You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
PostprocessGermlineCNVCalls performs a check of the denoising/calling hyperparameter configs used to generate the model in GermlineCNVCaller cohort mode against those used to generate the case-mode result passed to PostprocessGermlineCNVCalls. However, although some of these hyperparameters are not exposed in case mode (since they have no effect on the sample-level parameters inferred in case mode, e.g., psi_t_scale), their python default values are nevertheless written to the case-mode config.
I think that this results in a spurious mismatch between the cohort/case mode configs, which causes PostprocessGermlineCNVCalls to emit the following warnings in case mode when non-default values are used:
WARNING gcnvkernel.postprocess.viterbi_segmentation - Different denoising configuration between model and calls -- proceeding at your own risk!
WARNING gcnvkernel.postprocess.viterbi_segmentation - Different calling configuration between model and calls -- proceeding at your own risk!
I'm pretty sure that inference is actually performed correctly, but we may want to double check and clean up these warnings. We should probably just copy the non-exposed values from the model config on the python side when running GermlineCNVCaller in case mode.
Not sure if there's any way to emit sensible warnings on the Java side. These hyperparameters are still exposed to the Java command line in case mode, they just aren't passed on to the python command line. So the user can change their values from their engine defaults without having any effect at all, but this is probably what we want. Perhaps we can document, though.
The text was updated successfully, but these errors were encountered:
samuelklee
changed the title
Inaccurate warning about mismatching parameter configs in PostprocessGermlineCNVCalls.
Possibly inaccurate warning about mismatching parameter configs in PostprocessGermlineCNVCalls.
Dec 10, 2020
PostprocessGermlineCNVCalls performs a check of the denoising/calling hyperparameter configs used to generate the model in GermlineCNVCaller cohort mode against those used to generate the case-mode result passed to PostprocessGermlineCNVCalls. However, although some of these hyperparameters are not exposed in case mode (since they have no effect on the sample-level parameters inferred in case mode, e.g.,
psi_t_scale
), their python default values are nevertheless written to the case-mode config.I think that this results in a spurious mismatch between the cohort/case mode configs, which causes PostprocessGermlineCNVCalls to emit the following warnings in case mode when non-default values are used:
I'm pretty sure that inference is actually performed correctly, but we may want to double check and clean up these warnings. We should probably just copy the non-exposed values from the model config on the python side when running GermlineCNVCaller in case mode.
Not sure if there's any way to emit sensible warnings on the Java side. These hyperparameters are still exposed to the Java command line in case mode, they just aren't passed on to the python command line. So the user can change their values from their engine defaults without having any effect at all, but this is probably what we want. Perhaps we can document, though.
The text was updated successfully, but these errors were encountered: