Can you use a different image size during transfer learning?

Question

I have made a switch from TensorFlow to PyTorch recently. I use a famous Github repo for training on EfficientNets. I wrote the model initiation class as follows:

class CustomEfficientNet(nn.Module):
    def __init__(self, config: type, pretrained: bool=True):
        super().__init__()
        self.config = config
        self.model = geffnet.create_model(
            model_name='EfficientNetB5',
            pretrained=pretrained)
        n_features = self.model.classifier.in_features
        self.model.classifier = nn.Linear(n_features, num_classes=5)
        

    def forward(self, input_neurons):
        output_predictions = self.model(input_neurons)
        return output_predictions

In addition, in my transforms, I tend to useResize(img_size = 512, img_size=512) for my training on certain image classification tasks (Mostly Kaggle Competitions). So the question here is, the official input size for EfficientNetB5 is 456x456, but I used 512x512 or even 256x256 and get very decent results. Is this normal? Or did I miss out the source code where the author will resize into the native resolution for you?

PS: This seems to be the norm in all the PyTorch Tutorials I saw on Kaggle. My full code can be seen here in this notebook here; I like to not leave logic gaps and therefore this question popped up.

Answer 1

Yes you can use different input sizes when it comes to transfer learning, after all the model that you load is just the set of weights of the fixed sequence of layers and fixed convolution kernel sizes. But I believe that there is some sort of minimum size that the model needs to work efficiently. You would still need to re-train the model but it will still converge quite quickly.

You would have to check the official implementation on the minimum size of the model like the one in VGG16 where they specify that the width and height need to be at least 32.