Hi All, I am working with a dataset where the distribution of the columns and rows looks something like this
In past i have only worked with images like 256,256 or 512,512. Any pointers on how to handle such large dimensions. Is resizing to 256x256 is a viable option. I am hesitant on this as It would squish things and model might not learn relative positions properly ?? Thoughts and pointers much appreciated.
In my experience with fastai v1 so far I’ve been able to use images even larger than that. I have tested up to 570x570, and compared them to smaller image sizes. Best results were achieved with larger image sizes.
I have a post where I show the data I’ve gathered.
An important aspect you need to consider with images though is memory. I’m using GCP p100 which I think has 30 GB of memory, and had to lower batch size with resnet50 to 16.