add a new generic class to combine classifiers

[jecampagne]

In tomo_challenge/classifiers/jec_GB.py I have define a Gradient Boosting DT classifier, and in tomo_challenge/classifiers/jec_CombineClassifier.py I have define a rather simple way to combine 2 classifiers.
To run these new classes in I have created example/jec_GB.yaml and example/jec_multiclf.yaml files

[joezuntz]

Could you uncomment the things in the requirements file, as they help the continuous integration work. I'll add something else describing the conda installation.

[jecampagne]

Concerning the results as a general remark based on several tests, the FOM_3x2 seems to have some very strange behaviour. Now If I focus on SNR_3x2 my first two best results are (23June20-14:45Paris):

• myCombinedClassifiers: 50 Gradient Boost Models combined with 50 Random Forest Models with 6 redshift bins, griz bands: 'SNR_3x2': 1528.2427120796633,

• myGradientBosster with 6 redshift bins, griz bands, and 1000 models: 'SNR_3x2': 1524.0354116695296

For these two settings the FOM_3x2 is of the order of 11000, but I get an incredible 114325 score with an other settings....

[jecampagne]

Could you uncomment the things in the requirements file, as they help the continuous integration work. I'll add something else describing the conda installation.

I have "uncommented" the things that were not working in my context.

[jecampagne]

With DC2 data and JAX from @EiffL

Then, if one optimizes the "bins", one gets different sets for SNR_3x2, FOM_3x2 or FOM_DETF_3x2 optimization. For instance, for GRIZ, nbins=10, GB+RF (50 estimators each), when optimizing FOM__DETF_3x2
SNR_3x2 : 1770.6
FOM_3x2 : 11953.9
FOM__DETF_3x2 : 176.7
one gets the following "bins"

[jecampagne]

Buzzard data & JAX metrics.
GB+RF (50/50 estimators), GRIZ bands, 10 bins with equal n(z) bins: SNR_3x2 = 1742.0, FOM_3x2=6618.7, FOM_DETF_3x2 = 91.2 which means a drastic reduction form DC2 data. Optim bins to be done later.

[EiffL]

Ohhh nice! This is what we were hoping ^^' You beat me to it haha. Very curious to see what the resdhift distributions look like now.

[jecampagne]

Here to compare with @EiffL I use the data "Buzzard", bands RIZ, 6 bins using both the colours ans the errors with GB+RF classifiers (50+50 estimators) :

• using bins driven by equal n(z) I get:
  SNR_3x2 = 1460.30, FOM_3x2: 3437.56, FOM_DETF_3x2: 64.49
• Now after an optimization to get higher FOM_DETF_3x2, here is the results
  'SNR_3x2': 1394.09814453125, 'FOM_3x2': 3745.437255859375, 'FOM__DETF_3x2': 69.76288604736328}
  which leads to the following "bins"
  

[jecampagne]

Here a new result: with Buzzard data, opitmizing "10 bins" for FOM_DETF_3x2, GRIZ bands & no errors

This can be directly compared to the DC2 data