Train RNN model

[obitech]

• Move from plain convolutional to recurrent convolutional network
  • One prediction per 10s audio slice
• No more majority vote on received output classes
• First prototype working
• Also see https://github.com/gojibjib/jibjib-model/issues/6