Performance of CelebA constrained model

[bytosaur]

• Have I written custom code (as opposed to using stock example code provided): No

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Google Colab

• Fairness Indicators version: 0.30.1

• TensorFlow version: 2.5.0

• Python version: 3.7.10

• TFMA version: 0.30.0

Hey there,

I have noticed that the constrained model in the CelebA example Notebook has a horrible positive rate: 0.1 @ 0.5 threshold.

I expected that the model would perform at least equally good on this metric.