Operators (@ and ^) and side effects

[amilsted]

There's a nagging worry in my head that our overloading of @ and ^ might cause people problems, since these operations involve not only Nodes or Edges - they also modify the parent TensorNetwork. I can't think of another case where operators work like this.

Perhaps we should make it necessary to use these within a context, to make this clearer? For example:

net = TensorNetwork()
n1 = net.add_node(tensor1)
n2 = net.add_node(tensor2)

with net:
  n1[1] ^ n2[0]
  res = n1 @ n2

print(res.get_tensor())

Apart from providing __enter__ and __exit__ methods for TensorNetwork, AFAICT this would require defining a global variable to hold the "current network", but I don't see an obvious problem with that.

[chaserileyroberts]

I can't see how this is helpful. Edge and Node already point to their parent network, and we already throw an error if the networks are mismatched.

[chaserileyroberts]

require defining a global variable

No! Bad ash.

Tensorflow did something similar with their tf.Graph objevt, and they are gutting that feature in TF 2.0 because it confused the hell out of users and caused code to break in really subtle ways.

[amilsted]

I can't see how this is helpful. Edge and Node already point to their parent network, and we already throw an error if the networks are mismatched.

Indeed, it would just be a different syntax for the same thing. The idea was just to clarify the context-dependency (on the parent network).

Tensorflow did something similar with their tf.Graph objevt, and they are gutting that feature in TF 2.0 because it confused the hell out of users and caused code to break in really subtle ways.

Well, my thought was to have this global only set (and unset) by the TensorNetwork context manager. If an expression with @ or ^ were used outside such a context, this would still be an error. Probably one would want to make nesting such contexts an error also.

Anyway, I recognize I am probably just being a little paranoid here, so if nobody else agrees there is a problem, I promise I will try to forget about it :)

[chaserileyroberts]

Let's revisit it in 2 weeks. If you can find a code sample that doesn't give a clear error message already and can be fixed with the context manager, then we'll implement it. Sound good?