Skip to content

Improvements for dataloading, inference, regression, and export

Samuel Van Stroud requested to merge svanstro/salt:svanstro/update-dataloading into main

Biger things

  • Following on from !118 (merged), this MR moves the mean/std regression target normalisation to the regression task class, instead of being implemeted in both the dataloading and prediction writing stages. Now the two regression target scaling options (denominator and mean/std scaling) are both explicitly configured and this happens in the same place.
  • Add an option for the input scaling to happen inside each InitNet via data.norm_in_model: True. Tying the normalisation to an init net is more coherent and completes the factorisation between input and label scaling. The to_onnx script is simplified accordingly (as we no longer need to manually add the norm operations).
  • Improved inheritance for the Task classes, added run_inference methods as discussed in #30 (closed) to make the PredictionWriter class cleaner
  • Improve the support for jet-level tasks in the PredictionWriter - now multiple jet level tasks can be written out together.

smaller things

  • Add a NormDictScaler class which handles the input scaling. \
  • Disable the uncomitted changes check when running from pytest
  • rename data.inputs -> data.input_names

closes #31 (closed). see also #30 (closed)

Edited by Samuel Van Stroud

Merge request reports