Open yhr91 opened 6 years ago
Questions to think about:
Something else I have experimented with: based on the distribution of image sizes, define a number of fixed bins, e.g. '256', '512', '1024' and assign every image to the best-fitting bin and pad with zeros. Downside: you have to train for each bin separately and it will definitely increase overall memory and compute demands. I have working code for this that I wrote for my protein work - but there image sizes were fairly spread out. Maybe resizing all images to the most common dimension would be best in this case.
Feel free to assign yourself to this issue