have a look at this bug https://github.com/FluxML/Flux.jl/issues/982
Am I correct in thinking that is is really serious error?
Conv gives not repeatable, quite reandom results.
I was able to repeat the bug running in Atom/Juno on Mac with Julia1.3.1+Flux0.9.
Thinking it was time to give up on julia and run!
However, running the same examples in the terminal REPL, the bug did not happen. This is using the same Julia that Juno uses, 1.3.1+Flux0.9. Also tried in Julia1.3.0+Flux0.10 terminal on Linux, no issue there.
So it seems like this is something to do with running in Juno???
An even smaller way to show the problem:
batchsize = 2 ii = randn(Float32, 5,4,3,batchsize) ll = Conv((3,3), 3=>3) maximum(ll(ii)-ll(ii))
Repeat that last line about 10 times. Usually the result is zero (correct) but sometimes it is a large number. Increased batchsize makes it happen sooner.