DotLayer argument simplifications

[albertz]

This issue is just to collect recent orthogonal but related proposals.

DotLayer specific

• DotLayer defaults should not use integers, change defaults #627
• DotLayer, use single reduce argument #636
• DotLayer could infer var1/var2 automatically #738
• DotLayer: maybe use 1D conv CuDNN kernel #740
• Generalize DotLayer to allow sparse inputs #741

Related on dim tags

• Specify dim tags for layers that create new axes #597
• Enforce dim tags to be unique in tensor #632
• Definition of equality of dim tags #634
• Specify expected output shape for verification #706

Other related

• Implicit broadcasting for all inputs might be bad #691
• GatherLayer
• LinearLayer

[albertz]

Note, in principle, DotLayer would not need any options anymore, as we can infer all the information about what axes to reduce and what axes are shared. However, I think it is good to have to specify the reduce axes explicitly. Maybe we could however omit this when the user specifies out_shape.

[Zettelkasten]

The reduce arg is always needed to distinguish it from the common axes, right?
(Okay, you could infer it from out_shape as it is not present there but the common axes are.)

[albertz]

The reduce arg is always needed to distinguish it from the common axes, right? (Okay, you could infer it from out_shape as it is not present there but the common axes are.)

Ah yes you are right.

[albertz]

I think all the main features for DotLayer are implemented now, so this can be closed.