MAINT validate parameters in TSNE - #23845

@Micky774 Thank you for your comments and suggestions :)

Currently failing the test_estimators[TSNE()-check_fit2d_1sample] test, which makes sense because: perplexity should be strictly less than the number of samples (one), but also no less than one. In this case there's no valid value of perplexity to satisfy a one-sample fit. Should we consider excusing TSNE from this check? @jeremiedbb @glemaitre

Alternatively, we could allow perplexity < 1 to not break backward compat and reconsider this issue when changing the lower bound in a separate PR

Alternatively, we could allow perplexity < 1 to not break backward compat and reconsider this issue when changing the lower bound in a separate PR

Sounds good, no need to hold up this PR

I changed the interval of the perplexity to allow perplexity < 1.

@Micky774
When running utils.estimator_checks I noticed that the current version throws an error during the test "check_estimators_dtypes". As far as I understood it is related to this PR . The test uses X_train of size 20, which is lower than the default perplexity (=30), hence the function "_check_params_vs_input" raises an error. Is it the expected behavior of the test?

When running utils.estimator_checks I noticed that the current version throws an error during the test "check_estimators_dtypes". As far as I understood it is related to this PR . The test uses X_train of size 20, which is lower than the default perplexity (=30), hence the function "_check_params_vs_input" raises an error. Is it the expected behavior of the test?

I'm not sure what you're referring to here -- the tests pass on both main and this PR. Under what circumstances are you getting this error?

I'm not sure what you're referring to here -- the tests pass on both main and this PR. Under what circumstances are you getting this error?

I think it is a misunderstanding from my side.
I was checking this doc (under "Rolling your own estimator") and tried the following code:

from sklearn.utils.estimator_checks import check_estimator
from sklearn.manifold import TSNE
check_estimator(TSNE())

which throws the error below:
ValueError: perplexity must be less than n_samples

But when running the tests with pytest:
pytest -vl scikit-learn/sklearn/tests/test_common.py
the tests pass.

Actually that's a good catch! We should fix that typo in its own PR (would you be interested 👀).

Regarding why our common tests pass: before running that check we set the parameters of the estimators to values that won't raise errors for it, e.g. in this case we actually explicity set the perplexity to something lower for the reasons you observed!

Edit: for clarification, by "fix that typo" I mean replace TSNE with another estimator that doesn't need custom-set parameters for the check.

Edit: for clarification, by "fix that typo" I mean replace TSNE with another estimator that doesn't need custom-set parameters for the check

I am not sure to understand this. In the doc "LinearSVC" estimator is used, which does not throw any error. I just tried the same piece of code with TSNE estimator and noticed the error. Where should I replace TSNE with another estimator?

I also noticed that in earlier versions of sklearn (before the implementation of the _check_params_vs_input function), the code check_estimator(TSNE()) did not throw any error.

I am not sure to understand this. In the doc "LinearSVC" estimator is used, which does not throw any error. I just tried the same piece of code with TSNE estimator and noticed the error. Where should I replace TSNE with another estimator?

Ah ignore me I was hasty and falsely thought that TSNE was used in the docs. All is good.

Also, the current failing CI is unrelated to this PR

Got it!

Hey there @chalulu, could you merge w/ main and push to prompt the CI/CD once more? I think the current error is not related to your PR.

@Micky774 I merged with main and pushed again.
Thank you for your review/feedback, I'm looking forward to trying to contribute on more issues.