ipyparallel string/unicode problems

[noahaskell]

I'm trying to fit pymc2 models in parallel using ipyparallel on a Mac Pro, and I'm getting weird errors related to strings. Here's my system information:

python -c "import IPython; print(IPython.sys_info())"
{'commit_hash': u'b573435',
 'commit_source': 'installation',
 'default_encoding': 'UTF-8',
 'ipython_path': '/Users/noahsilbert/anaconda/envs/py2/lib/python2.7/site-packages/IPython',
 'ipython_version': '4.1.2',
 'os_name': 'posix',
 'platform': 'Darwin-15.4.0-x86_64-i386-64bit',
 'sys_executable': '/Users/noahsilbert/anaconda/envs/py2/bin/python',
 'sys_platform': 'darwin',
 'sys_version': '2.7.11 |Anaconda 4.0.0 (x86_64)| (default, Dec  6 2015, 18:57:58) \n[GCC 4.2.1 (Apple Inc. build 5577)]'}

I'm using pymc version 2.3.6. Here's an example script that produces the error(s) in question:

import pymc as pm
import numpy as np
import time

N = 200
x = np.random.normal(loc=5,scale=2,size=N)
X = np.ones((N,2))
X[:,1] = x

beta = np.array([10,-3])

e = np.random.normal(scale=5,size=N) 

y = np.dot(X,beta) + e

def regression(yy,XX):

    a = pm.Normal('a',mu=0,tau=1,size=1)
    b = pm.Normal('b',mu=0,tau=1,size=1)
    t = pm.InverseGamma('t',alpha=1,beta=1,size=1)

    @pm.deterministic()
    def y_hat(Xm=XX,A=a,B=b):
        y_hat = A*XX[:,0] + B*XX[:,1]
        return y_hat

    Y = pm.Normal('Y', mu=y_hat, tau=t, observed=True, value=yy)

nchain = 1
nkeep = 200
nburn = 500
nthin = 10
niter = nburn + nkeep*nthin

clock_string = str(int(time.time()*1000000))[-6:]
date_string = time.ctime().split()[1] + '_' + time.ctime().split()[2]
time_string = date_string + '_' + clock_string

db_name = 'regression_MC_' + time_string + '.hdf5'
MC = pm.MCMC(input=regression(yy=y,XX=X),db='hdf5',dbname=db_name,dbmode='w',verbose=1)
MC.use_step_method(pm.AdaptiveMetropolis,MC.stochastics)
MC.sample(iter=niter,burn=nburn,thin=nthin)
MC.db.close()

I have that saved as test_script.py, and then I try to run the following script, either from the command prompt or in IPython:

import ipyparallel as ipp
c = ipp.Client()
d = c[:]
d.block = True
model_script = open('test_script.py').read()
d.execute(model_script)

At which point I get the following Traceback:

---------------------------------------------------------------------------ValueError                                Traceback (most recent call last)<ipython-input-1-cd0f32256b19> in <module>()
     38
     39 db_name = 'regression_MC_' + time_string + '.hdf5'
---> 40 MC = pm.MCMC(input=regression(yy=y,XX=X),db='hdf5',dbname=db_name,dbmode='w',verbose=1)
     41 MC.use_step_method(pm.AdaptiveMetropolis,MC.stochastics)
     42 MC.sample(iter=niter,burn=nburn,thin=nthin)
<ipython-input-1-cd0f32256b19> in regression(yy, XX)
     16 def regression(yy,XX):
     17
---> 18     a = pm.Normal('a',mu=0,tau=1,size=1)
     19     b = pm.Normal('b',mu=0,tau=1,size=1)
     20     t = pm.InverseGamma('t',alpha=1,beta=1,size=1)
/Users/noahsilbert/anaconda/envs/py2/lib/python2.7/site-packages/pymc/distributions.pyc in __init__(self, *args, **kwds)
    318                     logp_partial_gradients=logp_partial_gradients,
    319                     dtype=dtype,
--> 320                     **arg_dict_out)
    321
    322     new_class.__name__ = name
/Users/noahsilbert/anaconda/envs/py2/lib/python2.7/site-packages/pymc/PyMCObjects.pyc in __init__(self, logp, doc, name, parents, random, trace, value, dtype, rseed, observed, cache_depth, plot, verbose, isdata, check_logp, logp_partial_gradients)
    762                           dtype=dtype,
    763                           plot=plot,
--> 764                           verbose=verbose)
    765
    766         # self._logp.force_compute()
/Users/noahsilbert/anaconda/envs/py2/lib/python2.7/site-packages/pymc/Node.pyc in __init__(self, doc, name, parents, cache_depth, trace, dtype, plot, verbose)
    212         self.extended_children = set()
    213
--> 214         Node.__init__(self, doc, name, parents, cache_depth, verbose=verbose)
    215
    216         if self.dtype is None:
/Users/noahsilbert/anaconda/envs/py2/lib/python2.7/site-packages/pymc/Node.pyc in __init__(self, doc, name, parents, cache_depth, verbose)
    117             raise ValueError(
    118                 'The name argument must be a string, but received %s.' %
--> 119                 name)
    120         self.__name__ = name
    121
ValueError: The name argument must be a string, but received a.

