I want to train a resnet 50 model for an image clasification task. The default model requires 224×224 images, my dataset has 64×64 images and it seems wasteful to first upscale them then train on that data.

How should I modify the layers so that it not only accepts the 64×64 image, but retains a decent size by the final layer, when it reaches the fully conected classification layer?

This is my keggle notebook: https://www.kaggle.com/code/cristiciorba/cv-resnet-lego-bricks