Menu

nnet2 dnn training with a low rank layer

Help
Yan Yin
2015-07-10
2015-07-10
  • Yan Yin

    Yan Yin - 2015-07-10

    Hi All,

    As I see from existing Kaldi recipe, we usually start with network with one hidden layer and hidden layers are gradually inserted into network during network training. Now I want to train a network with a smaller-size low rank layer between top hidden and output layer. In nnet.config the input dimension of output layer need to be same as output dim of hidden layer, and then later during training when the low rank layer is inserted at the end, will this ends up with issue (output dim of low rank layer will not match input dim of output layer (defined in nnet.config)?
    If this is the case, I am wondering how this can be done properly?

    thanks,
    Yan

     

    Last edit: Yan Yin 2015-07-10
  • Yan Yin

    Yan Yin - 2015-07-10

    an extreme case is, If I want to build a shrinking network (hidden layer size keep decreasing), how to achieve this?

     

    Last edit: Yan Yin 2015-07-10
    • Daniel Povey

      Daniel Povey - 2015-07-10

      One easy way to do this is simply to set up the neural network with
      two affine components after the final hidden nonlinearity and before
      the output softmax, with a smaller dimension in between. The program
      nnet-am-limit-rank-final will do this for you by performing an SVD on
      the last layer, if you are using the nnet2 setup (assuming that
      program still works..). I never ended up using this in the standard
      recipes because it didn't improve the WER results and at the time I
      wasn't so interested in pure speed improvements.

      Dan

      On Fri, Jul 10, 2015 at 9:46 AM, Yan Yin riyijiye1976@users.sf.net wrote:

      I guess what I can do is, initiliaze network with input, low rank, and
      output layers, then keep inserting hidden layer before low rank layer.


      nnet2 dnn training with a low rank layer


      Sent from sourceforge.net because you indicated interest in
      https://sourceforge.net/p/kaldi/discussion/1355348/

      To unsubscribe from further messages, please visit
      https://sourceforge.net/auth/subscriptions/