Input size for EfficientNet versions from torchvision.models

vision
1

Hi guys!
I’m doing some experiments with the EfficientNet as a backbone.
I’m using the pre-trained EfficientNet models from torchvision.models.
As I found from the paper and the docs of Keras, the EfficientNet variants have different input sizes as below.

image
image734×355 11.2 KB

Is it true for the models in Pytorch?
If I want to keep the same input size for all the EfficientNet variants, will it affect the performance of the models?

2

Yes, the torchvision EfficienNet models also expect a different input shape as described in the docs:

The sizes of the EfficientNet models depend on the variant. For the exact input sizes check here

where the link points to:

    elif args.model.startswith('efficientnet_'):
        sizes = {
            'b0': (256, 224), 'b1': (256, 240), 'b2': (288, 288), 'b3': (320, 300),
            'b4': (384, 380), 'b5': (489, 456), 'b6': (561, 528), 'b7': (633, 600),
        }
2 Likes
3

Hello @ptrblck.

Do you know if that sizes are still used in torchvision v.12.0 ?
I can not find any information or reference in the documentation or v0.12.0 code

Thanks

4

You can read more information about each model by selecting it from here. which also describes the input size.

1 Like
5

Thanks for your answer.
As you said, that is in the v0.13 but I am still in the v0.12 and documentation says nothing, check out this example or this :slight_smile:

That’s why I wonder about if the pre-processes in v0.13 and v0.12 are the same

6

For older versions you could use my previously posted link and change the branch to 0.12: Models and pre-trained weights — Torchvision 0.12 documentation

1 Like