I have a question regarding the EfficientNetV2 family of models. If my understanding is correct there are 6 models under this family - B0 to B1 & S are the comparatively smaller models while M & L are the larger ones. However, I'm having difficulty understanding the expected input dimensions (image resolution) for these models.

In my dataset, the image size is 400x400. I can convolve the images to reduce the size or add zero padding to increase it, but I'm unsure about the model's actual expectations.

I have been referring to the paper EfficientNetV2: Smaller Models and Faster Training for guidance, but I might be missing something. I would greatly appreciate it if anyone could provide some insights or direct me to the correct information.

Thank you in advance for your help! Cheers!

(Edit: While looking further into the problem, I think we can simply provide the target_size in the tf.keras.preprocessing.image.ImageDataGenerator function & it will resize the input image to the given resolution? However, I'd still like to get an answer to my original question, if possible)


1 Answer 1


There is no single image size that the EfficientNetV2 models expect going just of the method described in the paper. As shown in figure 4 from the paper, the models are trained by starting with an image of a small size, and as training progresses the image size increases in addition to increased regularization. The minimum and maximum image size used to train the model depends on the model size, table 6 shows the parameters used in the progressive training approach. The minimum size for all models (S, M, and L) is 128 pixels, with the maximum size being 300 pixels for the S model and 380 pixels for the M and L models.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.