If you find a way to do this, please let me know. For bonus points you should follow Roubaud's principle whereby the code "speaks" about its own algorithms.
Ohh, in that case, we could train the LSTM only with code that was used to generate it. Unfortunately, I think that's just not enough data to get interesting results. Maybe we could train it on multiple different kinds of implementations of LSTMs, but that may or may not be cheating...