Draft: py-pynn: add patch for numpy@1.24:

changed title from Draft: py-pynn: add patch for numpy@1:24: to Draft: py-pynn: add patch for numpy@1.24:

added 5 commits

14bc208f...d86c095c - 4 commits from branch spack_v0.21.1
33bc2aef - py-pynn: add patch for numpy@1:24:

During our last Spack update, that included a change from py-numpy@1.23.5 to py-numpy@1.24.4 in the env, one of the pyNN tests for the brian2 backend failed:

=================================== FAILURES ===================================
_______________________ TestProjection.test_partitioning _______________________

self = <unittests.test_brian.TestProjection testMethod=test_partitioning>

    def test_partitioning(self):
        p1 = sim.Population(5, sim.IF_cond_exp())
        p2 = sim.Population(7, sim.IF_cond_exp())
        a = p1 + p2[1:4]
        # [0 2 3 4 5][x 1 2 3 x x x]
        prj = sim.Projection(a, a, MockConnector(), synapse_type=self.syn)
        presynaptic_indices = np.array([0, 3, 4, 6, 7])
>       partitions = prj._partition(presynaptic_indices)

test/unittests/test_brian.py:38: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = Projection("assembly67→assembly67"), indices = array([0, 3, 4, 6, 7])

    def _partition(self, indices):
        """
        partition indices, in case of Assemblies
        """
        if isinstance(self.pre, common.Assembly):
            boundaries = np.cumsum([0] + [p.size for p in self.pre.populations])
            assert indices.max() < boundaries[-1]
>           partitions = np.split(indices, np.searchsorted(
                indices, boundaries[1:-1])) - boundaries[:-1]
E           ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (2,) + inhomogeneous part.

/mnt/spack_v0.21.1/spack/opt/spack/linux-ubuntu20.04-x86_64/gcc-10.3.0/py-pynn-0.11.0-63gwwjhiu5vr2d7gz7kbzwknl4basfsd/lib/python3.8/site-packages/pyNN/brian2/projections.py:172: ValueError

From what I can see, this is because starting from py-numpy@1.24, Ragged array creation will now always raise a ValueError unless dtype=object is passed. (see here)

In py-numpy@1.23.5 there was already a warning:

VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when creating the ndarray

With this small patch, all tests are passing. Do you think it's enough? if something else is needed please let me know

I think this patch is fine, but on reflection the original code is anyway not very easy to understand. I've fixed this a slightly different way, and released PyNN 0.12.2

even better, thanks!

closed

Draft: py-pynn: add patch for numpy@1.24:

Merge request reports

Activity