Is ChromBPNet trained with additional data?

[gonzalobenegas]

Thanks for your work, excited to see the paper out!

From reading the paper I understand that for Tasks 4 and 5, only the predicted total reads are used from ChromBPNet. However, you describe that ChromBPNet was also trained on profile prediction.

The count prediction was trained using mean squared error loss, while the profile head was trained using log-likelihood loss based on a multinomial distribution.

We utilized the same training setup for our probing and fine-tuning models so that inputs and labels were identical to
those for ChromBPNet.

Would it be fair to say that ChromBPNet is given additional, profile-level labels?