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.