Extra srcdep parameters and limit the offset of gamma MC for training #1077
Conversation
…rectness) for source-dep analysis and limit the offset of gamma MC for the RF training
| @@ -478,10 +478,13 @@ def build_models(filegammas, fileprotons, | |||
| # Apply the temporary disp norm regressor and sign classifier to the test set | |||
| disp_norm = tmp_reg_disp_norm.predict(test[config['disp_regression_features']]) | |||
| disp_sign = tmp_cls_disp_sign.predict(test[config['disp_classification_features']]) | |||
| disp_sign_proba = tmp_cls_disp_sign.predict_proba(test[config['disp_classification_features']]) | |||
There was a problem hiding this comment.
There is no need to call predict and predict proba, that's wasteful. Since we now want the proba, you should change the disp_sign calculation to use the proba to not evaluate the random forest twice:
col = list(tmp_cls_disp_sign.classes_).index(1)
disp_sign = np.where(disp_sign_proba[:, col] > 0.5, 1, -1)
There was a problem hiding this comment.
Thanks for the comment! Actually, I just copy & paste the codes of gammaness computation, so this part can also be updated to reduce computation time
https://github.com/cta-observatory/cta-lstchain/blob/master/lstchain/reco/dl1_to_dl2.py#L652-L653
|
CI failed since only a single event survives after off-axis cut (src_r < 1 degree) ... |
I think it was @maxnoe the one creating those datasets, right? Would it be possible to create one including more inner events now that we have seen an improvement in the performance of src-dependent analysis? |
|
so @maxnoe would it be possible to get a different test dataset with more off-axis events or shall we change the tests for them to pass after these conditions are applied? |
…to zeros) to keep events after src_r cut)
No sure what simulations you are talking about, but I can't remember creating any. If you need a new small test file with specific settings, you should probably ask the simulation team to create it, like I did here: |
|
ah I see Max's comment now, but now I prepared fake DL1 MC gamma file (like fake proton MC) by changing |
Codecov ReportPatch coverage:
Additional details and impacted files@@ Coverage Diff @@
## master #1077 +/- ##
==========================================
+ Coverage 74.02% 74.12% +0.10%
==========================================
Files 123 123
Lines 11869 11915 +46
==========================================
+ Hits 8786 8832 +46
Misses 3083 3083
☔ View full report in Codecov by Sentry. |
This PR implements extra two parameters for source-dep analysis defined in the following slides.
Any comments are welcome, especially the naming of the parameters...!
https://indico.cta-observatory.org/event/4631/contributions/38096/attachments/23113/33203/20230213_srcdep_update.pdf
In addition, this PR changes the default value of `src_r_cut of MC gamma for the RF training.