week 27.03 02.04.2017

• half functional Lasagne TIMIT
• tensorflow SR: Wavenet -> example how to use sequence length?
• inputs with variable length input sequence
• audio input to LSTM: format
• lots of stuff improved on last week's work (preprocessing etc) -> see there for explanations

Connecting Data and network structure

• wrong input shape? not all input files same size (some WAV longer than others) -> when stored in numpy array, data.shape = (4512,) instead of (4512,300,26)

  • need to add a Sequence_length to each input?
  • need to pad with zeros and add mask somehow? How to train properly?
    -> see http://colinraffel.com/talks/hammer2015recurrent.pdf. CODE
    -> Lasagne basics

• You can get the output shape of a Lasagne layer by calling print(l_dense.output_shape)

Combining Audio and Image

• Multiple inputs to Lasagne
• Multiple networks combined