Since we can pickle graphs, we should do some fuzz testing.
Construct random graphs, apply them a bunch of times, and see that we get the same results if we pcikle and unpickle in the middle of training as if we just keep training.