Remote Sensing Example

[nilsleh]

Hi, Thanks for providing the code to your paper. I was wondering if there is an ETA for the remote sensing example you provide in your paper. We are trying to add your method to our UQ library. Our background is Earth Observation data and we are planning a new release soonish, so we would like include your method and create a tutorial notebook showcasing the method for a remote sensing task. Cheers and thanks again!

[nikitadurasov]

Hey @nilsleh, thanks for the interest in our work!

Sure, I think it would be great adding our approach to the library. How do you think the tutorial would look like?

We were thinking about sharing pretrained model with some input images showcasing confident / uncertain cases (similar to the teaser image in the paper).

I'm not sure about time, but let's say we'll try to figure it out within a couple of weeks?

Let me know if it sounds good!

Best, Nikita

[nilsleh]

Thanks for your reply, the question was really out of curiosity and didn't mean to pressure time-wise! So personally, there would be two aspects that we like to consider for a notebook:

• one, training a module with our integrated Lightning setup, to make sure we can reproduce your results, such that we have "some" confidence about having it implemented correctly. So for that case I suppose it would be checking that we can get similar quantitative results and generate a figure like the one you have in your paper
• showing how a user can plug in their own dataset to apply the method for their task of interest

We tried doing that with the ZigZag MNIST Example (adding some explanation and mathematical details to the notebook is still on our ToDO List). So for the remote sensing example with the Iterative Method, we would:

• train on the dataset with our implementation
• evaluate and check that we can reproduce the results
• have the training code in the notebook, so people know what happened, but for the purpose only load the pretrained checkpoint
• with the pretrained checkpoint give some visual intuitions about the UQ with some nice figures that demonstrate the concepts

Generally we try to structure the notebooks as follows, for example in this SWAG example:

• give a quick mathematical introduction to a method
• show how you can use that method with code on a dataset
• show some qualitative and quantitative results

From our perspective, we would only need to have details about the dataset and training scheme you used (some script similar to the ones you have provided for ZigZag MNIST or the regression cases here, would be helpful), so we can try to reproduce it, and then we can add the notebook. Of course, if you like and have any time availability, we'd also be happy if you want to add some details to a notebook or provide feedback/suggestions etc. Thanks in advance :)

[nikitadurasov]

Hey @nilsleh,

Sounds good! The issue with the code release for remote sensing experiments is that my colleague responsible for them recently left the lab. However, I've obtained the code from him. We have two options now: I can share the model architecture, training pipeline, data, etc., with you so you can re-implement it. Meanwhile, I'll prepare the code to add to this repository. I believe that's the best option; please email me at my university address :)

Additionally, I could release the MNIST experiments for the ICML paper (similar to what you have for Zigzag). It's a good idea too.

Best, Nikita