Is there a model conversion tool?


(Netxiao) #1

I have a trained model(layers = 2, rnn_size = 512), is there a tools convert it to another network topology (layers = 4, rnn_size = 1024)?

thanks!


(Guillaume Klein) #2

There is no such tool currently.

However, @Dakun experimented on this subject. Maybe he could share his findings and how he proceeded.


#3

It’s not difficult to expand/convert from e.g. layers =2, rnn_size = 512 to layers = 4, rnn_size = 1024.

  1. you may output all the parameters in each encoder/decoder for different configurations.
  2. load the previous model
  3. create a new encoder/decoder
  4. copy or expand corresponding parameters for each matrix.
  5. train with the newly generated encoder/decoder

BTW, personally, I don’t think it is a good way to directly jump from 2512 to 41024.
It’s really big for 41024 network. Maybe you can start from 2512 to 4*512 for instance.
That will be much easier.


(Netxiao) #4

Thanks for your great answer!


(Terence Lewis) #5

Great post - I was thinking about this problem this morning :slight_smile: