Optimization A Bit Too Clever/Document it

[chewxy]

This graph:

g := NewGraph()
xV, _ := tensor.New(tensor.WithBacking([]float64{1,2,3,4})
yV, _ := tensor.New(tensor.WithBacking([]float64{1,2,3,4})
x := NodeFromAny(g, xV)
y := NodeFromAny(g, yV)

yields this:

This can be fixed of course, with this:

g := NewGraph()
xV, _ := tensor.New(tensor.WithBacking([]float64{1,2,3,4})
yV, _ := tensor.New(tensor.WithBacking([]float64{1,2,3,4})
x := NodeFromAny(g, xV, WithName("x")) // note the WithName()
y := NodeFromAny(g, yV, WithName("y")) // note the WithName()

which will yield this:

While it is in fact the correct thing to do (i.e. optimize away and leave with with only one node), this is not user friendly (this bug was encountered by a new user).

Solutions

Either one of these

1. Document such behaviours.
2. Add an exception for NodeFromAny to not use newUniqueNode.

Status

Thinking.

[chewxy]

TODO: investigate the feasibility of applying #81 to this as well