I’m working through Week 3 of the Convolutional Neural Networks course with Andrew Ng. and confused about the implementation of a convolutional sliding window (https://www.coursera.org/learn/convolutional-neural-networks/lecture/6UnU4/convolutional-implementation-of-sliding-windows).
I understand the why behind it, and understand that this will be one forward-propagation pass that will be significantly more efficient than running multiple forward passes across a large number of sliding window convolutions…but I’m failing to understand how this is done in practice or how to implement.
As an aside, can we apply this same convolutional sliding window technique to time series?
Have you seen the week 1 assignment 1 ( Convolutional Neural Networks: Step by Step) of this course?
Yes, we can use Conv1D for time series.
Yes I have and completed that with 100%. I’m not confused about the standard sliding method but rather the new method introduced in this particular lecture.
It’s been a while since I went through these lectures, so cannot recall what new method introduced here. Hopefully, one of the other mentors will help you out.
Like Saif, it’s been 3 or 4 years since I watched those lectures. Fortunately I’m retired, so I had time this morning to go back and watch them again to refresh my memory.
I think the idea is actually pretty simple: the way convolutions work is that they naturally include the “stepping across and down” the image, right? That’s fundamental to the process, so “sliding window” is just the action of the filter and the “stride”. I think that’s all he’s saying. In the real “sliding window” implementation, I think what he means is that you run a classifier separately on each position of the window. If you do it as a convolution on the whole image, then we are training the filters to work in all locations in the normal convolution process striding across the whole image.
Of course we later get to the full glory of YOLO and recognizing multiple objects within each grid cell, so stay tuned for that.
1 Like