search

deepjavalibrary / djl

An Engine-Agnostic Deep Learning Framework in Java
https://djl.ai
Apache License 2.0
4.12k stars 654 forks source link

SigmoidBinaryCrossEntropyLoss wrong calculation #1316

Closed huhuh1234567 closed 2 years ago

huhuh1234567 commented 3 years ago

Description

In class djl/api/src/main/java/ai/djl/training/loss/SigmoidBinaryCrossEntropyLoss, when set fromSigmoid to true, the calculation would be wrong. It is at line 77 that a log operation is missing after NDArrays.sub(1., pred).add(eps)

Expected Behavior

The loss should be positive.

Error Message

The loss may be negative.

How to Reproduce?

Steps to reproduce

What have you tried to solve it?

Environment Info

zachgk commented 3 years ago

Thanks for pointing this out @huhuh1234567! Would you be interested in submitting a PR to fix this?