So I’m building an app that counts money from a photo, but honestly, local training has been a total nightmare. Spent 12 hours straight just cleaning and stitching datasets together, only for training on my 3070 to go horribly because of VRAM limits forcing me to downscale everything. Looking at the test image, you can see it’s detecting something, but it completely misses the fine details and fails to classify them because it has zero concept of relative sizes. I’m pivoting to a Transformer-based model (RF-DETR) so it can actually look at the whole picture and compare coin sizes better. Renting a beefy GPU in the cloud to do the training.🙄