Compute receptive field and tiling automatically

[jdeschamps]

Problem

Currently, users have to guess the tiling size and overlaps during prediction. There is a fine balance to have between GPU memory, prediction speed, UNets' pseudo translation-invariance and receptive field.

When it comes to the UNet, the receptive field should be easy to compute, and the constraints on the tiling and overlaps are mathematically well-defined. GPU memory can be guessed in the system and so it should be easy to automatically compute coherent tile and overlap sizes.

References

• a package to determine receptive fields: https://github.com/shelfwise/receptivefield
• CSBDeep receptive field calculation
• CSBDeep tries large tiles and half them if an error is raised (iteratively)