Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

EfficietNet + Generalized Mean Pooling causes NaN loss!

See original GitHub issue

Config:

TF: 2.6
Environment: Colab

Issue

I’ve plugged a GeM pooling layer into the EfficientNet model but for an unknown reason, it causes nan loss. But with other image-net models, it works just fine.

Reproducible Code

Issue Analytics

State:
Created 2 years ago
Comments:8 (1 by maintainers)

Top GitHub Comments

1reaction

mattdangerwcommented, Nov 11, 2021

Given that this bug is not hit when trainable is false, we believe this might be an issue with zeros in self.p, which look like they would trigger a nan. You may need to try protecting against zero values in p in the custom GeM layer.

0reactions

innatcommented, Feb 21, 2022

@Jobayer-000 I think for GeM, this p needs to be trainable. Anyway, It’s solved now and I closed the issue.

ref. https://github.com/innat/DOLG-TensorFlow/blob/main/layers/GeM.py