transfer learning using Pre-trained vgg-16

How can I train my own CNN using the pre-trained imagenet-vgg-verydeep-16 network. I have tried to use the fully connected layer 'fc7' to train my network and I get the following error.

Undefined function or variable 'activations'. trainingFeatures = activations(convnet, trainingSet, featureLayer, ...

can anybody help please.

Answer 1

[mizuki]

The function "activations" is in Neural Network Toolbox. Do you see NNTB in your license? First check with the following command:

>> ver

Also, the path might not be set correctly. Please follow the below article.

https://jp.mathworks.com/matlabcentral/answers/96005

[Burhan]

thank you mizuki

[Burhan]

Dear Mizuki, if NNTB is meant for Neural Network Toolbox then it was present when I typed >>ver from command line and that is

Neural Network Toolbox Version 10.0 (R2017a)

I also used 'which' function that returned 'activations not found. It however works for Alexnet pre-trained network. I can't figure out, how?

[Burhan]

sir, one thing more! from Matalab documentation i checked activations function and it exists with same name (and letter case too).

[mizuki]

NNTB appeared to the output of VER, which means NNTB is installed. I did not get what you have tried exactly. Did you type

>> which -all activations

and resulted in the error message " activations not found"? If so, that is fine since activations is a class, which is not usually searched with doc command.

The situation that AlexNet worked fine but VGG16 does not work sounds that your installation of VGG16 did not work well. How about the result of

>> which -all vgg16

Have you tried the following commands, which is on the page that I mentioned before?

>> restoredefaultpath 
>> rehash toolboxcache
>> savepath

Also, ACTIVATIONS is: - activations https://www.mathworks.com/help/releases/R2017a/nnet/ref/seriesnetwork.activations.html

Answer 2

[Burhan]

1) I typed "ver" and Following line was one of the outputs.

Neural Network Toolbox Version 10.0 (R2017a)

2) I then checked "which -all activations" and the result is: activations not found

3) The result for "which -all vgg16" is: C:\Program Files\MATLAB\R2017a\toolbox\nnet\cnn\vgg16.m

4) Yes, I tried >> restoredefaultpath >> rehash toolboxcache >> savepath

Yet again the following line of my code:

trainingFeatures = activations(convnet, trainSet, featureLayer, ... 'MiniBatchSize', 32, 'OutputAs', 'rows');

produces error: Undefined function or variable 'activations'.