Transfer learning input layer size

Hi all,

When we select a pretrained model for transfer learning, does it have to use the same input layer size as the fine tuning model too? For example is it possible to use models trained on images of different dimensions/sizes for OCR applications?

Thanks in advance!

Hi @Jerry_Zheng,

I have never seen any pre-trained model that doesn’t have a fixed requirement on input, so the general understanding is, we need to resize our photos to fit to the model’s input layer. However, you may try to test what resizing techniques are feasible, and by “feasible” I mean that it will not suffer a huge performance loss. When your photos are smaller than expected, besides stretching them, it is worthwhile to test if padding the extra space with white (or other) color does better because in that case you are not changing the ratio of the content in your original photos.