AttributeError: 'Scratchpad' object has no attribute 'ufunc'

Questions v5
1

Hello. I am new to PyMC. I wrote a state space model in a straight-forward fashion and ran it, then I encountered the error.
Minimal code is this:

import pymc as pm
import numpy as np

N = 32  # N >= 32 causes "AttributeError: 'Scratchpad' object has no attribute 'ufunc'"
Y = np.random.randint(0, 20, size=N)

with pm.Model() as model:
    alpha = [pm.Uniform('alpha[1]', 0, 10)]
    sigma = pm.Uniform('sigma', 0, 100)

    for i in range(1, N):
        alpha.append(pm.Normal(f'alpha[{i+1}]', mu=alpha[i-1], sigma=sigma))

    for i in range(N):
        l = pm.Deterministic(f'lambda[{i+1}]', alpha[i].exp())
        Yobs = pm.Poisson(f'Y[{i+1}]', mu=l, observed=Y[i])

with model:
    trace = pm.sample(1000, progressbar=False)

Here is the error message:

AttributeError: ‘Scratchpad’ object has no attribute ‘ufunc’

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File “/Users/ynomura/tmp/pymc_test4.py”, line 19, in
trace = pm.sample(1000, progressbar=False)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pymc/sampling/mcmc.py”, line 745, in sample
initial_points, step = init_nuts(
^^^^^^^^^^
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pymc/sampling/mcmc.py”, line 1426, in init_nuts
initial_points = _init_jitter(
^^^^^^^^^^^^^
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pymc/sampling/mcmc.py”, line 1317, in _init_jitter
point = ipfn(seed)
^^^^^^^^^^
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pymc/initial_point.py”, line 169, in inner
values = func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pytensor/compile/function/types.py”, line 970, in call
self.vm()
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pytensor/link/vm.py”, line 411, in call
raise_with_op(self.fgraph, node, thunk)
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pytensor/link/utils.py”, line 528, in raise_with_op
raise exc_value.with_traceback(exc_trace)
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pytensor/link/vm.py”, line 407, in call
thunk()
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pytensor/graph/op.py”, line 524, in rval
r = p(n, [x[0] for x in i], o)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pytensor/tensor/elemwise.py”, line 747, in perform
ufunc = node.tag.ufunc
^^^^^^^^^^^^^^
File “/Users/ynomura/miniforge3/envs/pymc_env/lib/python3.12/site-packages/pytensor/graph/utils.py”, line 286, in getattribute
return super().getattribute(name)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: ‘Scratchpad’ object has no attribute ‘ufunc’
Apply node that caused the error: Add(Mul.0, alpha[2]_jitter, alpha[3]_jitter, alpha[4]_jitter, alpha[5]_jitter, alpha[6]_jitter, alpha[7]_jitter, alpha[8]_jitter, alpha[9]_jitter, alpha[10]_jitter, alpha[11]_jitter, alpha[12]_jitter, alpha[13]_jitter, alpha[14]_jitter, alpha[15]_jitter, alpha[16]_jitter, alpha[17]_jitter, alpha[18]_jitter, alpha[19]_jitter, alpha[20]_jitter, alpha[21]_jitter, alpha[22]_jitter, alpha[23]_jitter, alpha[24]_jitter, alpha[25]_jitter, alpha[26]_jitter, alpha[27]_jitter, alpha[28]_jitter, alpha[29]_jitter, alpha[30]_jitter, alpha[31]_jitter, alpha[32]_jitter)
Toposort index: 65
Inputs types: [TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=())]
Inputs shapes: [(), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), ()]
Inputs strides: [(), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), (), ()]
Inputs values: [array(7.28997374), array(0.38199763), array(-0.1476399), array(-0.24766453), array(-0.24406243), array(-0.58236784), array(-0.00086235), array(0.52218202), array(-0.27085093), array(0.35206935), array(0.86599777), array(-0.76071219), array(-0.13915781), array(-0.980413), array(0.68555598), array(-0.20443367), array(-0.09218605), array(-0.8407623), array(0.68579167), array(0.63068857), array(0.46683988), array(-0.01151997), array(0.66353266), array(-0.46197365), array(-0.00129028), array(-0.43414777), array(0.94450627), array(0.56740174), array(0.84685636), array(-0.29390383), array(0.95625211), array(-0.49638046)]
Outputs clients: [[‘output’]]

HINT: Re-running with most PyTensor optimizations disabled could provide a back-trace showing when this node was created. This can be done by setting the PyTensor flag ‘optimizer=fast_compile’. If that does not work, PyTensor optimizations can be disabled with ‘optimizer=None’.
HINT: Use the PyTensor flag exception_verbosity=high for a debug print-out and storage map footprint of this Apply node.

This error doesn’t occur if N <= 31.

I around the line causing this error in Elemwise.perform, elemwise.py in PyTensor and found that calling self.prerapre_node doesn’t set node.tag.ufunc.
There is a comment saying “NumPy ufunc support only up to 32 operands (inputs and outputs) But our c code support more.” in the top of prepare_node() but I couldn’t find the c code.

Is there any workaround?
(I know I can use pm.GaussianRandomWalk for this example, but that is not what I mean)

Thanks in advance.

PyMC version: 5.15.1
PyTensor version: 2.22.1
Python version: 3.12.3
I Installed PyMC with conda.

2

For looping in pytensor, you should use the scan Op. See here for the docs and examples of usage

3

Hi Jesse,

Thank you so much for your advice. I learned about pytensor.scan and the “boilerplate” code you provided in Declaring Priors using Loops for State Space Models - #2 by jessegrabowski, and they solved my problem.

Revised code which worked without error:

import numpy as np
import pymc as pm
from pymc.pytensorf import collect_default_updates
import pytensor
import pytensor.tensor as pt

N = 32
Y = np.random.randint(0, 20, size=N)

# Helper function for pm.CustomDist
n_steps = N - 1
def statespace_dist(mu_init, sigma_level, size):

    def grw_step(mu_tm1, sigma_level):
        mu_t = mu_tm1 + pm.Normal.dist(sigma=sigma_level)
        return mu_t, collect_default_updates(outputs=[mu_t])

    mu, updates = pytensor.scan(fn=grw_step, 
                                outputs_info=[{"initial": mu_init}],
                                non_sequences=[sigma_level], 
                                n_steps=n_steps,
                                name='statespace',
                                strict=True)

    return mu

with pm.Model() as model:
    alpha1 = pm.Uniform('alpha[1]', 0, 10)
    sigma = pm.Uniform('sigma', 0, 100)

    alpha = pm.CustomDist('alpha', alpha1, sigma, dist=statespace_dist)
    alpha = pt.concatenate([[alpha1], alpha])

    l = pm.Deterministic('lambda', alpha.exp())
    Yobs = pm.Poisson('Y', mu=l, observed=Y)

with model:
    trace = pm.sample(1000, progressbar=False)
1 Like