Output processing function REMOVECONSTANTROWS is not supported with GPU.

Hello,

I am following MATLAB tutorial on neural networks with GPU parallelization and I am not able to use my own data to start the training process, only the example data set works. The code is this:

 %[xa,ta] = house_dataset;

 [X_Train,Y_Train,T_Train] = loadData({'Data2008.mat'},  0);
 % Year to be predicted
 [X_PredYear,Y_PredYear,T_PredYear] = loadData('Data2009.mat', 0);

 xg = nndata2gpu(X_Train);
 tg = nndata2gpu(Y_Train);

 net1 = feedforwardnet(10);
 net2 = configure(net1,X_Train,Y_Train);  % Configure with MATLAB arrays
 net2 = train(net2,xg,tg);    % Execute on GPU with NNET formatted gpuArrays
 yg = net2(xg);               % Execute on GPU
 y = gpu2nndata(yg);          % Transfer array to local workspace

When I use the house data_set, everything works fine. However, when I load my data from the .mat file (X_Train and Y_Train are columns extracted from a table, so they are basically 2D arrays), I get the error "Output processing function REMOVECONSTANTROWS is not supported with GPU." in the line:

 net2 = train(net2,xg,tg);

As if this extra functionality "REMOVECONSTANTROWS" was added to the arrays for some reason. I actually tried a "dumb" approach to copy each element from X_Train and Y_Train to double arrays and pass in the5se arrays to the function nndata2gpu but the error persists. Could anyone tell what is preventing me to use my input data in this code?

Thank you.

Answer 1

[Amanjit Dulai]

The error message you are getting suggests that there are rows in your output data which are constant. Without knowing what your data is, it's hard to know for sure, but here are a few suggestions:

1) You mentioned that X_Train and Y_Train are columns from a table. Remember, the 'train' function in the Neural Network Toolbox requires that your data is formatted so that each column is an example (e.g. X_Train should be a matrix where the number of columns is the number of training samples, and the number of rows is the size of each training sample). The same applies to Y_Train. From the error, I would guess that the format for Y_Train is wrong.

2) You could switch off the output function 'removeconstantrows', although your network might encounter issues when training if some of the output data rows are constant. You can remove the output function 'removeconstantrows' with the following command:

net.output.processFcns = {'mapminmax'};

[Andre Calfa]

Sorry I actually have recently fixed my problem but forgot to reply here explaining what I did. There were two things to fix. The first one was as Amanjit said on item 1. I had to transpose both matrices when calling nndata2gpu. The removeconstantrows error was gone after that. However, I came across another error when calling the train function: cuda illegal memory access. This was because my matrices elements were not with double type, so after I casted them to double everything started working as it should.

Thank you for the reply!

[Lindsay Wong]

Thanks. I meet the same error and found that the output column is constant. Problem solved.

[Yunyu Hu]

I also had this problem and remove the constant input columns , then problem solved

Answer 2

[Remington Reid]

I ran into the same issue but required a slightly different solution. I was training a 2 layer feedforward net that produces a single output. The input training data dimensions were 6 x N and the output training dimensions were 1 x N. The output data was definitely not constant. I tried filtering repeated values in the output to be safe with no luck. Both input and output were classed as 'double'. Disabling the REMOVECONSTANTROWS option worked but I had to disable it both on the input and on the output.

net.input.processFcns = {'mapminmax'};

net.output.processFcns = {'mapminmax'};

In my case network training and performance did not seem to suffer, but since there were no constant rows I suppose that isn't to surprising. Thank you for the instructions to disable the option!