About resolution generalization

[Silensea]

Thank you for offering this incredible repository! It has been a great help in my learning of neural operators.

I tried the UNO_Tutorial.ipynb and wanted to test the resolution generalization. I then tested the model with an input size of [32, 32] (while the training data size is [64, 64]). However, I encountered an error related to tensor matching issues.

I initially thought that UNO could directly support inputs of different resolutions, but it seems that additional adjustments might be necessary. Could you please advise if the model architecture needs to be modified to handle experiments with new resolutions?

[ashiq24]

Hi @Silensea

Thank you for bringing up this issue!

You are absolutely correct — the current implementation only supports scaling up to higher resolutions, such as training at a resolution of $64 \times 64$ and testing at $100 \times 100$. While the model is intended to also support scaling down to lower resolutions than the training data, this functionality is not implemented here.

Typically, most use cases involve training on lower resolutions (since generating and training on high-resolution data is computationally expensive) and then evaluating on higher resolutions. However, if your specific requirement is to train on higher resolutions and test on lower ones, we have a couple of options:

1. You could use the UNO model available in the neuraloperator library, which supports this setting
2. If you prefer to use this implementation, I can make the necessary changes to add support for lower-resolution evaluation.

Let me know which option works best for you, and I’ll be happy to assist further!

Best regards,
Ashiq