TypeError: zip argument #1 must support iteration

[aakhmetz]

Description of your problem

Please provide a minimal, self-contained, and reproducible example.

import theano
from theano.compile.ops import as_op
from theano import shared
import theano.tensor as tt
import pymc3 as pm
import numpy as np
from scipy.special import gammainc

@as_op(itypes=[tt.dscalar, tt.dscalar, tt.lvector], otypes=[tt.dmatrix])
def pgamma_for_projection(delay_mean, delay_sd, taus):
    τs_to_forecast = np.append([0],taus[taus<0])
    
    res = np.array([],dtype=np.float64).reshape(0,τs_to_forecast.shape[0])
    for τ in taus[taus>=0]:
        res0 = gammainc((delay_mean/delay_sd)**2, delay_mean*(τ-τs_to_forecast)/(delay_sd**2))
        res0 = np.r_[np.diff(res0),1-res0[-1]]
        res = np.vstack([res,res0])
    return res

τs = np.arange(17,-28,-7,dtype=np.int64)

with pm.Model() as model:
    delaymean = pm.Uniform('delaymean',13,16)
    delaysd = pm.Uniform('delaysd',7,10)
    
    trace = pm.sample(1000, tune=200)
    
    probs = pm.Deterministic('probs',pgamma_for_projection(delaymean,delaysd,shared(τs)))
    counts = pm.NegativeBinomial('counts',10,probs)
    ttl = pm.Deterministic('ttl',counts)

    ppc = pm.sample_posterior_predictive(trace, vars=[probs,counts,ttl], samples=500)

Please provide the full traceback.

Auto-assigning NUTS sampler...
Initializing NUTS using jitter+adapt_diag...
Multiprocess sampling (2 chains in 2 jobs)
NUTS: [delaysd, delaymean]
Sampling 2 chains: 100%|██████████| 2400/2400 [00:01<00:00, 1706.70draws/s]
  0%|          | 0/500 [00:00<?, ?it/s]
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-72-1a6f7c9ddc3b> in <module>
     22     ttl = pm.Deterministic('ttl',counts)
     23 
---> 24     ppc = pm.sample_posterior_predictive(trace, vars=[probs,counts,ttl], samples=500)

~/anaconda3/lib/python3.6/site-packages/pymc3/sampling.py in sample_posterior_predictive(trace, samples, model, vars, size, random_seed, progressbar)
   1102                 param = trace[idx % len_trace]
   1103 
-> 1104             values = draw_values(vars, point=param, size=size)
   1105             for k, v in zip(vars, values):
   1106                 ppc_trace[k.name].append(v)

~/anaconda3/lib/python3.6/site-packages/pymc3/distributions/distribution.py in draw_values(params, point, size)
    394                                             point=point,
    395                                             givens=givens.values(),
--> 396                                             size=size)
    397                         evaluated[param_idx] = drawn[param] = value
    398                         givens[param.name] = (param, value)

~/anaconda3/lib/python3.6/site-packages/pymc3/distributions/distribution.py in _draw_value(param, point, givens, size)
    502                 not all(var.dshape == getattr(val, 'shape', tuple())
    503                         for var, val in zip(variables, values))):
--> 504                 output = np.array([func(*v) for v in zip(*values)])
    505             elif (size is not None and any((val.ndim > var.ndim)
    506                   for var, val in zip(variables, values))):

TypeError: zip argument #1 must support iteration

Please provide any additional information below.

I wanted to sample sums over columns of counts, but sum() didn't work for me and then even simple creation of Deterministic node led to a Type error.

Versions and main components

• PyMC3 Version: 3.6
• Theano Version: 1.0.3
• Python Version: 3.6.6
• Operating system: osX
• How did you install PyMC3: (conda/pip) pip with the latest version from github

[twiecki]

It seems to have problems with sending samples through the determinstic. Maybe @lucianopaz has an idea?

[lucianopaz]

I'll look into this in more detail but from a first glance, I think I had run into this also in the PR #3293. The problem I had run into was that the givens dictionary included some variables that were not ancestors of the variable that was going to be compiled and then executed. These non-ancestors could be scalars or things whose shape did not broadcast well with the true ancestors, and that lead to the zip error. If this error ends up being caused by the same thing, then #3293 should fix it.

[lucianopaz]

I checked the issue more in depth. The problem is caused because counts is not given a shape during its construction. In the present implementation of random variates and distributions done in pymc3, the variable's shape is not inferred through broadcasting of the distribution's parameters. The only solution to explicitly set the variable's shape by hand during its construction like this:

...
    counts = pm.NegativeBinomial('counts', 10, probs,
                                 shape=(np.sum(τs >= 0),
                                        np.sum(τs < 0) + 1))
...

This is a limitation of pymc3's implementation of random variables and distributions. @brandonwillard's ideas of writing random variates as theano.Op's would eventually solve problems like this one, but I think that this issue can be closed.

[aakhmetz]

Thanks for checking! Sorry I haven't thought of that shape-problem. Probably, it would be a nice practice of coding just to define the shape parameter everywhere, so that possible errors could be avoided