Refactor

[AMairesse]

Hi,

Here is a bit of re-factorization :

• config file management sent into hyperparameters.py
• training sent into the model

I have also worked on the file size problem (cutting) and optimization for small configs. I came up with two parameters set : one is recommended because it can use most of the dataset, one is a optimal config for using 8 Gb of memory without swapping.

Now train.py and stt.py could be merged into one file, the resulting file could manage training with a '--train' parameter. I was thinking to rename stt.py to "rnn-speech" (project name) or keep stt.py (meaning "speech to text", shorter in everyday use :-) ). Do you have any preference ? Do you prefer to keep two separate launchers ?

Antoine.

[domerin0]

Thanks again for your excellent work Antoine. I like the idea for using a flag to train or sample from one file, reduces code duplication and is probably good for future maintainability. Choose whatever name you like most, considering all the work you've just recently put into it.

On another note, would you like to be added as a collaborator?

[AMairesse]

Hi, why not, I'm still new to git so my workflow is sometimes messy with the pr thing :-)