max nargin decision problem in code

[mcu1988]

In cnn-svm/model/cnn_svm.py , output of one versus rest svm is written as :
output = tf.identity(tf.sign(output), name='prediction')
correct_prediction = tf.equal(tf.argmax(output, 1), tf.argmax(y_input, 1))

tf.sign(output) seems incorrect. It should be removed. SVM here use max margin decision. When output margins of multiple binary linear SVM is converted to 1 and -1, tf.argmax(output, 1) may return a false value, because more than one binary linear SVM's output may be converted to 1 as they may all output a positive margin.

[AFAgarap]

@mcu1988 Will you please describe the problem you see in the implementation? Thank you. :)

[mcu1988]

from Deep Learning using Linear Support Vector Machines (2013)
Multiple SVM is described in this picture. Output of Multiple SVM is defined in (9). There is no need to convert output margins of each binary SVM to 1 and -1. When more than one svm outputs a positive margin value, these positive margin are converted to 1 by using output = tf.identity(tf.sign(output), name='prediction'). correct_prediction = tf.equal(tf.argmax(output, 1), tf.argmax(y_input, 1)) may return a false value.

[AFAgarap]

The conversion to {+1, -1} is as per the SVM model design written by Vapnik and Cortes in their paper.