Differences in max and sum values when using pre-trained model for testing？

[Fengxin0401]

Hello,I'm encountering an issue while using the Borzoi pre-trained model for testing. The peak trend of the coverage I predicted (e.g., in Figure 2) aligns with the trend shown in your article, but there are significant differences in the numerical values of max and sum compared to those provided in the article.
Questions:
Could these numerical differences be caused by input data formatting/processing (e.g., missing normalization steps, incorrect feature scaling)?
Are there official example scripts or reference code that I can review for consistency?

[johli]

Hi,

I am attaching a jupyter notebook that generates the coverage prediction visualization from Fig 2a of the paper (the ADGRE1 gene example). The decimal numbers change slightly depending on e.g. tensorflow/cuda version and the exact GPU used.

borzoi_example_ADGRE1.ipynb

Best,

Johannes