data

[sunshine795]

Hello, can you provide some data（'../../../../../../mnt/data/scheepensd94dm/'）?

[dscheepens]

Hi, best if you download the data yourself from the Copernicus Climate Data Store. I've added the file grib_to_numpy.py which includes the link to the data and transforms from .grib to .npy. Don't forget to change the 'root' to your root (also in main.py).

[sunshine795]

Thank you very much. I'll try it on

[sunshine795]

Hello, my Windows system cannot install pygrib library, can I consider using NetCDF data format?

[dscheepens]

Do you have pip installed and tried:
pip install pyproj
pip install pygrib
?
The library should work for windows.

[sunshine795]

Hello, I have downloaded the data to the required directory, how should I train it next? For example, what is the order in which files are run?

[dscheepens]

Hi, just set up in main.py which parameters you want to use (e.g. loss function, number of layers etc.) and then run: python3 main.py

Make sure you have all the relevant packages installed

[dscheepens]

Please let me know if there are any other issues. The code still needs a bit of curating to be sure.

[sunshine795]

Hi,in file "loss_functions.py", "ser_t = ((preds[indices]-targets[indices])**2).mean()"position error is "IndexError: too many indices for tensor of dimension 3".

How to solve it?

[dscheepens]

Hi, is your data of dimension (N, 12, 64, 64) where N is the number of samples?

You can also just use the standard MSE loss if you just want to get the code running. We tested the SERA loss in our paper but you can of course use any loss.