Zhihan Gao

As the shape of the data increases, the memory costs also increase.

#### Question 1 We found that the experimental settings of the [MovingMNIST benchmark](https://paperswithcode.com/sota/video-prediction-on-moving-mnist) are not standardized. Many methods use an infinite training set (generated on the fly). To establish a...

Thanks for your interest! We have conducted preliminary empirical studies on Earthformer AR and found that the performance is far from satisfactory though it was able to generate predictions in...

Thanks for your question. You can find the details of data splits in the [`setup()`](https://github.com/amazon-science/earth-forecasting-transformer/blob/0303b37178940dd27acbcb619ffd9c381731a224/src/earthformer/datasets/enso/enso_dataloader.py#L254) function of the corresponding LightningDataModule. Also, as commented at [the beginning of the file](https://github.com/amazon-science/earth-forecasting-transformer/blob/0303b37178940dd27acbcb619ffd9c381731a224/src/earthformer/datasets/enso/enso_dataloader.py#L1), this...

Thank you for bringing up this issue. Your observation is correct regarding the default config on ICAR-ENSO, which slightly differs from that of SEVIR, N-body MNIST, and Moving MNIST. We...

Thank you for your question. Yes, `num_global_vectors=0` indicates not using global vectors. The training of Earthformer with global vectors (`num_global_vectors=8`) on ICAR-ENSO is rather unstable. To help alleviate this, I...

Thank you for your interest and your question. It may help to refer to the following piece of our implementation: https://github.com/amazon-science/earth-forecasting-transformer/blob/7732b03bdb366110563516c3502315deab4c2026/src/earthformer/cuboid_transformer/cuboid_transformer.py#L3192-L3195 The decoder typically takes as input the outputs from...

Thank you for your question. Earthformer does not directly predict the Niño3.4 index, but rather forecasts future sea surface temperature (SST) anomalies across a certain region used to calculate the...

The dataset we used has a maximum support of `out_len=26` (two years). For the experiments in our paper, we only explored a setting with `out_len=14` (one year).

This error does not appear to be related to our code. You may want to run `sudo dmesg` to view system log messages and help identify what happened.