[Track v1.4] New PR branch to serve as submodule for scikit-tree

[adam2392]

Reference Issues/PRs

As of v0.2 for sktree, we have decided we do not need a custom built and released via pypi scikit-learn fork. Instead, we just have to keep an updated fork branch here that maintains the changes under tree/ and ensemble/.

This branch has significantly lower diff and less complexity compared to e.g. #44

What does this implement/fix? Explain your changes.

Any other comments?

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: dda0df6. Link to the linter CI: here}

[PSSF23]

I cannot reproduce the error on my local machine with following code:

from sklearn.ensemble import RandomForestClassifier

import numpy as np
from inspect import signature

rnd = np.random.RandomState(0)
n_samples = 30
X = rnd.uniform(size=(n_samples, 3))
y = np.arange(n_samples)

clf_1 = RandomForestClassifier()
clf_1.set_params(random_state=0)

func = getattr(clf_1, "fit", None)
func(X,y)
args = [p.name for p in signature(func).parameters.values()]

func = getattr(clf_1, "score", None)
func(X,y)
args = [p.name for p in signature(func).parameters.values()]

func = getattr(clf_1, "partial_fit", None)
func(X,y)
args = [p.name for p in signature(func).parameters.values()]

The code should replicate most of the check_fit_score_takes_y test, but it runs smoothly every time. I also don't understand why only these 2 CIs failed when all should have the same test library.

• Linux_Runs pylatest_conda_forge_mkl
• macOS pylatest_conda_mkl_no_openmp

[adam2392]

I cannot reproduce the error on my local machine with following code:

from sklearn.ensemble import RandomForestClassifier

import numpy as np
from inspect import signature

rnd = np.random.RandomState(0)
n_samples = 30
X = rnd.uniform(size=(n_samples, 3))
y = np.arange(n_samples)

clf_1 = RandomForestClassifier()
clf_1.set_params(random_state=0)

func = getattr(clf_1, "fit", None)
func(X,y)
args = [p.name for p in signature(func).parameters.values()]

func = getattr(clf_1, "score", None)
func(X,y)
args = [p.name for p in signature(func).parameters.values()]

func = getattr(clf_1, "partial_fit", None)
func(X,y)
args = [p.name for p in signature(func).parameters.values()]

The code should replicate most of the check_fit_score_takes_y test, but it runs smoothly every time. I also don't understand why only these 2 CIs failed when all should have the same test library.

• Linux_Runs pylatest_conda_forge_mkl
• macOS pylatest_conda_mkl_no_openmp

If you try running the HonestForestClassifier instead, then the error seems to be able to be produced on my computer

[PSSF23]

@adam2392 Is sktree updated with the current submodule? After the resizing bug is fixed?

[adam2392]

Yeah it should be. The test is commented out tho rn.

[PSSF23]

All CIs passed.

[adam2392]

TODO: Change

int -> intp_t
double -> float64_t
SIZE_t -> intp_t
DTYPE_t -> float32_t
INT32_t -> int32_t

see: https://github.com/scikit-learn/scikit-learn/pull/27352/files and related PRs

[adam2392]

TODO: Change

int -> intp_t double -> float64_t SIZE_t -> intp_t DTYPE_t -> float32_t INT32_t -> int32_t

see: https://github.com/scikit-learn/scikit-learn/pull/27352/files and related PRs

This was accomplished in 9a5d91b

[adam2392]

5ccd00f introduces a major change, where n_constant_features is migrated to WITHIN the SplitRecord. This is to enable over-riding what gets passed from parent node to child node.

I.e. MultiviewSplitRecord could in principle override SplitRecord and store additional n_constant_features. One per feature set. Ofc, I don't think this is actually the best design cuz the idea of "multi-view" really breaks away from the underlying assumption that there is only one feature set.

[adam2392]

We should remove binning for now as it is a untested feature...