Different behavior of 'activations' function of Neural Network Toolbox between 2017b and 2018a.

Question

HIDEYA MATSUDA 2018-5-4

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/398965-different-behavior-of-activations-function-of-neural-network-toolbox-between-2017b-and-2018a

评论： HIDEYA MATSUDA 2018-5-4

If I tried CNN sample code obtained here, I got error on 2018a and not on 2017a. Is it a bug or expected behavior? How can I fixed it on 2018a?

Obtained sample code here.https://jp.mathworks.com/matlabcentral/fileexchange/57280-cnn---transfer-learning-example
Run classifyFlowers.m
Hit following error

Error: classifyFlowers (line 39)
classifier = fitcsvm(trainingFeatures, trainingSet.Labels);

Findings:

I found that the size of matrix of returned value from 'activations' function was different.

trainingFeatures = activations(convnet, trainingSet, featureLayer, 'MiniBatchSize', 1);

2017b (Expected):

>> size(trainingFeatures)
ans =
           112        4096

2018a (Not Expected):

>> size(trainingFeatures)
ans =
           1           1        4096         112

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

Johannes Bergstrom 2018-5-4

1
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/398965-different-behavior-of-activations-function-of-neural-network-toolbox-between-2017b-and-2018a#answer_318614

This is the expected behavior. Please see the release notes under 'Functionality Being Removed or Changed', https://www.mathworks.com/help/nnet/release-notes.html

Default output format of activations: In R2018a, the new default output format of activations is a 4-D array. To reproduce the old default behavior, set the 'OutputAs' value to 'rows'.