import error （AuroraAirPollution）

[yhchen1112]

Hi,
First of all, thank you for your amazing work on Aurora and for releasing the Air pollution model !

According to the tutorial (https://microsoft.github.io/aurora/example_cams.html) ，I tested the link AuroraAirPollution model, but now there was a mistake ImportError: cannot import name 'AuroraAirPollution' from 'aurora' (/ data/Downloads/anaconda3 envs / 1 / lib/python3.10 / site - packages/aurora/set py), I tried to use PIP and conda installation are not, The version of aurora I'm using is 1.5.2. Is there a problem with my installation version or some other possible error? Looking forward to receiving your reply and help！

[wesselb]

I've just released a new version, 1.6.0, which should include the new models. Let me know if that works!

[yhchen1112]

I've just released a new version, 1.6.0, which should include the new models. Let me know if that works!

Hi，thank you very much for your reply and help. I have now successfully run the sample code and obtained the output result. Additionally, I would like to ask, I only have four 4090 cards with 24G of video memory each. How much do you suggest I set the batch size to appropriately?

[wesselb]

@yhchen1112 We typically run with batch size just one. With 24 GB GPUs, you'll likely not be able to run the model on a GPU. You'll therefore need to split the model across multiple GPUs with model parallelism. This is unfortunately not supported out of the box.

[yhchen1112]

@yhchen1112 We typically run with batch size just one. With 24 GB GPUs, you'll likely not be able to run the model on a GPU. You'll therefore need to split the model across multiple GPUs with model parallelism. This is unfortunately not supported out of the box.

@wesselb Thanks for your reply!It seems that my method of temporarily saving the predicted results to the cpu by retaining only the minimum number of prediction steps on the GPU has run successfully!
But now I have a new question. When I was using the Air Pollution model to make predictions by rollout from 2019-11-01-0:00, I found that outliers would appear after 16 days and all would diverge after 23 days. Is this normal? There can still be other prediction methods to solve it?I'm really looking forward to receiving your reply!

[wesselb]

@yhchen1112 The air pollution model has been trained for predictions up to 5 days. We have not actually attempted to roll the model out further than 5 days. It's very interesting to see that you are attempting to do this! I'm not very surprised that the model eventually diverges, given how scarce and volatile the CAMS data is.

If you would want to make predictions for more than 5 days into the future, I'd recommend some additional roll-out fine-tuning.