Frozen Synthetic Dataset

[hnaik]

Based on the code in gengraph and synthetic_structsim modules, I gather that when running training with the synthetic datasets, input graphs are generated on the fly. Based on the randomness routines present in methods such as synthetic_structsim.build_graph() it is evident that every invocation of the method will likely generate a completely different instance of the input graph.

1. Wouldn't it be more prudent to run a few tens of these iterations, save representations to files and then use those files for all experiments? Yes, there will be some loss in surprise and variety, but at least the experiments would then be repeatable/reproducible, wouldn't you agree?
2. Are there other downsides to this frozen dataset approach that I might have missed.