can't get proper prediction score from python when --boosting N option used at training time

[jprobichaud]

Describe the bug

I've trained a small model using the --boosting N option, like for example --loss_function logistic -b 18 --l1 0.1 --l2 0.0001 --nn 50 --boosting 5 using python or the command line.

Then, when I try to get prediction scores out of it, I get "1.0" all the time when using python or using -p on the command line. When I use the -r command line parameter with vw cli executable, I get the list of all features with their scores and a final meaninfgul score at the end.

It seems impossible to get these values from within the python classes, no matter what I try (playing with the various PredictionType )

This seems either a documentation bug or a more serious issue w.r.t. boosting reduction perhaps.

How to reproduce

1. train with --loss_function logistic -b 18 --l1 0.1 --l2 0.0001 --nn 50 --boosting 5
2. try to get predictions within python that matches the scores you get from using vw -i model.vw -t test.txt -r /dev/stdout

Version

9.2.0

OS

Linux

Language

Python

Additional context

No response

[ataymano]

Seems like a bug with boosting (and nn/loss function are not relevant here).
Simpler repro (can be done with binary and -p as well, so problem is not python specific):

from vowpalwabbit import pyvw
import numpy as np

vw = pyvw.Workspace('--boosting 1 -b 1')

yhat = lambda x: 2 * x + 3

for i in range(10000):
    x = np.random.rand()
    vw.learn(f'{yhat(x)} | 1:{x}')

print(f'w_x = {vw.get_weight(1, 0)}')
print(f'constant = {vw.get_weight(0, 0)}')

x = 2
print(f'y({x}) = {vw.predict("| 1:{x}")}')

Output:

w_x = 1.9999972581863403
constant = 3.000001907348633
y(2) = 1.0

[ataymano]

yeah seems like some binary classification logic is hardcoded there:

vowpal_wabbit/vowpalwabbit/core/src/reductions/boosting.cc

Line 170 in 63090d6

ec.pred.scalar = VW::math::sign(final_prediction);