Standardize errors for when pre-trained weights are not available

[datumbox]

Our current implementation throws a ValueError if pre-trained weights are requested for an architecture that we don't offer checkpoints. In this PR we switch to NotImplementedError exceptions and we standardize the error messages.

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit 394aa03 (more details on the Dr. CI page):

---

💚 💚 Looks good so far! There are no failures yet. 💚 💚

---

This comment was automatically generated by Dr. CI (expand for details).

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

[NicolasHug]

Thanks for the PR @datumbox . What is the logic for using a NotImplementedError here? At first sight it seems that ValueError may be more relevant. From the docs of the NotImplementedError exception:

It should not be used to indicate that an operator or method is not meant to be supported at all

[jdsgomes]

I left a bit but overall I'm aligned with Nicolas point that NotImplemented might not be the best candidate here

[datumbox]

@NicolasHug I think you changed your opinion on this because this change was first proposed by you here: #5022 (comment)

Personally I think throwing ValueError is not the right option for this. The models can throw multiple ValueError exceptions if not parameterized correctly. Throwing NotImplementedError for something that is not supported is in the spirit of the docs.

[NicolasHug]

@NicolasHug I think you changed your opinion on this because this change was first proposed by you here: #5022 (comment)

For the models modified in this PR, are we absolutely sure that we will implement the pretrained weights in the future (as it was the case in #5022 (comment)) ? If not, the Python docs explicitly recommend against using NotImplementedError.

[datumbox]

@NicolasHug Yes, there is nothing stopping us from providing these models on the future and nor I can think of a case were we offer a model builder but we absolutely don't want to offer weights. Also note that the models missing are on the list to be added and can be contributed by the community #2707

[NicolasHug]

I believe we rely on ValueError much more often than on NotImplementedError in the current code-base for such scenarios. I won't fight against it but sorry, I prefer not to approve it.