Nice to see a preexisting model ship with the code. Always surprised me how difficult it is to find and compare trained RNNs, given how tiny a typical char-RNN

Nice to see a preexisting model ship with the code. Always surprised me how difficult it is to find and compare trained RNNs, given how tiny a typical char-RNN model is, and how long they usually take to train. I hope in the near future it becomes viable to load different models on demand — models perhaps trained on the same corpus but differing in their choices of parameters, so that it might be possible to glimpse the qualitative impact of, say, window size on the model's output. (I'm especially curious what it looks like at n>1000, does it degrade? Pay diminishing returns? Do longer window sizes help RNNs to stay on topic for longer than a sentence or two? I'm really interested in the answer, but CUDA poops out on my machine generally around the n=256 threshold. So I'm hoping more projects like yours take off :)