Implement an easy way to incoporate transforms into the inference of a trained model.
- make N transforms
- maybe only apply one TA transform for inference?
At the end, probably the average of these predictions should be the final prediction.
Optionally: use transforms also for validation set (but maybe with different logic than training?) to ensure the best validation checkpoint actually has OOD performance?
Implement an easy way to incoporate transforms into the inference of a trained model.
At the end, probably the average of these predictions should be the final prediction.
Optionally: use transforms also for validation set (but maybe with different logic than training?) to ensure the best validation checkpoint actually has OOD performance?