It applies a function, like lambda in python.
thanks, got it. Just was confused about the explanation what this tensor is going to do (filter bottom-right edges).
1 Like
The kernel will pick out brightness changes in the vertical direction, which corresponds to horizontal edges.
1 Like
So is a feature map that final 512 vector ? Or the layer juste before of 11x11x512 ? I’ve read that term several times in papers and I’m never too sure of what it is
What is being explained now is done before flattening out the image right?
It looks like the adaptive pooling is at the end of the network. But how does the beginning of the network handle image of different sizes? It looks like from the summary that the network expects an image of a certain size at the beginning. is that correct?
The feature map is the 11x11x512 grid. The 512 are the features and the 11x11 are the map.
3 Likes
No, the summary just picked the size of your data. The model in fastai can deal with any input size (as long as it’s bigger than 32x32).
so a row corresponds to vertical edge in image ? and column vertical
avg_acts or avg_cats? 
Average pooling corresponds to “smoothing” the image.
but i think i have seen Jeremy change the size of the inputs.
do training with small images. then train with larger images, repeat.
Isnt maximum of activation a better measure to check than average for the heatmap?
Again, the model doesn’t care about the size of images and can work on any size, that’s the beauty of a CNN.
1 Like
The convolutional layers don’t care what size the image is
Can you explain how this works?
1 Like
why would you use both average and max pool ?
1 Like
This corresponds to choosing “top sobel” for the kernel here if you want to test it out: http://setosa.io/ev/image-kernels/
Both would work.
1 Like
To take the better part of each of them
2 Likes