darktex / char-rnn Goto Github PK

The "size" parameter for the RNN is not the same for the two models (Torch and JS). In Torch, each Linear layer has its own bias, so there are as many bias vectors as there are weight matrices (so 1 per gate per layer for both x and h). In JS, each gate has only 1 bias. We might either change the JS code or we might try to "hack" the format a little bit by adding 1's to the input vectors.

The JS code was made to enable training in a browser in reasonable time. This meant that Karpathy added an extra layer before the LSTMs that mapped the vocabulary (of size ~50-100) to a much smaller embedding (the example uses size 5). We won't care about training, so given that we will only do forward propagation we should be able to afford skipping this embedding step, but the JS code expects to find a matrix Wil that contains the weights for the embedding. I have therefore hacked it a bit by setting an embedding size to be the same as the vocabulary size, and setting Wil to be an identity matrix. This needs testing.

It's simple to convert a Torch table into JSON using json.encode(), but that does not preserve the order of the elements. LUA tables are inherently unordered when their key is not an integer. If you do use an integer as key that's cool, but then you have to be more careful when building the JSON file. This may or may not have an impact on correctness and should be tested.