CODEWITHSUNDEEP
deep-learningconv-neural-network
user3668129
user3668129

Reputation: 4810

What are the cases we should use stride which is not 1?

I read a lot of tutorials and view some posts which used CNN with, stride 1 or stride 2 (for down sampling).

But I didn't saw any examples with stride of length 3 or above.

Upvotes: 0

Views: 964

Answers (2)

Hossein
Hossein

Reputation: 25924

larger strides mean you are effectively gaining more speed/using less memory footprint at the expense of losing much more valuable information that can ultimately influence your accuracy.

Also, note that larger strides can also be used when you want to take advantage of the (usually large) input at hand in your model and do the downsampling while the learning process is underway as opposed to resizing the images to much smaller dimensions and starting off using more smaller strides.

apart from this, refrain from using large strides throughout the network blindly as it hinders your model's learning capacity.

There is always a cost/benefit scenario associated with actions like this. It's up to you, the architecture designer, that based on the requirements at hand, to make the call to do what or which compromises to make.

Upvotes: 1

Malo
Malo

Reputation: 1313

I suppose it depends on the size of your filter. The choice is a compromise between speed and accuracy.

Upvotes: 0

Related Questions