For cascade event-generator models it's beneficial to utilize symmetries about the shower maximum rather than the cascade vertex as these are more appropriate. Currently, this is done by modifying the labels during the training procedure, if a model activates the parameter shift_cascade_vertex in the label_settings of the data_handler_settings. In this case, the labels of the cascade vertex are shifted and training is performed for these shifted labels. The result is that the cascade model expects the cascade parameters to be the shifted ones. This may create confusion later on for end users. As of now, the user has to manually check the setting of the label module and shift the injected cascade or reconstructed cascade appropriately.
A better alternative would be to make the shift_cascade_vertex a property of the model architecture itself. The model can then perform the shift internally and still use the appropriate symmetries, while the definition of the cascade labels remains clear. This would reduce the risk of mis-interpretation
mhuen
commented
3 months ago
For cascade event-generator models it's beneficial to utilize symmetries about the shower maximum rather than the cascade vertex as these are more appropriate. Currently, this is done by modifying the labels during the training procedure, if a model activates the parameter
shift_cascade_vertexin thelabel_settingsof thedata_handler_settings. In this case, the labels of the cascade vertex are shifted and training is performed for these shifted labels. The result is that the cascade model expects the cascade parameters to be the shifted ones. This may create confusion later on for end users. As of now, the user has to manually check the setting of the label module and shift the injected cascade or reconstructed cascade appropriately.A better alternative would be to make the
shift_cascade_vertexa property of the model architecture itself. The model can then perform the shift internally and still use the appropriate symmetries, while the definition of the cascade labels remains clear. This would reduce the risk of mis-interpretation