Closed yellowcap closed 3 months ago
Good idea but not on the roadmap for now.
Note: While doing multinode with slurm, we have to adjust num_tasks, num_workers while changing the chip_size, this was not trivial to implement, will skip this for now.
To train the multi size images architecture, we need to either stop/restart the model every ~10 epochs, changing progressively the image size in the data loader.
Preferrably we would implement a pytorch hook to make this change automatically.