In the notation of filter f^[l] = 5. Does this mean a filter of size 5 x 5 x n where n = number of channels or does it mean a filter of size 5 x 5 ?
Hello Nithin,
It would be great if you could guide me to the Week-Video-Assignment that is related to your questions so I can help you more.
Thanks!
Course 4, week 1, “simple convolution network example” video at 2:49 Andrew says 5 x 5 filter but how is it possible to apply a 2D filter to a 3D volume. This is my exact doubt hope u clear my doubt.
Thanks!
This is a great question, my friend.
As you know, Andrew decided to use 5x5 filters with a stride of 2. Thus, he will be merging every five matrix cells into one and will move with a step of two cells.
We had 37 cells, if we divided them into groups of 5 and moved your stride with a step of two, you will be left with 17 cells. It would be helpful if you use a one-dimensional example where you have an array of 37 cells and you apply a 5 filter on it.
Good enough? If not, feel free to ask me to clarify more.
Regards,
1 Like
I understood what u said but…
In the pic I attached the size of filter f^[2] (which I marked in red box) is 5 x 5 (2D filter) but activation a^[1] is a 3D Volume. How can a 2D filter f^[2] be applied on a 3D volume a[1].
Thanks!
1 Like
The filters are always applied across all the input channels at once. In the example you show, the input is 39 x 39 x 3, so there are three input channels. Given the filter size of 3 for layer 1, then each filter is 3 x 3 x 3. Since there are a total of 10 filters, we end up with 10 channels in the second layer. In the second layer, the filter size is 5, so each filter is 5 x 5 x 10.
2 Likes
My doubt is clarified.
Thanks
1 Like