week 03.04 09.04.2017

Network building:

• making LSTM networks
• making bidirectional LSTM networks
• making deep (bidirectional) LSTM networks
• also with dense layer after it

Network functions:

• to train in batches, all inputs in a batch need to be the same size -> we need padding and masks
  see general_tools for padding and mask generation. This is used in iterate_minibatch to fix the input data so everything is the same length.
• also get valid frames for later

Network evaluation, training and testing:

• automatic dataset generation, choose TIMIT/TCDTIMIT/combined. Also possible to just choose a wav directory.
• just specify data_dir and store_dir
• automatic network training and generation. Options:
  - uni or bidirectional?
  - number of LSTM layers, nb of units.
  - add dense layers afterward?
  - store log files, network parameters, training info etc in specified directories.
• automatic network evaluation on specified data. See 'evaluateManyModels.py'. Attempt at evaluation of wavs without targets, but not finished (can finish later if time left, it's not very hard).

I compared all kinds of different network models on all the datasets (TCD, TIMIT, combined). See Verslag/modelPerformances.jpg