Some questions about the code of the svm prediction part about cnn-svm HOT 8 CLOSED

@anbo1024 it can be used in a one-versus-all manner, i.e. in a one-hot encoded vector, instead of using 0 for the classes that a sample does not belong, we use -1 : [0, 1, 0, 0] --> [-1, 1, -1, 1]

from cnn-svm.

@AFAgarap I still have some doubts.
1、You will [0, 1, 0, 0] --> [-1, 1, -1, 1]，Softmax regards the maximum subscript as the final classification result. How do you determine the final classification result with this encoding method?
2、I think that in CNN, whether using SVM or Softmax, the network output has been classified. SVM and Softmax calculate loss in different ways.

output = tf.identity(tf.sign(output), name='prediction')
output = tf.identity((output), name='prediction')
correct_prediction = tf.equal(tf.argmax(output, 1), tf.argmax(y_input, 1))

I think tf. sign and tf. nn. softmax functions can be removed.The following Correct_prediction function already contains the maximum operation.
I hope I can get your advice. Thank you very much.

from cnn-svm.

Well, in terms of having a one-hot encoded vector with a -1 in place of 0, e.g. [-1, 1, -1, -1], which one is greater, -1 or 1? 1, yes? So, it's still fine to be there. But you're right, it can be discarded for the purposes of getting the training accuracy.

from cnn-svm.

@AFAgarap I understand what you mean. Now, I use SVM as a loss function on my own dataset, but the performance is very poor, and Softmax performs very well. Why is this? Is the data still to be preprocessed?

from cnn-svm.

In Yichuan Tang's paper, Deep Learning using Linear Support Vector Machines, he used PCA and added Gaussian noise for MNIST. But that was on a feed-forward neural network with 2 layers having 512 units each.

from cnn-svm.

The paper is here

from cnn-svm.

Yes, I have read this article. Good results have been achieved in your code. Did you preprocess it?

from cnn-svm.

No, I didn't

from cnn-svm.