Coda_output function doesn't work.

[griff122]

Hi,

So after going through the utils.coda_output function I noticed that observed nodes are being skipped with a try/except statement. The caught on the except is a TypeError, and I've noticed that when an observed node is used, I doesn't have the attribute .trace() , so an AttributeError is thrown instead. Since no catch is made, the program crashes.

I set the information in the gelman_biossary.py in a function called tester(), which returns locals():

Here are my results...

>>> from pymc.examples import gelman_bioassay
>>> bar = pymc.MCMC(gelman_bioassay.tester(), db="ram")
>>> bar.stochastics
set([<pymc.distributions.Normal 'alpha' at 0x000000001673C780>, <pymc.distributions.Normal 'beta' at 0x000000001673C128>])
>>> bar.sample(100)
[                  0%                  ] 
[****************100%******************]  100 of 100 complete 

>>> pymc.utils.coda_output(bar)

Generating CODA output
==================================================
Processing alpha
Processing beta
Processing LD50
Processing deaths

Traceback (most recent call last):
  File "<pyshell#174>", line 1, in <module>
    pymc.utils.coda_output(bar)
  File "C:\Python27\lib\site-packages\pymc\utils.py", line 726, in coda_output
    index = _process_trace(trace_file, index_file, v.trace(), vname, index)
AttributeError: 'Binomial' object has no attribute 'trace'

I've made an edit, which will remove all observed nodes from the variables list, before the loop:

def coda_output(pymc_object):
    """Generate output files that are compatible with CODA

    :Arguments:
      pymc_object : Model or Node
          A PyMC object containing MCMC output.

    """

    print_()
    print_("Generating CODA output")
    print_('='*50)

    name = pymc_object.__name__

    # Open trace file
    trace_file = open(name+'_coda.out', 'w')

    # Open index file
    index_file = open(name+'_coda.ind', 'w')

    ## observedNodes is a list of any observed node name in the model
    observedNodes = [node.__name__ for node in pymc_object.observed_stochastics]

    variables = [pymc_object]
    if hasattr(pymc_object, 'variables'):
        ## changed this call to the one below
        #variables = pymc_object.variables
        variables = [v for v in pymc_object.variables if v.__name__ not in observedNodes]


    # Initialize index
    index = 1

    # Loop over all parameters
    for v in variables:    
        vname = v.__name__
        print_("Processing", vname)
        try:
            index = _process_trace(trace_file, index_file, v.trace(), vname, index)
        except TypeError:
            pass

    # Close files
    trace_file.close()
    index_file.close()

# Lazy shortcut
coda = coda_output

This works for me. I'm not sure if anyone else has had issues with this, but here's a fix.

[griff122]

Also I believe (haven't tested yet though) that if you change the "except TypeError:" to "except Exception:" it will work as well.

[fonnesbeck]

Can you please post your code for tester? I currently cannot replicate this behavior. What version of PyMC are you running?

[griff122]

I am using pymc2.2 64bit.
Here is the code for tester,

from pymc import *
from numpy import ones, array

def tester():
    # Samples for each dose level
    n = 5*ones(4,dtype=int)
    # Log-dose
    dose=array([-.86,-.3,-.05,.73])

    # Logit-linear model parameters
    alpha = Normal('alpha', 0, 0.01)
    beta = Normal('beta', 0, 0.01)

    # Calculate probabilities of death
    theta = Lambda('theta', lambda a=alpha, b=beta, d=dose: invlogit(a + b*d))

    # Data likelihood
    deaths = Binomial('deaths', n=n, p=theta, value=array([0,1,3,5], dtype=float), observed=True)

    # Calculate LD50
    LD50 = Lambda('LD50', lambda a=alpha, b=beta: -a/b)

    return locals()

[fonnesbeck]

Can you build a version from the GitHub repository and see if it goes away? The code in tester() does not generate any sort of error when I run coda_output on the model.

[griff122]

I wasn't able to get the cloned zip to build on windows, but I had used the version from here for the 32bit code:

http://pypi.python.org/pypi/pymc/

I got the 64bit from:

http://www.lfd.uci.edu/~gohlke/pythonlibs/#pyzmq
which was updated May 7th, 2012

[gaborgulya]

I have the same problem. I could use the above workaround code after the following preparations:

def print_(*args):
  print args

_process_trace = pymc.utils._process_trace

[fonnesbeck]

The test code passes in the current master. Closing.