[ENH] Add the option to turn off all type checking and conversion and go straight to _fit/_predict/_transform

[TonyBagnall]

Describe the feature or idea you want to propose

in estimator base classes (looking at BaseCollectionTransformer but same elsewhere) we do checks, get meta data and convert to inner type. All good, but it does introduce an overhead. For example, this is minirocket with and without the checks

the checks involve several scans of the data. For larger data it would be good to have an option to just go straight to inner methods and fail without informative error.

This is minirocket in safe and unsafe modes (time in secs, to transform length 500 series varying number of cases). Note the difference is linear in n

<style> </style>

	With	Unsafe	Diff
500	0.49	0.16	-0.33
1000	0.88	0.20	-0.68
1500	1.30	0.26	-1.04
2000	1.74	0.32	-1.43
2500	2.02	0.49	-1.53
3000	2.46	0.44	-2.02
3500	2.81	0.49	-2.32
4000	3.19	0.54	-2.65
4500	3.66	0.60	-3.06
5000	4.08	0.73	-3.35
5500	4.57	0.75	-3.82
6000	4.75	0.83	-3.92
6500	5.17	0.83	-4.33
7000	5.80	0.95	-4.85
7500	6.14	0.99	-5.15
8000	6.51	1.01	-5.50
8500	6.85	1.05	-5.80
9000	7.33	1.13	-6.21
9500	7.55	1.23	-6.32
10000	7.92	1.22	-6.70

Describe your proposed solution

I would have a parameter in constructor, say "unsafe", that defaults to False, then just test in fit etc.

    def fit(self, X, y=None):
          if self.unsafe:
                 self._fit(X,y)
                 return self
        if self.get_tag("requires_y"):
            if y is None:
                raise ValueError("Tag requires_y is true, but fit called with y=None")
        # skip the rest if fit_is_empty is True
        if self.get_tag("fit_is_empty"):
            self._is_fitted = True
            return self
        self.reset()

etc

Describe alternatives you've considered, if relevant

No response

Additional context

No response

[TonyBagnall]

ok can ignore this, it was in fact an internal bug in mini rocket :)