Funny, inconsistent behavior of depthwiseconv

[vargonis]

I'm getting inconsistent results among different calls for a depthwise convolution, with frequent NaNs. It can be reproduced, for instance, with:

using Flux
c = DepthwiseConv((3,3), 3=>6, relu; pad=1)
x = rand(10,10,3,1);
all(c(x) .== c(x))

NNlib is 0.6.0. (Flux is 0.10 and Julia 1.3, but that seems irrelevant).