Johan

Thanks, I updated the failing test. Since reduce without axis is implemented with multiple nodes I think that `@tf Ysum1 = reduce_sum(Y)` should create a scope that contains these nodes....

Ah, just because there is a node named `reduce_sum/rank` doesn't prevent us from creating a node named `reduce_sum`. I thought it did. However before this PR the node was named...

Here are some graphs. tensorflow.py v1.0 and v1.8 were near identical. TensorFlow.jl known rank:  tensorflow.py v1.8 known rank:  TensorFlow.jl unknown rank:  tensorflow.py v1.8 unknown...

Code for the above: julia ``` using TensorFlow; tf = TensorFlow v = tf.placeholder(Float32, shape=[2,3,5]) myop = tf.with_op_name("myop") do tf.reduce_sum(v) + tf.reduce_sum(v) + tf.reduce_sum(v, name="s") end myop_dim = tf.with_op_name("myop3") do...

This is what the graph from TensorFlow.jl looks like before this PR (both the case with known rank and the case with unknown rank produces the same graph): 

If I'm not mistaken `MyArrayType(a::AbstractArray)` takes priority over `convert(::Type{MyArrayType}, a::CategoricalArray)` since julia searches for a matching constructor before the fallback to `convert`. Since `Array` has a method with this signature...

Thanks. I guess Ctrl+Enter does "Julia Client: Run Block" on Windows and Linux? (I only know about OS X where it's Cmd+Enter). I was referring to the behaviour of "Julia...

Nope, not by me. But it seems @adrianbroher copied the block for parsing `--texture` arguments when he created the block for parsing `--rendermode` arguments here: https://github.com/rougier/freetype-gl/commit/75a05c4db4ea6a6db70f30d7a81d2a5626689ebf