issue with constraint_logpdf(auxdata, pars)

[IzaV10]

Hi,

I have some problem using constraint_logpdf(auxdata, pars). I setup the example like this:

inits = inModel_SM.config.suggested_init()
inits[ inModel_SM.config.poi_index ] = 1.0

exp_data_model=inModel_SM.expected_data( inits ) 
auxdata_sm=inModel_SM.expected_auxdata( inits )

inModel_SM.logpdf(inits, exp_data_model))
inModel_SM.constraint_logpdf(auxdata=auxdata_sm,pars=inits)

While inModel_SM.logpdf(inits, exp_data_model)) works fine I get the following error when calling constraint_logpdf(auxdata, pars)

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-45-af486baff749> in <module>
----> 1 inModel_SM.constraint_logpdf(auxdata=auxdata_sm,pars=inits)

/cvmfs/sft.cern.ch/lcg/views/LCG_100/x86_64-centos7-gcc8-opt/lib/python3.8/site-packages/pyhf/pdf.py in constraint_logpdf(self, auxdata, pars)
    667 
    668         """
--> 669         return self.make_pdf(pars)[1].log_prob(auxdata)
    670 
    671     def mainlogpdf(self, maindata, pars):

/cvmfs/sft.cern.ch/lcg/views/LCG_100/x86_64-centos7-gcc8-opt/lib/python3.8/site-packages/pyhf/pdf.py in make_pdf(self, pars)
    700         if mainpdf:
    701             pdfobjs.append(mainpdf)
--> 702         constraintpdf = self.constraint_model.make_pdf(pars)
    703         if constraintpdf:
    704             pdfobjs.append(constraintpdf)

/cvmfs/sft.cern.ch/lcg/views/LCG_100/x86_64-centos7-gcc8-opt/lib/python3.8/site-packages/pyhf/pdf.py in make_pdf(self, pars)
    376         pdfobjs = []
    377 
--> 378         gaussian_pdf = self.constraints_gaussian.make_pdf(pars)
    379         if gaussian_pdf:
    380             pdfobjs.append(gaussian_pdf)

/cvmfs/sft.cern.ch/lcg/views/LCG_100/x86_64-centos7-gcc8-opt/lib/python3.8/site-packages/pyhf/constraints.py in make_pdf(self, pars)
    105             flat_pars = tensorlib.reshape(pars, (-1,))
    106 
--> 107         normal_means = tensorlib.gather(flat_pars, self.access_field)
    108 
    109         # pdf pars are done, now get data and compute

/cvmfs/sft.cern.ch/lcg/views/LCG_100/x86_64-centos7-gcc8-opt/lib/python3.8/site-packages/pyhf/tensor/jax_backend.py in gather(self, tensor, indices)
    190 
    191     def gather(self, tensor, indices):
--> 192         return tensor[indices]
    193 
    194     def boolean_mask(self, tensor, mask):

/cvmfs/sft.cern.ch/lcg/views/LCG_100/x86_64-centos7-gcc8-opt/lib/python3.8/site-packages/jax/interpreters/xla.py in _forward_method(attrname, self, fun, *args)
   1008 
   1009 def _forward_method(attrname, self, fun, *args):
-> 1010   return fun(getattr(self, attrname), *args)
   1011 _forward_to_value = partial(_forward_method, "_value")
   1012 

TypeError: only integer scalar arrays can be converted to a scalar index

Thanks for your help!

Cheers,
Iza

[matthewfeickert]

@IzaV10 Sorry I missed this — just seeing it now. The problem is that pyhf.pdf.Model.constraint_logpdf requires tensor objects and your inits is a list. I understand your confusion as to why logpdf works but constraint_logpdf doesn't and that's our fault for making sure that some parts of the API will work even with non-tensors when we should strictly enforce tensors everywhere.

The good news is that this is very simple to fix: just wrap your inits in pyhf.tensorlib.astensor.

As a full example of this, if we take the "Hello World" example from the README and modify it a bit

# discussion_1608.py 
import pyhf

pyhf.set_backend("jax")

model = pyhf.simplemodels.uncorrelated_background(
    signal=[12.0, 11.0], bkg=[50.0, 52.0], bkg_uncertainty=[3.0, 7.0]
)
data = [51, 48] + model.config.auxdata

init_params = model.config.suggested_init()
init_params[model.config.poi_index] = 1.0
init_params = pyhf.tensorlib.astensor(init_params)

expected_data = model.expected_data(init_params)
expected_auxdata = model.expected_auxdata(init_params)

print(f"model.logpdf: {model.logpdf(init_params, expected_data)}")
print(
    f"model.constraint_logpdf: {model.constraint_logpdf(expected_auxdata, pars=init_params)}"
)

we get

$ python discussion_1608.py 
model.logpdf: [-12.63410988]
model.constraint_logpdf: -6.658431444518868

The above is done with pyhf v0.6.3, and from your output I can see that you're using the version of pyhf from what looks like LCV view LCG_100 on CVMFS which we don't recommend as it is an old version of pyhf (we don't do the distribution of pyhf of CVMFS)

[feickert@login ~]$ lsetup "views LCG_100 x86_64-centos7-gcc8-opt"
************************************************************************
Requested:  views ... 
 Setting up views LCG_100:x86_64-centos7-gcc8-opt ... 
>>>>>>>>>>>>>>>>>>>>>>>>> Information for user <<<<<<<<<<<<<<<<<<<<<<<<<
************************************************************************
[feickert@login ~]$ pyhf --version
pyhf, version 0.5.4

Though in this case it doesn't matter as the API for pyhf.pdf.Model.constraint_logpdf in v0.5.4 is the same. 👍

Hope that helps.

[kratsg]

@IzaV10 if this fixes it, please mark it as the answer for your discussion. As Matthew pointed out, we have not perfected our API everywhere so there are rough edges that could be smoothed out.