Using saved neural network on different size input
显示 更早的评论
Hello,
I am new to MATLAB Neural Networks and have successfully trained and saved a network using input with 100 entries and associated targets. I would like to use this trained network with a smaller number of inputs, say 15 inputs - when I try I get the following error: " Input data sizes do not match net.inputs{1}.size"
I am pretty sure this is because the network was trained on different number of elements - is there a way to use my previously trained network on the new input size, or will I have to retrain for the new size?
Thank you in advance for any help/hints.
采纳的回答
Greg Heath
2018-5-15
编辑：Greg Heath
2018-5-15
2 个投票
The previous answer is not correct.
The topic of removing inputs that are correlated with other inputs is not trivial.
Unfortunately, ignoring features after training to just eliminate their influence is not effectively accomplished by just zeroing out the respective inputs.
If the input to be zeroed out is not orthogonal to the other inputs, the optimal approach is to COMPLETELY retrain the net from scratch.
However, very often the suboptimal choice of continued training after the removal is used.
I posted one or more examples using the BIOID dataset. Check BOTH tHe NEWSGROUP and ANSWERS.
Hope this helps.
Thank you for formally accepting my answer
Greg
更多回答（1 个）
Ameer Hamza
2018-5-14
编辑：Ameer Hamza
2018-5-14
1 个投票
A particular network architecture corresponds to a particular input and output mapping. The input to neural network represents the features which you are using for classification. Decreasing the number of input does not make much sense. Are trying to ignore some features? In that case, convert 15 element vector to 100 element vector by adding 85 zeros. The feature vector must remain of the same size, as used at the time of training.
类别
在 帮助中心 和 File Exchange 中查找有关 Deep Learning Toolbox 的更多信息
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!
Translated by 
