Tree: node_value could be float instead of double

[sdpython]

Description

Predict function in _tree.pyx is comparing a float feature to a double threshold (node_value). The double threshold could be replaced by a float threshold using the following function:

import numpy

def float_threshold(dy):
    fy = numpy.float32(dy)
    if fy == dy:
        return fy
    if fy < dy:
        return fy
    eps = max(abs(fy), numpy.finfo(numpy.float32).eps) * 10
    nfy = numpy.nextafter([fy], [fy - eps], dtype=numpy.float32)[0]
    return nfy

More explanations: Tricky detail when converting a random forest from scikit-learn into ONNX.

Expected Results

Exactly the same results but with less memory consumption and maybe with some speed gain as no cast to double would be required when comparing both values.

Versions

Any version >= 020.

[cmarmo]

Hi @sdpython, thanks for reaching out, and sorry for the late answer. Are you interested in proposing a PR?

[NicolasHug]

@sdpython instead of node_value, do you mean node.threshold? I do not understand yet how the node's value (i.e. the prediction) relates to this threshold comparison.