Avoid log(0) in KL divergence

[kevin1kevin1k]

Repository commit

03a4251

Python version (python --version)

Python 3.10.15

Dependencies version (pip freeze)

absl-py==2.1.0
astunparse==1.6.3
beautifulsoup4==4.12.3
certifi==2024.8.30
charset-normalizer==3.4.0
contourpy==1.3.0
cycler==0.12.1
dill==0.3.9
dom_toml==2.0.0
domdf-python-tools==3.9.0
fake-useragent==1.5.1
flatbuffers==24.3.25
fonttools==4.54.1
gast==0.6.0
google-pasta==0.2.0
grpcio==1.67.0
h5py==3.12.1
idna==3.10
imageio==2.36.0
joblib==1.4.2
keras==3.6.0
kiwisolver==1.4.7
libclang==18.1.1
lxml==5.3.0
Markdown==3.7
markdown-it-py==3.0.0
MarkupSafe==3.0.2
matplotlib==3.9.2
mdurl==0.1.2
ml-dtypes==0.3.2
mpmath==1.3.0
namex==0.0.8
natsort==8.4.0
numpy==1.26.4
oauthlib==3.2.2
opencv-python==4.10.0.84
opt_einsum==3.4.0
optree==0.13.0
packaging==24.1
pandas==2.2.3
patsy==0.5.6
pbr==6.1.0
pillow==11.0.0
pip==24.2
protobuf==4.25.5
psutil==6.1.0
Pygments==2.18.0
pyparsing==3.2.0
python-dateutil==2.9.0.post0
pytz==2024.2
qiskit==1.2.4
qiskit-aer==0.15.1
requests==2.32.3
requests-oauthlib==1.3.1
rich==13.9.2
rustworkx==0.15.1
scikit-learn==1.5.2
scipy==1.14.1
setuptools==74.1.2
six==1.16.0
soupsieve==2.6
sphinx-pyproject==0.3.0
statsmodels==0.14.4
stevedore==5.3.0
symengine==0.13.0
sympy==1.13.3
tensorboard==2.16.2
tensorboard-data-server==0.7.2
tensorflow==2.16.2
tensorflow-io-gcs-filesystem==0.37.1
termcolor==2.5.0
threadpoolctl==3.5.0
tomli==2.0.2
tweepy==4.14.0
typing_extensions==4.12.2
tzdata==2024.2
urllib3==2.2.3
Werkzeug==3.0.4
wheel==0.44.0
wrapt==1.16.0
xgboost==2.1.1

Expected behavior

The entries where y_true is 0 should be ignored in the summation (see Actual behavior)

Actual behavior

In

Python/machine_learning/loss_functions.py

Lines 662 to 663 in 03a4251

	kl_loss = y_true * np.log(y_true / y_pred)
	return np.sum(kl_loss)

if any entry of y_true is 0, the output of np.log would become -inf and thus the method returns nan.
Maybe it would be better to exclude those entries where y_true is 0?

[vedprakash226]

if y_true is 0 than what we have to return

[kevin1kevin1k]

I would like to work on this.

I left comments on your PR. Basically we can only use the entries where y_true is nonzero, without adding epsilon.

if y_true is 0 than what we have to return

y_true is an array instead of a number here, so we can still use the remaining entries.

[brambhattabhishek]

1.# Ensure that y_pred doesn't have zero values to avoid division by zero
2.# Clip y_pred to a small positive value to avoid log(0)
3.# Calculate the KL divergence only for non-zero y_true entries

def kl_divergence(y_true, y_pred):
    y_pred = np.clip(y_pred, 1e-10, None) to avoid log(0)
    kl_loss = np.where(y_true != 0, y_true * np.log(y_true / y_pred), 0)
    return np.sum(kl_loss)

[brambhattabhishek]

I would like to work on this.

I left comments on your PR. Basically we can only use the entries where y_true is nonzero, without adding epsilon.

if y_true is 0 than what we have to return

y_true is an array instead of a number here, so we can still use the remaining entries.

is my solution is correct ?

[kevin1kevin1k]

I would like to work on this.

I left comments on your PR. Basically we can only use the entries where y_true is nonzero, without adding epsilon.

if y_true is 0 than what we have to return

y_true is an array instead of a number here, so we can still use the remaining entries.

is my solution is correct ?

I think it's correct, and additionally handles the case where y_pred is 0. Great job.

[brambhattabhishek]

I would like to work on this.

I left comments on your PR. Basically we can only use the entries where y_true is nonzero, without adding epsilon.

if y_true is 0 than what we have to return

y_true is an array instead of a number here, so we can still use the remaining entries.

is my solution is correct ?

I think it's correct, and additionally handles the case where y_pred is 0. Great job.

thanks sure