Commonly we extract features using:
net = googlenet()
%Extract features
featureLayer = 'pool5-drop_7x7_s1'; #the last layer before softmax
How to extract features from a different layer earlier in the network? I am attempting to extract from a different layer but the output of the layer is in 4D. I tried to use squeeze function but its not working for those layers other than the last layer. Did anyone find success in this task? Kindly assist.
If you are using Caffe framework. follow the link
https://ahmedibrahimvt.wordpress.com/2016/04/18/caffe-in-matlab-matcaffe-simple-tutorial/
if you are using matconvnet framework. follow this link
http://www.vlfeat.org/matconvnet/this
you can easily extract features through Matconvnet framework.
hi?
Altogether, max Pooling pooling units are also able to implement other functions such as average pooling or even L2-norm pooling. Average pooling was initially most often used until recently, compared with max pooling, which in practice gave better performance. The trend has vanished.
As other friends have already suggested, I think you are using MatConvNet.
Yes the output of various layers are 4D. for feature extraction you need 2D vector I guess. You can use Reshape in Matlab to change the 4D matrix to a 2D vector. Then you can apply some dimensionality reduction methods like FV or PCA or BoW and so on, to reduce the size of feature vector.
Hi Sivaramakrishnan Rajaraman,
I hope you have your answer already. However, if you didn't yet, here is my answer might help you. Assume you have your trained GoogLeNet model on you MATLAB and you want to extract features from featureLayer = 'pool5-drop_7x7_s1'; layer. All you have to do is very simple thing, just activate the layer to extract features from that particular layers ans save as a veritable.
trainingFeatures = activations(net, TrainingSet, featureLayer, ...
'MiniBatchSize', 32, 'OutputAs', 'columns');
The command mentioned above will extract the features from the layer you have defined as "featureLayer", and save it as columns matrix. To explore more about it, please find yourself over "https://au.mathworks.com/help/nnet/ref/activations.html".
Regards,
Tohid
- Badges
- Science topic