Non-deterministic inference question

[spadarian]

Hi,

First, thanks for the amazing job you have done!

I managed to train my own model and, while I was running an external validation, I noticed that the inference is non-deterministic. I'm following the example notebook (4.4. Applying SPT), specifically using voxel_semantic_pred.

I'm assuming that there is some augmentation happening at inference time but not sure. I added tta_runs: null to my config (after training, since I just noticed it was happening), but still get different predictions every time I run the model.

Any advice would be greatly appreciated!

[spadarian]

I just noticed #186

I will have a closer look at that

[spadarian]

I'm a bit confused about the best way to run the model at inference time (with a different dataset). Is there a recommended way of doing it?

Do I need to create a new "experiment" and set all the augmentations to 0 and disable the sample_* parameters?

Thanks!