Regnet beat Efficientnet and is 5x faster on GPU

fastai
1

The paper seems very interesting: Designing Network Design Spaces.

12 Likes
2

Thanks for posting this - it looks very promising.
The github for it is here:

but code for the actual RegNets is not there yet. Hopefully soon.

3 Likes
3

@LessW2020 I wrote an article about it.

Also the code will be released here


as it is, the codebase only contains AnyNet, the original design space for RegNetX. In principle it would be possible to take their design philosophies to reconstruct RegNetX but i’m sure we wont have to wait too long for them to release the code.

6 Likes
4

cool, enjoyed the article!
I am going to pull the AnyNet code now and run through it just to get ready for RegNet proper.
I’ve been using EfficientNet extensively so very interested to see a strong competitor for the first time.

4 Likes
5

Just tweeted about RegNet as I ran a few experiments on the weekend. Nothing groundbreaking, a decent model in the ResNet family, quite close to SE-ResNeXT for the Y and ResNeXT for the X variants. Nothing close to a 5x difference with EfficientNet. Similar to comparing with other ResNet family models you get generally comparable throughputs +/-20% at comparable accuracies, so nothing too different here.

4 Likes