I have an image classifier trained on resnet34, using image size 234 on my own image dataset. If I simply change the architecture to be resnet50 and increase the image size to be 300 instead just like in the course, I get worst performance (3% less accuracy). I have around 950k images… I would have thought that resnet50 having more parameters could take advantage of the larger size of the images.
What could cause a worst performance from resnet50?
Thanks,