How can I generate my own convolution table?

135 views
Skip to first unread message

Wang Yu

unread,
Apr 8, 2014, 1:54:11 AM4/8/14
to ebl...@googlegroups.com
There is a maketable tool for eblearn. But I found no describing about how to use it. Can any one give me some tips?
For example, I want to generate a conv layer with 8 feature maps as input and output 24 feature maps. How do I generate the table with maketable tool?

Steven

unread,
Apr 10, 2014, 7:17:53 PM4/10/14
to ebl...@googlegroups.com
Hi,

I believe if you just want to fully connect the 8 maps to the 24 maps (i.e. each of the 24 maps receives input from each of the 8 maps), then you don't need maketables. Just set the following arguments in your .conf file:

conv2_table_in = 8
conv2_table_out = 24

(Note: this assumes the 24 feature maps correspond to your conv2 layer)

If you want to create your own connections between the feature maps, then make a text file of the form:

input_map_# , output_map_number_# \n

for each pair of feature maps to be connected - e.g.

0, 0;
2, 20;
3, 22;
4,1;
4, 2;
4, 3;

and so on. Then, save it as something like "MyTables.txt" Now, run maketables with the "text" argument, as follows:

maketables -text MyTables.txt

This will produce an output .mat file. For the sake of argument say this file is "table_0_0_connect_60_fanin_density_inf_MyTables.txt.mat" Next,, in your .conf, remove the table_in & table_out parameters, and instead use the parameter:

conv2_table = table_0_0_connect_60_fanin_density_inf_MyTables.txt.mat

This, I believe, will let your net use the connections you specified.

HTH.

Wang Yu

unread,
Apr 13, 2014, 9:52:50 PM4/13/14
to ebl...@googlegroups.com
Thanks a lot, Steven.
That's very helpful to me. 

在 2014年4月11日星期五UTC+8上午7时17分53秒,Steven写道:

Steven

unread,
May 6, 2014, 11:08:26 AM5/6/14
to ebl...@googlegroups.com
NOTE: Just looked at this again and realized I was writing "maketables", when the command name is really "maketable" ...Ooops...... Just wanted to correct this thread.
Reply all
Reply to author
Forward
0 new messages