Shawn Tan

We wanted to run it with cuda by default, so we flipped the meaning of that flag. You can also use --cuda to run _without_ cuda. Sorry for the confusion.

From the paper: > As the master gates only focus on coarse-grained control, modeling them with the same dimensions as the hidden states is computationally expensive and unnecessary. In practice,...

I don't. The current architecture for the controller is just the simplest guess (slapping everything on the hidden layer). You can train a model that works well, it's just I...

Same here.

Yes and I've tried RMSProp. I'm looking for more specifics, like initialisation parameters, gradient clipping, etc.

The LSTM has the same idea of 'forgetting' and then adding the new input. That was what was meant by that line in the paper. Just to be clear: I...

Are you suggesting separating files on the ECG classification into different directories? The labels are applied for a particular window, and they don't always line up with a file.

Yeah, it was one of the things I was working on before life got in the way. Heh. I might get back round to fiddling with it when I have...

I'm not sure what you mean. Do you mean in terms of computational cost? You can take a look at this issue https://github.com/shawntan/theano-ctc/issues/1, the asker did a small benchmark.

You may need to obtain the PennTreebank parse trees before running test_phrase_grammar.py