If I replace the string in line 18 with str('a') (and similarly for the other named variables in the model), I then get this error:

---------------------------------------------------------------------------AttributeError                            Traceback (most recent call last)<ipython-input-1-8041b7b164c1> in <module>()
     38
     39 db_name = 'regression_MC_' + time_string + '.hdf5'
---> 40 MC = pm.MCMC(input=regression(yy=y,XX=X),db='hdf5',dbname=db_name,dbmode='w',verbose=1)
     41 MC.use_step_method(pm.AdaptiveMetropolis,MC.stochastics)
     42 MC.sample(iter=niter,burn=nburn,thin=nthin)
/Users/noahsilbert/anaconda/envs/py2/lib/python2.7/site-packages/pymc/MCMC.pyc in __init__(self, input, db, name, calc_deviance, **kwds)
     80             name,
     81             calc_deviance=calc_deviance,
---> 82             **kwds)
     83
     84         self._sm_assigned = False
/Users/noahsilbert/anaconda/envs/py2/lib/python2.7/site-packages/pymc/Model.pyc in __init__(self, input, db, name, reinit_model, calc_deviance, verbose, **kwds)
    205         # Specify database backend and save its keywords
    206         self._db_args = kwds
--> 207         self._assign_database_backend(db)
    208
    209         # Flag for model state
/Users/noahsilbert/anaconda/envs/py2/lib/python2.7/site-packages/pymc/Model.pyc in _assign_database_backend(self, db)
    583             self.restore_sampler_state()
    584         else:   # What is this for? DH.
--> 585             self.db = db.Database(**self._db_args)
    586
    587     def pause(self):
AttributeError: 'unicode' object has no attribute 'Database'

If I change line 39 to db_name = str('regression_MC_' + time_string + '.hdf5'), or if I try to use str() with the individual parts of time_string, I get the same error.

This only happens on my Mac Pro. I can get pymc2 to cooperate with ipyparallel on an iMac, for example. But given the larger number of processors in the Mac Pro, I would dearly love to get this working on that computer. Thanks.

[fonnesbeck]

Can you check to see that the quotes around 'a' are actual ascii quotes and not unicode quotes? This can sometimes happen if you've cut an pasted code from a pdf or a website, for example. Alternately you could try install in Python 3, which I recommend anyway. Easy to do via Anaconda.

[noahaskell]

I typed it all out in vim, so I'm not sure how the quotes could be unicode and not ascii. But how can I check to make sure?

Does pymc2 work with Python 3? The model I'm actually trying to fit uses an imported multivariate normal cdf function in one of the deterministic nodes. I have a version of it implemented in pymc3 (using Python 3), but I have to use metropolis sampling because of the cdf function (which is used in an as_op function). The added compilation time for pymc3 doesn't seem to be worth it, and pymc2 seems to be working fairly well for this model. I just want to fit it in parallel.

I can invoke multiple instances of IPython and run the script from each of those, and it works fine, even on my Mac Pro, but I don't know if that's actually running the chains in parallel in a useful way.

[fonnesbeck]

The version numbers of PyMC and Python are not related to one another. I haven't used Python 2.7 for a couple years now.

[noahaskell]

Okay, good to know. I thought I had run into problems trying to use PyMC 2 with Python 3, but maybe I'm misremembering something. I just recently upgraded to Python 3.

[noahaskell]

Everything works fine in Python 3. For what it's worth, installing PyMC2 broke my PyMC3 installation, but I uninstalled PyMC3 and Theano then reinstalled PyMC3 (which also reinstalled Theano, of course), and both PyMC2 and PyMC3 seem to work fine.