Why RMSE doens't decrease in deeplearning toolbox?

Question

영진 김 2024-2-19

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2083693-why-rmse-doens-t-decrease-in-deeplearning-toolbox

回答： Angelo Yeo 2024-2-19

training_set2.csv

在 MATLAB Online 中打开

Hi, I'm stduying Matlab deep learning toolbox, and confused about using regression layer.

I want to make regression model for 8 input and 1 output. Input and output are all feature.

How can I make model which RMSE goes dereasing?

Should I control hyperparameter to fix this situation?

Plz help my project. Thanks.

(data file is uploaded)

clc;clear;
filename = "training_set2.csv";
tbl = readtable(filename,'TextType','string');
head(tbl)
edges = 0.41:0.01:0.54; %discretize data
responses=cell2mat(table2cell(tbl(:,"FM")));
tbl(:,"FM") = cell2table(num2cell(discretize(responses,edges)));
labelName = "FM";
numObservations = size(tbl,1);
numObservationsTrain = floor(0.8*numObservations);
numObservationsValidation = floor(0.15*numObservations);
numObservationsTest = numObservations - numObservationsTrain - numObservationsValidation;
idx = randperm(numObservations);
idxTrain = idx(1:numObservationsTrain);
idxValidation = idx(numObservationsTrain+1:numObservationsTrain+numObservationsValidation);
idxTest = idx(numObservationsTrain+numObservationsValidation+1:end);
tblTrain = cell2mat(table2cell(tbl(idxTrain,1:9)));
responseTrain = cell2mat(table2cell(tbl(idxTrain,10)));
tblValidation = cell2mat(table2cell(tbl(idxValidation,1:9)));
responseValidation = cell2mat(table2cell(tbl(idxValidation,10)));
tblTest = cell2mat(table2cell(tbl(idxTest,1:9)));
responseTest = cell2mat(table2cell(tbl(idxTest,10)));
numFeatures  = size(tbl,2) - 1;
numClasses = 1;
layers = [
    featureInputLayer(numFeatures)
    fullyConnectedLayer(50)
    batchNormalizationLayer
    reluLayer
    fullyConnectedLayer(numClasses)
    softmaxLayer
    regressionLayer]; 
miniBatchSize = 1;
options = trainingOptions('adam', ...
    'InitialLearnRate',0.0001, ...
    'MiniBatchSize',miniBatchSize, ...
    'MaxEpochs',50,...
    'Shuffle','every-epoch', ...
    'Plots','training-progress', ...
    'ValidationData',{tblValidation,responseValidation}, ...
    'Verbose',false);
net = trainNetwork(tblTrain,responseTrain,layers,options);

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Angelo Yeo 2024-2-19

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2083693-why-rmse-doens-t-decrease-in-deeplearning-toolbox#answer_1411468

The training environment is not ideal to use deep neural networks. A few comments:

Why did you discretize the output, "FM", and use softmaxLayer and regressionLayer at the same time? There is no point to discretize "FM" but to use "regressionLayer".
The number of samples is too small. The total number of data is 152. Which is too small to train a neural network.
Also, it is hard to distinguish between the samples. Most of features are identical or correlated.