pm.Data breaks broadcasting of different shapes

[paw-lu]

Description of your problem

Typically when you add shapes that are mismatched, broadcasting is expected.

# Numpy example of expected broadcasting
>>> np.array([[1, 2]]) + np.array([[1], [2]])
array([[2, 3],
       [3, 4]])

This works as expected when inputting a numpy array, but when fed through via pm.Data, the broadcasting does not work as expected, and raises an exception on the mismatched shapes.

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

"""This works!"""
import numpy as np
import pymc3 as pm

raw_data = np.array([[1], [2], [3]])

with pm.Model() as model:
    val = pm.Normal("val", mu=0, sd=10, shape=(1, 2))
    pm.Deterministic("add", var=val + raw_data)

"""This does not work!"""
import numpy as np
import pymc3 as pm

raw_data = np.array([[1], [2], [3]])

with pm.Model() as model:
    val = pm.Normal("val", mu=0, sd=10, shape=(1, 2))
    data = pm.Data("data", raw_data)  # New line
    pm.Deterministic("add", var=val + data)

Please provide the full traceback.

Complete error traceback

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-231-ccd5d08fd0d4> in <module>
      4     val = pm.Normal("val", mu=0, sd=10, shape=(1, 2))
      5     data = pm.Data("data", raw_data)
----> 6     pm.Deterministic("add", var=val + data)

~/.cache/pypoetry/virtualenvs/time-series-demand-5LaNRWUQ-py3.9/lib/python3.9/site-packages/theano/tensor/var.py in __add__(self, other)
     99     def __add__(self, other):
    100         try:
--> 101             return theano.tensor.basic.add(self, other)
    102         # We should catch the minimum number of exception here.
    103         # Otherwise this will convert error when Theano flags

~/.cache/pypoetry/virtualenvs/time-series-demand-5LaNRWUQ-py3.9/lib/python3.9/site-packages/theano/graph/op.py in __call__(self, *inputs, **kwargs)
    251 
    252         if config.compute_test_value != "off":
--> 253             compute_test_value(node)
    254 
    255         if self.default_output is not None:

~/.cache/pypoetry/virtualenvs/time-series-demand-5LaNRWUQ-py3.9/lib/python3.9/site-packages/theano/graph/op.py in compute_test_value(node)
    128     thunk.outputs = [storage_map[v] for v in node.outputs]
    129 
--> 130     required = thunk()
    131     assert not required  # We provided all inputs
    132 

~/.cache/pypoetry/virtualenvs/time-series-demand-5LaNRWUQ-py3.9/lib/python3.9/site-packages/theano/graph/op.py in rval()
    604 
    605         def rval():
--> 606             thunk()
    607             for o in node.outputs:
    608                 compute_map[o][0] = True

~/.cache/pypoetry/virtualenvs/time-series-demand-5LaNRWUQ-py3.9/lib/python3.9/site-packages/theano/link/c/basic.py in __call__(self)
   1769                 print(self.error_storage, file=sys.stderr)
   1770                 raise
-> 1771             raise exc_value.with_traceback(exc_trace)
   1772 
   1773 

ValueError: Input dimension mis-match. (input[0].shape[1] = 2, input[1].shape[1] = 1)

Please provide any additional information below.

Versions and main components

• PyMC3 Version: 3.11.2
• Aesara/Theano Version: 1.1.2
• Python Version: 3.9.5
• Operating system: CentOS Stream 8
• How did you install PyMC3: (conda/pip) pip via poetry

[paw-lu]

Looked into it a bit more at it seems a quick fix is to use theano.shared and manually set the broadcastable arg.

Example:

"""This works!"""
import numpy as np
import pymc3 as pm
import theano

raw_data = np.array([[1], [2], [3]])

with pm.Model() as model:
    val = pm.Normal("val", mu=0, sd=10, shape=(1, 2))
    data = theano.shared(raw_data, broadcastable=(False, True))
    pm.Deterministic("add", var=val + data)

Problem is that I can't find a way to do this via pm.Data. Would a solution here be to add a broadcastable arg to pm.Data?

[michaelosthege]

Problem is that I can't find a way to do this via pm.Data. Would a solution here be to add a broadcastable arg to pm.Data?

Maybe even forwarding unknown kwargs to shared() alltogether.
Another workaround is to pass only a vector and slice it with [:, None] or [None, :].

[paw-lu]

Maybe even forwarding unknown kwargs to shared() alltogether.

yeah that might be nice, I didn't know if shared was the primary thing pm.Data wrapped around, but if it's so, forwarding kwargs would be nice.

Good point on the None slicing!

[twiecki]

I agree we should fwd args to shared.