[BUG]: Memory Leak on inference

[mbegue-cr]

Before You Report a Bug, Please Confirm You Have Done The Following...

• I have updated to the latest version of the packages.
• I have searched for both existing issues and closed issues and found none that matched my issue.

RetinaFace's Version

0.0.17

Python Version

3.11.11

Operating System

No response

Dependencies

tensorflow==2.16.2

Reproducible example

from retinaface import RetinaFace
# 423MB of Real Memory used after this line

url = "https://media.cnn.com/api/v1/images/stellar/prod/210610161942-05-g7-through-history.jpg?q=w_1160,c_fill/f_webp"

faces = RetinaFace.detect_faces(url)
# 1.45 GB Real Memory used after this line
faces = RetinaFace.detect_faces(url)
faces = RetinaFace.detect_faces(url)
faces = RetinaFace.detect_faces(url)
# 1.59GB Real Memory use after this call.
# Real Memory size for this python process increases after each call

Relevant Log Output

No response

Expected Result

I expect that any intermediate variables and tensors be cleared from memory after each detect_faces call. Furthermore, I inspected the retinaface model and we clearly have trainable=False, so I don't think we're accumulating gradients.
This is not desirable at all especially if I want to run hundreds of face detection calls in a single python session.

What happened instead?

After importing retinaface there is 423MB of Real Memory used for that python process.
After the first detect_faces call, 1.45GB Real memory is used for that process.
After several detect_faces calls, the memory steadily increases monotonically.

Additional Info

No response