search

NVIDIA / modulus

Open-source deep-learning framework for building, training, and fine-tuning deep learning models using state-of-the-art Physics-ML methods
https://developer.nvidia.com/modulus
Apache License 2.0
1.05k stars 249 forks source link

Corrdiff: About the resolution of ERA5 and CWB data #725

Open bugsuse opened 3 days ago

bugsuse commented 3 days ago

It's very nice work!

I downloaded the CWB dataset and found that the spatial resolution of the ERA5 variables is entirely consistent with that of the CWB variables, with identical latitude and longitude ranges. However, both the CWB dataset overview and the corrdiff paper indicate that the resolution of ERA5 is 0.25°, and it was interpolated by a factor of 4 to facilitate training, which should result in a resolution of 0.0625° instead of the 2 km resolution. Could you clarify why the ERA5 variables were interpolated to match the resolution of the CWB variables?

zomosky commented 2 days ago

In the mini dataset (hrrr-mini), they used a funtion (/datasets/hrrrmini.py line167) to interpolated (perhaps considering the size of the dataset). In the training, the target and the low_resolution data (ERA5) needs to be a good correspondence. So they interploted the ERA5 first, i guess