Some modification required for source-dependent MC production

[chaimain]

Following the discussion in #188 we will need to add --source-dep option for generating IRFs, and maybe some more things, as the discussion with @SeiyaNozaki progresses, regarding producing source dependent MC production.

[LukasNickel]

Choosing pointlike gammas for RF right? Anything else? Lets discuss it in the call

[SeiyaNozaki]

I think we need to modify for source-dependent analysis: https://github.com/cta-observatory/lstmcpipe/blob/master/lstmcpipe/stages/mc_train.py#L52-L53 to use point gamma MC as RF training input and https://github.com/cta-observatory/lstmcpipe/blob/master/lstmcpipe/stages/mc_dl2_to_irfs.py#L284 to add --source-dep flag if soure-dep analysis is selected

[LukasNickel]

Pointlike gammas for all random forests. Is that a good idea?

Is it enough to put source dependent in the config and use the normal for dl1/dl2 or is there something missing?

[SeiyaNozaki]

Pointlike gammas for all random forests. Is that a good idea?

yes, this is the way for source-dependent analysis. It would make some bias for proton/electrom MC, but we can obtain better performance for point gamma source.

Is it enough to put source dependent in the config and use the normal for dl1/dl2 or is there something missing?

It should be enough only to activate source-dep flag in the config.

[vuillaut]

I believe this can now be easily supported by the new PathConfig logic. Is there still a need @SeiyaNozaki @chaimain ?

[vuillaut]

Hi @SeiyaNozaki With the AllSky production, what gamma do you use for training for the source-dep analysis? (since the TrainingDataset includes only GammaDiffuse)

[SeiyaNozaki]

For all sky MC, I'm using diffuse gamma as RF training for source-dependent analysis (though I used point gamma before all sky MC).

[vuillaut]

For all sky MC, I'm using diffuse gamma as RF training for source-dependent analysis (though I used point gamma before all sky MC).

ha ok great, so #375 is in line with your processing and it also means there is no changes required in lstmcpipe to process source-dependent !! You can request a prod for your source-dep analysis, the only difference is in the lstchain config.

[SeiyaNozaki]

Hi @vuillaut , I want to ask base production for source-dependent analysis (like #449 ), but in general, I also want to improve the code for source-dependent analysis:

• Now source-dependent parameters are automatically added during RF training step and DL1 to DL2 step. So we can re-use DL1 training MC from the standard base production. We just need to process Training step and TestMC production. It would be nice to add a new config_class including train_pipe, dl1_to_dl2, dl2_to_irfs
• Somehow I also want to add an input argument for lstmcpipe_generate_config for source-dependent analysis, then automatically source-dependent MC data are loaded. I think we need to modify somewhere related to the following function, but I want to hear your opinion! https://github.com/cta-observatory/lstmcpipe/blob/74919f0c0c726acee7efe57c9c183d52806fc2ca/lstmcpipe/utils.py#L156 https://github.com/cta-observatory/cta-lstchain/blob/7c13ed0333eafadcefed17d5ec899f3c75b5f34e/lstchain/io/config.py#L47-L72