	lemmikkieläimet	0.17345558106899261
<http:https://www.yso.fi/onto/yso/p22993>	naksutinkoulutus	0.15796299278736115

# Now load (a different) vocab for fasttext (which has the same vocab setting in projects.cfg as tfidf):
$ annif loadvoc fasttext-fi tests/corpora/archaeology/subjects.tsv
$ echo kissa | annif suggest tfidf-fi
$ 
(No results)

Annif could give a warning when reloading a vocabulary, which could list all the projects that share the vocabulary, or at least all the projects that have been already trained using the vocabulary that now changes. Or there could even be a confirmation prompt for the latter projects case.

[osma]

This is a good idea. However, if we implemented #274 first then the problem would be at least partly mitigated.

[juhoinkinen]

As Osma pointed today, a similar problem probably arises also when projects with different vocabularies are combined to an ensemble model.

[juhoinkinen]

I think this can be closed, because since #614 the argument to the load-vocabulary command is a vocabulary ID, not a project ID, so it is a less surprise that the operation affects (or can affect) multiple projects. Also since #274 it could be possible to "undo" loading a wrong vocabulary, because the original URIs are retained in the internal vocabulary, so just loading the original vocabulary back could reset the situation. (Disclaimer: I'm not sure about this.)