Pangu: Wong surface - upper air concatenation order?

[mikael10j]

Hello,

It seems that there is a typo in the order of the concatenation between surface and upper air tokens in Pangu and Pangu lite. In pangu.py, line 427 (or 546):
x = torch.concat([surface.unsqueeze(2), upper_air], dim=2)
However, in the Pangu lite example, the upper air pressure levels are organized in ascending order, from 50 to 1000 because of pLevels = sorted(list(upper_air_mean.keys())) in WeatherLearn/tree/master/examples/pangu_lite/data_utils.py.

So this way, attention is computed between surface and the highest atmospheric level, which is not desirable from a physical point of view.

Am I missing something?
Best,

[lizhuoq]

@mikael10j Thank you for your issue. Please refer to the official pseudocode