Is it valid to enter a Non-Image input in Convolutional Neural Network toolbox?
9 views (last 30 days)
Show older comments
Javier Pinzón
on 27 Apr 2017
Commented: Javier Pinzón
on 22 Feb 2018
Hello All, I was wondering wether it is possible to enter an input that is not an image in a CNN using the toolbox (2016b or later), i.e., I have a [:,:,3] matrix containing data of a signal through the time (every 20 ms), however, this data contains negative numbers, some numbers that are bigger than 255, and they are "double". Also, they are stored in a .mat file... so, I used the "imageDatastore" function to store the data:
imds = imageDatastore(location, 'FileExtensions', '.mat', 'IncludeSubfolders',1, ...
'LabelSource','foldernames');
but changing the "readFcn" for the next code:
function I = readFcn1(filename)
% Load data and get matrices from the structure
I = load(filename);
I = I.signaldata;
I send it to the "trainNetwork" function and "works", however, i do not know if internally it changes data of negative numbers, changes form double to uint16, or it trains the network with the original ones. Or there is another way to use a Non-image input for a CNN.
Thanks in advance ;)
3 Comments
SRUTHY SKARIA
on 21 Feb 2018
Hi, I have a similar problem. I tried use the code but my code doesn't read the function, so I am unable to get the cell array from the mat file. Also the function should end with an 'end' rigt? Also with function I = readFcn1(filename), by filename it meant the entire path with the mat file right? Would you be able to give me advice one this. Thank you
Accepted Answer
Javier Pinzón
on 13 Jul 2017
2 Comments
sruthy shankar
on 6 Feb 2018
hello sir,Iam having a similar problem ..can u help me to solve it? iam working on CNN,but my input datas are not images but some numeric datas which are stored in an excel sheet,i have to split the excel sheet data to two sets one for training and other for testing.and also in each excel sheet each row is my input to the network,how can i provide such a data to a CNN?
More Answers (1)
Zachary David
on 14 Sep 2017
Hi Javier,
I've been experimenting with a similar case. A seemingly decent solution is to pre-process your values by mapping them to a specific range. You can use the procedures described in Matlab's mapminmax to do this for you.
Or (preferred) if you want to do it manually you can do:
yMapped = (ymax-ymin).*(x-xmin)./(xmax-xmin) + ymin;
where ymin and ymax are your target range, and xmin and xmax are the minimum and maximum values of your data.
The benefit of using this technique is that it preserves the shape of the distributions for each of your variables.
Enjoy
0 Comments
See Also
Categories
Find more on Deep Learning for Image Processing in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!