Setting bias to forget gates in LSTM networks and data contained in blobs

[Marshall Worth]

All, I have been looking over the following example

http://christopher5106.github.io/deep/learning/2016/06/07/recurrent-neural-net-with-Caffe.html

and I am trying to figure out this line

Set the bias to the forget gate to 5.0 as explained in the clockwork RNN paper

solver.net.params['lstm1'][2].data[15:30]=5

Is there any documentation that shows what is contained in that location, why did he choose [15:30]? How did he know to that the bias to the forget gates was there? Is the data stacked with all of the weight, bias, etc information? What order is it in?

What are in these locations

solver.net.params['lstm1'][0]
solver.net.params['lstm1'][1]
solver.net.params['lstm1'][2]

I ran across an include file that stated it could be the input and output blobs of 'lstm1'?

Thanks for your help, if you know of any relevant locations that explain some more detailed documentation of the LSTM network I would greatly appreciate it.

[Rsaeed]

Hi,

I was following the same tutorial but unable to make it work. Please let me know if it was working for you

Thanks