Lingxi Xie

Hi, 1. The V100 chips we used have 32GB memory. 2. During the training, around 25GB memory is occupied on each card (batch size is 1). 3. We used the...

Hi, There might be some inconsistency in the pseudocode. What I can say is: 1. The quantity of `type_of_windows` is indeed M_lat * M_pl. 2. The dimensionality of `attention` is...

Hi, Sorry for misleading you and other readers. The code was actually correct but not well organized. We integrated the normalization function into the model (e.g. PanguModel24()) as its first...

Hi, we are sorry that we could not release these contents due to the policy of our company.

Hi, thanks for the clarification. As far as I can see, the understanding is correct. Just to say, the `window_partition` function naturally returns a 4D tensor because we are dealing...

Hi, Your data are likely correct and, as far as we know, ERA5 is not updated. We checked the data and found that the discrepancy may be because we used...

Hi, the pre-trained models should receive pressure levels in the *descending* order.

We are sorry, but it is difficult for us to send out any data from our internal server. You may try to read the ONNX file instead: the mean and...

Hello, please see L170-173 of pseudocode.py.

目前的模型，必须要13层数据完整，才可以进行预报。如果缺失了部分数据，比如只有EC零场的9层，可以用再分析场补齐余下的层，这样对结果影响不大。