"Invalid training data" for multiclass classification

2 views (last 30 days)
I am trying to use classification, to classify three classes, using a simple fully connected NN.
My training data consists of 69979 examples with 12 input data 'X* (real numbers), and I wish to predict the classes '0','1', and '2', stored in 'Y'
>> whos X Y
Name Size Bytes Class Attributes
X 69979x12 6717984 double
Y 69979x1 70299 categorical
>> summary(Y)
0 27469
1 28078
2 14432
My network is like this,
layers =
9×1 Layer array with layers:
1 'fcIN' Feature Input 12 features
2 '' ReLU ReLU
3 'hl1' Fully Connected 40 fully connected layer
4 '' ReLU ReLU
5 'hl2' Fully Connected 40 fully connected layer
6 '' ReLU ReLU
7 'fcOUT' Fully Connected 3 fully connected layer
8 'SoftMax' Softmax softmax
9 'class_output' Classification Output crossentropyex
and training works fine using
[ml.net,ml.info] = trainNetwork(X,D,layers,options);
But what I really would like to do is multiclass classification. I want to estimayte the probability of classes '0' '1' and '2' for 125 parameters.
My 'raw' in and out data have sizes
>> whos X Y
Name Size Bytes Class Attributes
X 69979x12 6717984 double
Y 69979x125 8747695 categorical
So, for each 'x' in the input I have 12 data points, and I know the class of 125 (and not just 1) parameters.
I use the same network as above, but this time with 3*125 nodes in the last fully connected layer, as I have 3 classes for each of the 125 parameters.
But, when I try to train the network I get the error "Invalid training data":
>> [ml.net,ml.info] = trainNetwork(X,Y,ml.layers,ml.options)
Error using trainNetwork (line 184)
Invalid training data. For classification tasks, responses must be a vector of categorical responses. For regression
tasks, responses must be a vector, a matrix, or a 4-D array of numeric responses which must not contain NaNs.
IN order to get a "vector of categorical responses" I try to use
Y = num2cell(categorical(Y),2);
whos Y
Name Size Bytes Class Attributes
Y 69979x1 38418471 cell
But I get exactly the same error message.
So how to I format the Y data so I can train this type of data to predict the class of the 125 parameters for a set of input data (of size 1x12)?
Is it possible?

Answers (1)

Srivardhan Gadila
Srivardhan Gadila on 14 Jun 2021
In order to implement the above problem the output of the softmax layer should be 125x3 or 125x1x3 etc., you can create a custom layer to reshape the output 375 (= 3*125) from fullyconnected layer and insert it before the softmax layer. Make sure that the dimension of 125 corresponds to the dimension label 'S' and that of 3 corresponds to 'C'. You can refer to Define Custom Deep Learning Layers and addLayers for implementing the said functionality. But again later you may encounter issues with figuring out what should be the data format to give to the trainNetwork function. Hence I would suggest to go with dlnetwork & Deep Learning Custom Training Loops based approach instead of layerGraph and trainNetwork based workflow.
Apart from the above suggestions I think in general this may not be a good idea of predicting the class labels of higher dimensional data (125) from lower dimensional data (12), it may or may not give the good results.

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!