BLAS error at backward of F.sum of L.Convolution2d

The code below terminates unsuccessfully with the message:

BLAS error: Parameter incX passed to cblas_sgemv was 0, which is invalid.

With h = h[:,:,:,:], it causes no errors. This behavior seems to be a bug.

import chainer
import chainer.functions as F
import chainer.links as L
import numpy as np

conv = L.Convolution2D(3, 1, 3)

conv.cleargrads()

x = np.random.rand(300).astype(np.float32).reshape(1,3,10,10)
h = conv(x)
print(h.data)
# h = h[:,:,:,:]
y = F.sum(h, axis=(2,3))
print(y.data)
y.backward()

log.txt

[muupan]

FYI:

• F.mean is also affected by this issue.
• You can avoid this issue by adding h = h[:] before F.sum(h) or F.mean(h).
• We might need to wait for https://github.com/numpy/numpy/issues/9165 to be solved.

[kmaehashi]

Ah, thanks, I overlooked the link. The corresponding PR is #9177.