Skip to content

Add a clipping option to KL divergence inputs for numerical stability. #345

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add a clipping option to KL divergence inputs for numerical stability. #345

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 8 additions & 4 deletions edward2/tensorflow/metrics.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,12 +171,13 @@ def logit_kl_divergence(logits_1, logits_2):
return tf.reduce_mean(vals)


def kl_divergence(p, q):
def kl_divergence(p, q, clip=True):
"""Generalized KL divergence [1] for unnormalized distributions.

Args:
p: tf.Tensor.
q: tf.Tensor
q: tf.Tensor.
clip: bool.

Returns:
tf.Tensor of the Kullback-Leibler divergences per example.
Expand All @@ -187,7 +188,10 @@ def kl_divergence(p, q):
matrix factorization." Advances in neural information processing systems.
2001.
"""
return tf.reduce_sum(p * tf.math.log(p / q) - p + q, axis=-1)
if clip:
p = tf.clip_by_value(p, tf.keras.backend.epsilon(), 1)
q = tf.clip_by_value(q, tf.keras.backend.epsilon(), 1)
return tf.reduce_sum(p * tf.math.log(p / q), axis=-1)


def lp_distance(x, y, p=1):
Expand Down Expand Up @@ -229,7 +233,7 @@ def average_pairwise_diversity(probs, num_models, error=None):
# TODO(ghassen): we could also return max and min pairwise metrics.
average_disagreement = tf.reduce_mean(tf.stack(pairwise_disagreement))
if error is not None:
average_disagreement /= (error + tf.keras.backend.epsilon())
average_disagreement /= (1 - error + tf.keras.backend.epsilon())
average_kl_divergence = tf.reduce_mean(tf.stack(pairwise_kl_divergence))
average_cosine_distance = tf.reduce_mean(tf.stack(pairwise_cosine_distance))

Expand Down