mirror of
https://github.com/luxfhe/ml-sdk.git
synced 2026-10-30 18:13:04 +00:00
zeekay
bde2e2531b
Machine Learning with FHE - fork of Zama Concrete-ML - Built-in FHE-friendly ML models - scikit-learn compatible API - PyTorch model conversion - Quantization-aware training
66 lines
2.1 KiB
Python
66 lines
2.1 KiB
Python
"""Tests for the r2 score test """
|
|
|
|
import numpy
|
|
import pytest
|
|
|
|
BIG_OFFSET = 10000
|
|
BIG_STDEV = 100
|
|
SMALL_VALUE = 0.00001
|
|
SMALL_STDEV = SMALL_VALUE
|
|
SMALL_OFFSET = SMALL_VALUE
|
|
|
|
|
|
@pytest.mark.parametrize("gt_range", [SMALL_STDEV, BIG_STDEV])
|
|
@pytest.mark.parametrize("gt_stdev", [SMALL_STDEV, BIG_STDEV])
|
|
@pytest.mark.parametrize(
|
|
"predicted_params",
|
|
zip(
|
|
[None, SMALL_STDEV, SMALL_STDEV, BIG_STDEV, BIG_STDEV],
|
|
[0, SMALL_OFFSET, BIG_OFFSET, SMALL_OFFSET, BIG_OFFSET],
|
|
),
|
|
)
|
|
@pytest.mark.parametrize("num_values", [1000])
|
|
def test_r2(
|
|
gt_range,
|
|
gt_stdev,
|
|
predicted_params,
|
|
num_values,
|
|
check_r2_score,
|
|
):
|
|
"""Test our modified r2 test on various data distributions"""
|
|
|
|
# Generate a uniform distribution and add a normal residual on top
|
|
gt_values = numpy.random.uniform(size=(num_values,)) * gt_range
|
|
gt_normal_residuals = numpy.random.normal(0, gt_stdev, size=(num_values,))
|
|
gt_values += gt_normal_residuals
|
|
|
|
predicted_stdev, predicted_offset = predicted_params
|
|
|
|
if predicted_stdev is not None:
|
|
# Correlated predicted value
|
|
pred_normal_residuals = numpy.random.normal(0, predicted_stdev, size=(num_values,))
|
|
predicted = gt_values + pred_normal_residuals
|
|
else:
|
|
# No correlation between gt and predicted
|
|
predicted = numpy.random.uniform(size=(num_values,))
|
|
|
|
predicted += predicted_offset
|
|
|
|
if predicted_stdev is None:
|
|
# If there is no correlation between gt and predicted, the test should fail
|
|
should_fail = True
|
|
elif predicted_stdev <= SMALL_VALUE and predicted_offset <= SMALL_VALUE:
|
|
# If the predicted values are very close to the target values (very low perturbation)
|
|
should_fail = False
|
|
else:
|
|
# If there is a difference in stdev or in the offset, the test should fail
|
|
# Note that the stdev should not be as high as the offset as by chance it could break the
|
|
# test
|
|
should_fail = True
|
|
|
|
if should_fail:
|
|
with pytest.raises(AssertionError):
|
|
check_r2_score(gt_values, predicted)
|
|
else:
|
|
check_r2_score(gt_values, predicted)
|