use trainnetwork for normal regression

Hi,

I have a dataset of 63 inputs and 1 output for a regression problem. Total sample 39686.

X: 63x39686

Y: 1x39686

I can easily use "net=fitnet(...)" and "train(net X,Y)" to train the model.

But I want to try the trainnetwork function. After configuring the layers like this:

layers = [

sequenceInputLayer(size(X,1),"Name","sequence_In","Normalization","rescale-zero-one")

fullyConnectedLayer(20,"Name","fc_1")

fullyConnectedLayer(20,"Name","fc_2")

regressionLayer("Name","regressionoutput")];

and options:

options = trainingOptions('sgdm', ...

'InitialLearnRate',0.001, ...

'Verbose',false, ...

'Plots','training-progress');

Then I train the model:

net = trainNetwork(X,Y,layers,options);

But it always shows :

To RESHAPE the number of elements must not change.

Error in NN_training_deep (line 33)

net = trainNetwork(X_,Y',layers,options);

Does anyone know how to solve this problem?

Answer 1

[Srivardhan Gadila]

The outputSize argument for the fullyConnectedLayer before the regressionLayer must be 1 as the number of ouputs for your regression problem is 1.

layers = [
     sequenceInputLayer(size(X,1),"Name","sequence_In","Normalization","rescale-zero-one") 
    fullyConnectedLayer(20,"Name","fc_1")
    fullyConnectedLayer(20,"Name","fc_2")
    fullyConnectedLayer(1,"Name","fc_3")
    regressionLayer("Name","regressionoutput")];

[Yunyu Hu]

Hi Srivardhan,

Thanks! It works. From the documenttation, the regression layer calculates only loss function for regression problem. So I can say it is not a real layer for the neural network, right?

Another question, when training the above problem, the iteration stopped after second iteration:

|======================================================================================================================|

| Epoch | Iteration | Time Elapsed | Mini-batch | Validation | Mini-batch | Validation | Base Learning |

| | | (hh:mm:ss) | RMSE | RMSE | Loss | Loss | Rate |

|======================================================================================================================|

| 1 | 1 | 00:00:00 | 488.06 | Inf | 119101.8125 | Inf | 0.0100 |

| 30 | 30 | 00:00:02 | NaN | NaN | NaN | NaN | 0.0100 |

|======================================================================================================================|

Do you know perhaps the reason?

BR

[Srivardhan Gadila]

The loss seems very high, try normalizing the data.

Also refer to the following links:

1. Improve Training Accuracy
2. Deep Learning Tips and Tricks
3. How to avoid NaN in the Mini-batch-loss from traning convolutional neural network?

[wahed fazeli]

Hi @Srivardhan Gadila can you solve my problem in this subject?

I have a dataset of 9 inputs and 1 output for training data. Total sample 488

B: 9x488

F: 1x488

I want to train my data using deep learning but when i want to do that .matlab r2018b give me nothing.

these are codes of matlab.

Firstly i have used this code but it gave me some errors.

layers = [
sequenceInputLayer(size(B,1),"Name","sequence_In","Normalization","rescale-zero-one")
fullyConnectedLayer(20,"Name","fc_1")
fullyConnectedLayer(20,"Name","fc_2")
fullyConnectedLayer(1,"Name","fc_3")
regressionLayer("Name","regressionoutput")];
Error using sequenceInputLayer>iParseInputArguments (line 41)
'Normalization' is not a recognized parameter. For a list of valid name-value pair arguments, see the documentation
for this function.
Error in sequenceInputLayer (line 26)
inputArguments = iParseInputArguments(varargin{:});

so i have changed the code and write this code.so it worked at first.

 layers = [
sequenceInputLayer(size(b,1),"Name","sequence_In")
fullyConnectedLayer(20,"Name","fc_1")
fullyConnectedLayer(20,"Name","fc_2")
fullyConnectedLayer(1,"Name","fc_3")
regressionLayer("Name","regressionoutput")];

and write this code for options.

option=trainingOptions('sgdm','MaxEpochs',20,'InitialLearnRate',0.001,'Verbose',false,'Plots','training-progress');
net=trainNetwork(B,F,layers,option);

when i run this code the matlab give me nothing in result.the version of matlab i have used is R2018b

validation RMSE: N/A and other parameters this is the snap shot of results.I dont know what is problem.can anyone help me fix this error .thanks.