Styjun

I'm trying to build my own regression network using Matlab. Although what I've got so far looks a bit pointless, I do want to expand it later into a slightly unusual network so I am doing it myself rather than getting something off the shelf.

I've written the following code:

% splitinto dev, val and test sets
[train_idxs,val_idxs,test_idxs] = dividerand(size(X,2));

training_X = X( : , train_idxs );
training_Y = Y( : , train_idxs );

val_X = X( : , val_idxs );
val_Y = Y( : , val_idxs );

test_X = X( : , test_idxs );
test_Y = Y( : , test_idxs );

input_count = size( training_X , 1 );
output_count = size( training_Y , 1 );

layers = [ ...
 sequenceInputLayer(input_count)
 fullyConnectedLayer(16)
 reluLayer
 fullyConnectedLayer(8)
 reluLayer
 fullyConnectedLayer(4)
 reluLayer
 fullyConnectedLayer(output_count)
 regressionLayer
 ];

options = trainingOptions('sgdm', ...
 'MaxEpochs',8, ...
 'MiniBatchSize', 1000 , ...
 'ValidationData',val_X,val_Y, ...
 'ValidationFrequency',30, ...
 'ValidationPatience',5, ...
 'Verbose',true, ...
 'Plots','training-progress');

size( training_X )
size( training_Y )
size( val_X )
size( val_Y )

layers

net = trainNetwork(training_X,training_Y,layers,options);

view( net );

pred_Y = predict(net,test_X)

I can't share what X and Y actually are, but the input X is a 3xn double array and the output is Y is a 2xn array which originally came from a Matlab table.

Here is the output:

ans =
 3 547993

ans =
 2 547993

ans =
 3 117427

ans =
 2 117427

layers = 
 9x1 Layer array with layers:

 1 '' Sequence Input Sequence input with 3 dimensions
 2 '' Fully Connected 16 fully connected layer
 3 '' ReLU ReLU
 4 '' Fully Connected 8 fully connected layer
 5 '' ReLU ReLU
 6 '' Fully Connected 4 fully connected layer
 7 '' ReLU ReLU
 8 '' Fully Connected 2 fully connected layer
 9 '' Regression Output mean-squared-error


Training on single CPU.
|======================================================================================================================|
| Epoch | Iteration | Time Elapsed | Mini-batch | Validation | Mini-batch | Validation | Base Learning |
| | | (hh:mm:ss) | RMSE | RMSE | Loss | Loss | Rate |
|======================================================================================================================|
| 1 | 1 | 00:00:02 | 0.88 | 4509.94 | 0.3911 | 1.0170e+07 | 0.0100 |
| 8 | 8 | 00:00:04 | NaN | NaN | NaN | NaN | 0.0100 |
|======================================================================================================================|
Error using view (line 73)
Invalid input arguments

Error in layer (line 85)
view( net );

Clearly something pathological is happening, since the training is almost instantaneous and I can't view the resulting network. Can anyone advise me what I am doing wrong ? Or perhaps give some debugging tips ?

Thanks,
Adam.

There are two problems here: the first one is, that the call view(net) fails. The reason is that view() function only works for network objects. The network class and corresponding methods have been a part of the Neural Network toolbox for years, and are intended for shallow, "classical" neural networks.

Your trained net however is a SeriesNetwork, which is a much newer class, used for Deep Learning. You can not mix functions for network and SeriesNetwork, so consequently view() doesn't work here.

There is a similar function called analyzeNetwork() to graphically view and analyze a deep neural network in the SeriesNetwork format:

analyzeNetwork(net)

The second problem is that the RMSE and the loss are NaN (not-a-number) after the training. The reason for this is difficult to diagnose without your actual data.

One possible reason: You have data containing NaN in the inputs or outputs. You can check this with the isnan() function:

any(isnan(training_X(:)))

If this is not the case, then you could e.g. check the weight and bias initialization or the learning rate.

You also should consider look at the 'InitialLearnRate' parameter in trainingOptions. By default it is 1e-3, it is sometimes necessary to choose a smaller value to avoid the optimization blowing up, like yours currently is.

Another option to look at with regression problems is the 'GradientThreshold' option in trainingOptions. Setting this will use gradient clipping to prevent gradients from exploding during training. This can also be beneficial/necessary in making RMSE optimization behave well.