Multi-view-neural-acoustic-words-embeddings

Prerequisites

Python --- 2.7.10, Tensorflow --- 0.9.0rc0, numpy, scipy

How to use

To help organizing the code, here is my working directory

/Multi-view_embedding
    /Data
    /Outputs
    /Saved_Models
    /Scripts

To train a model with 2-layer bidirectional LSTMs on both views:

python 2biLSTM.py <parameters>
available parameters include:
  -lr: learning rate, 0.001 by default (for LSTMs with more than 1 layer, 0.0001 is recommended)
  -obj: objectives, 0(obj0) by default, 02 = averaged sum of obj0 and obj2, must be a subset of 0 1 2 3
  -hs: hidden size, 512 by default
  -m: margin, 0.5 by default
  -kp: keep probability of dropout, 0.6 by default
  -bs: batch size, 20 by default
  -epo: epochs, 20 by default
  -inits: initialization scale of LSTM cells, 0.05 by default
to train our best model:
python 2biLSTM.py -lr 0.0001 -obj 02 -m 0.4 -epo 1000