Test-Time Augmentations (TTA) can improve accuracy during inference.
It would be interesting to allow the use of TTA in geo-inference, to further boost performance.
Example with pytorch: https://github.com/qubvel/ttach
Questions for discussions:
How would that look like in a Dask/Xarray world?
There would certainly have an impact of processing time?!
Test-Time Augmentations (TTA) can improve accuracy during inference.
It would be interesting to allow the use of TTA in geo-inference, to further boost performance.
Example with pytorch: https://github.com/qubvel/ttach
Questions for discussions: