setting input data division ratio
조회 수: 14 (최근 30일)
이전 댓글 표시
how can i change division ratio for input data ???...input data is divided by default into 60% for training data , 20 % for validation & 20 % test data....i want to change these values to 70% for training & 15 % for validation & test....i used the following commands to change them: net.divideparam.trainratio=0.7; net.divideparam.valratio=0.15; net.divideparam.testratio=0.15; but after running the program i didn't find any change occured in division of samples !!!
댓글 수: 1
Jonny
2011년 10월 7일
Hi Hoda,
Can you please,show me what you edit in divideparam in dividerand.m file?
Thanks
채택된 답변
Lucas García
2011년 8월 25일
Hi Hoda,
Let me paste an example of a simple 2-layer Feed-Forward network, to see if this works for you (you should be able to reproduce with the same dataset -cancer_dataset.mat-, it comes with the NN toolbox):
load cancer_dataset;
% 2 neurons in the first layer (tansig) and 1 neuron in the second layer
% (purelin).
% Levenberg-Maquardt Backpropagation Method is used
mlp_net = newff(cancerInputs,cancerTargets,2,{'tansig'},'trainlm');
% Different sets are randomly created for training, validation and testing
% the network
mlp_net.divideParam.trainRatio = 0.6;
mlp_net.divideParam.valRatio = 0.2;
mlp_net.divideParam.testRatio = 0.2;
mlp_net.trainparam.epochs = 100;
[mlp_net,tr] = train(mlp_net,cancerInputs,cancerTargets);
% Once the network has been trained, we can obtain the Mean Squared Error
% for the best epoch (time when the training has stopped in order to avoid
% overfitting the network).
mse_train = tr.perf(tr.best_epoch + 1); % There is epoch 0, but arrays in
% MATLAB start in 1.
mse_val = tr.vperf(tr.best_epoch + 1);
mse_test = tr.tperf(tr.best_epoch + 1);
Now, if you check train, validation and test ratios, after the training, you should get:
>> mlp_net.divideParam
ans =
Function Parameters for 'dividerand'
Training Ratio trainRatio: 0.6
Validation Ratio valRatio: 0.2
Test Ratio testRatio: 0.2
By the way, if you are using MATLAB R2011a, you should use feedforwardnet instead of newff.
Hope it helps.
댓글 수: 3
Lucas García
2011년 8월 26일
It's good to see that you found a way to set your own default values.
The plot epochs vs. gradient is related to the performance function. The network is minimizing some performance function (mean-square error of targets and outputs, by default).
The gradient plot is showing you the gradient of the performance function at each epoch. As the gradient (derivatives) gets smaller and closer to zero, the function will be minimized. That will imply that the outputs are very close to the targets and therefore the network is trained.
Jonny
2011년 10월 7일
Hi Lucas,
Are different sets randomly created for training, validation and testing from input (matrix) in network?
Also could you explain me what these three lines of code work
mse_train = tr.perf(tr.best_epoch + 1); % There is epoch 0, but arrays in
% MATLAB start in 1.
mse_val = tr.vperf(tr.best_epoch + 1);
mse_test = tr.tperf(tr.best_epoch + 1);
what is mse_train,mse_val, mse_test - for what they are used further?
thanks for your help :)
추가 답변 (1개)
grytan sarkar
2017년 5월 12일
Thanks Lucas.
mse_train = tr.perf(tr.best_epoch + 1); % There is epoch 0, but arrays in % MATLAB start in 1. mse_val = tr.vperf(tr.best_epoch + 1); mse_test = tr.tperf(tr.best_epoch + 1); Its really good for me.
I need more information. Is it possible to determine the mean, standard deviation, minimum and maximum value of each input in training, testing and validation data set. If i consider random selection.
댓글 수: 1
Greg Heath
2017년 5월 12일
Yes.
It should be straightforward.
What seems to be the problem?
Show your attempts and we will see if it looks ok.
Hope this helps.
Greg
참고 항목
카테고리
Help Center 및 File Exchange에서 Sequence and Numeric Feature Data Workflows에 대해 자세히 알아보기
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!