I use the nprtool to train a network on a 1 input -> 1 output class problem (actually there are many more input features, but for boiling down the issue I reduced it to one).
Using nprtool I can train a network even with 120k samples. No problems here.
When using the unmodified generated scripts (both simple and advanced script) things don't work. Meaning:
- >10k samples, Matlab crashes with a fatal error (Segmentation Violation in MEX file)
- >2k samples, performance/gradient are Infinite and the network does not train
- >1k samples, the network results in 90% error rate
With NPRtool I can get down to 16% error rate with the single input, so it's definitely possible to train from this data.
I have also tested saving the network from the NPRtool UI to a variable and then run train on this network, same result as above.
What are the differences between the NPRTool UI and the generated scripts? Has anyone ever encountered this problem or are there any known workarounds?
Thanks!
- Dennis
PS: Matlab version 2013a on OSX 10.8.4
1 Comment
Direct link to this comment:
http://www.mathworks.com/matlabcentral/answers/83668#comment_162497
How many hidden nodes?
How many random initial weight trials?