900 x 50 image size doubts

My dataset has 40000 images of 900x50 resolution, point of interest is across the whole image and they don’t look like Imagenet at all.

What special steps should I take to train a Resnet model as good as I can given these conditions?

You could straight use Resnet34 for instance, is size independent, but… you may be better with a custom Darknet for instance. Is your problem a classification or segmentation?
Also augmentations like square images, so some tunning may be needed.
You would have to modify some conv layers strides to be sure to not reduce below 1 the images. You could also cut the original images, maybe in 300x50 (so you get 3 of them).
If you need more help doin this, just send me a PM and I could help you.

Thanks, I sent you a PM.