Very nice code for this paper.
I found that the performance of cnn under attach is very unstable. Sometimes the performance can be much better than claimed in your table. The improvement from kervolution is limited.
I found a little mistake in your implementation that can affect the performance. The bias term should be added after the kernel trick, not as in your code. Better to change it.
Very nice code for this paper.
I found that the performance of cnn under attach is very unstable. Sometimes the performance can be much better than claimed in your table. The improvement from kervolution is limited.
I found a little mistake in your implementation that can affect the performance. The bias term should be added after the kernel trick, not as in your code. Better to change it.