[ENH] Estimators failing test check_fit_deterministic

[TonyBagnall]

Describe the feature or idea you want to propose

fails of check_fit_deterministic are usually related to seeding, and looking at the list may be numba related

• "SASTClassifier"
• "RSASTClassifier"
• "SAST"
• "RSAST"
• "SFA"

Describe your proposed solution

fix or create issue

Describe alternatives you've considered, if relevant

No response

Additional context

No response

[aryanpola]

@TonyBagnall how do I reproduce these errors?

[lucifer4073]

Hi @TonyBagnall,

I went through the check_fit_deterministic() function and clearly the issue was uneven seeding.

There are two approaches to resolve this:

1. Check for a seed attribute in the clone_estimator function; if it exists, initialize it with random_state, as done in PR [ENH] Resolve failing check_fit_deterministic #2575.
2. The second approach, which I believe might not be ideal, involves checking if seed is not None and assigning it a value during classifier initialization.

Bottleneck in the code:

self._random_state = (
    np.random.RandomState(self.seed)
    if not isinstance(self.seed, np.random.RandomState)
    else self.seed
)

[baraline]

I think the correct way to solve this issue would rather be to change these estimators non "aeon-standard" seed parameter by random_state parameter. The issue is that check_fit_deterministic expects a random_state parameters and not seed :

def check_fit_deterministic(estimator, datatype):
    """Check that calling fit twice is equivalent to calling it once."""
    estimator = _clone_estimator(estimator, random_state=0)

If the estimator don't use a random state parameter that might be the cause of the issue I think ? We would need to put up a deprecation notice if we do that.

[lucifer4073]

I think the correct way to solve this issue would rather be to change these estimators non "aeon-standard" seed parameter by random_state parameter. The issue is that check_fit_deterministic expects a random_state parameters and not seed :

def check_fit_deterministic(estimator, datatype):
"""Check that calling fit twice is equivalent to calling it once."""
estimator = _clone_estimator(estimator, random_state=0)
If the estimator don't use a random state parameter that might be the cause of the issue I think ? We would need to put up a deprecation notice if we do that.

Thanks for the reply.

As far my understanding, I shall change the "seed" parameter to random_state. Regarding the deprecation notice, I’m not entirely sure about the best approach.

[lucifer4073]

Hi @TonyBagnall @baraline ,

I have made the requested changes. Please have a look.