InaccessibleTensorError when running RGCN with Dropout

[SimonTopp]

When training the RGCN on TG GPU with dropout, you get the error:

InaccessibleTensorError: The tensor 'Tensor("lstm_cell/cond/Identity:0", shape=(None, 7), dtype=float32)' cannot be accessed here: it is defined in another function or code block. Use return values, explicit Python locals or TensorFlow collections to access it. Defined in: FuncGraph(name=call, id=46914785443568); accessed from: FuncGraph(name=keras_graph, id=46914397049016).

You don't get this error when setting dropout to zero. I think I've fixed it by updating the dropout mask in the forward pass as described here, but wanted to double check if this is new behavior since the training routine update PR #141. @jsadler2, have you run into this in the past/did you ever experiment with dropout? The defaults are set to 0, so I'm wondering if this has just never come up before.

[jsadler2]

I don't remember ever running into this. I also don't think that I ever tried dropout in the RGCN, now that I think of it. I might have only tried that in the plain LSTMs that I was using for the multi-task experiments.

[jzwart]

I also don't think we've run the RGCN with dropout >0. We've tested RGCN with state updating but I think just left dropout as default

[SimonTopp]

Fixed in #154.