Erratic Dilation behavior

[staticfloat]

Uh-oh, I should have tested this more thoroughly looks like. :/

# 2-length convolution with 1 channel in and 1 channel out
w = reshape(Float64[1, 1], (2, 1, 1))

# 128-length input vector with 1 channel and 1 batch.
x = reshape(Float64[1:128...], (128, 1, 1))

# Let's calculate the same thing a couple different times
ys = []
for idx in 1:10
    push!(ys, NNlib.conv(x, w; dilation = 2))
end

# Do some of these randomly have a bunch of NaN's in them?  Why yes they do:
[any(isnan.(ys[idx])) for idx in 1:10]

Running the above results in:

10-element Array{Bool,1}:
 false
  true
  true
 false
 false
 false
 false
 false
  true
  true

The fact that this is erratic makes me think the BLAS routine is reading from uninitialized/clobbered memory.

[staticfloat]

@tejank10 sorry to so quickly have to come back to you about this.

[tejank10]

@staticfloat, no problem. I changed the code on line 178 of impl/conv.jl to size(w) instead of dilation_dims(w, dilation) and it didn't give any NaN for the above inputs (I ran the loop for about 10k times). But what surprises me is that how could it (the code with dilation_dims) give different output for the same input, and can also gives the correct output in some cases?

[staticfloat]

Fixed by #94