I've checked your code and found that in C2AE-Multilabel-Classification/model/src/network.py, line 79, we should:
return tf.norm(Fx-Fe)
instread of:
return tf.reduce_mean(tf.square(Fx - Fe))
Since in the paper, the authors defined the loss as the Frobenius norm of Fx-Fe (the equation 2).
tf.reduce_mean just compute the expected value from all the elements of the input.
What do you think in this case 🍡 ?
Hi Dhruvramani,
I've checked your code and found that in C2AE-Multilabel-Classification/model/src/network.py, line 79, we should:
return tf.norm(Fx-Fe)instread of:return tf.reduce_mean(tf.square(Fx - Fe))Since in the paper, the authors defined the loss as the Frobenius norm of Fx-Fe (the equation 2). tf.reduce_mean just compute the expected value from all the elements of the input. What do you think in this case 🍡 ?