NCI WeatherBench Notebooks

This repository contains notebooks for NCI WeatherBench, which is a Deep learning benchmark developed at NCI, Australia. It is created using the Pangeo data specification (https://github.com/pangeo-data/WeatherBench); however, the data spans from 1959 to 2022, which is 50% more compared to that of the original one. It is about 10TB in size and was created completely using the NCI high-performance compute and disk resources, to the best of our knowledge it is first the Australian Deep learning benchmark of this size.
Details about the benchmark can be found here: https://geonetwork.nci.org.au/geonetwork/srv/eng/catalog.search#/metadata/f8295_5164_0873_0706

How to use

0. Before you start. You must be a member of the following NCI projects

   wb00
   dk92
   vp91

Additionally, you need a project that has enough resources to run V100 GPUs for several hours. One can apply to join NCI projects here: https://my.nci.org.au/

1. Download the notebooks in any suitable location on Gadi

cd /to/a/Gadi/location
git clone https://github.com/nci/NCI_WeatherBench

2. Launch an NCI ARE (https://are.nci.org.au) Jupyter instance with the following parameters.

Walltime (hours): <As required>
Queue: gpuvolta
Compute Size: 1gpu (minimum or custom for more memory)
Project: <Choose one that has enough resources>
Storage: gdata/dk92+gdata/z00+gdata/wb00+scratch/vp91+<All other storage that you need>

Module directories: /g/data/dk92/apps/Modules/modulefiles
Modules: NCI-ai-ml/23.10 NCI_weatherbench/2024.03.21 
Jobfs size: 200GB

3. From the ARE instance, navigate to the notebook's location and run them individually.

Notebooks

There are nine notebooks in the repository. The first three are baseline models, the next two are Keras/Tensorflow models, the next two are Pytorch implementations of the two previous notebooks, and the last two are for result visualization.  

• 1a.climatology-persistence.ipynb: Contain two baseline prediction methods: persistance, and climatology.

• 1b.climatology-persistence.ipynb: Same as the previous notebook but uses higher resolutions, 2.8125 and 1.40625. (Optional)

• 2.linear-regression-baseline.ipynb: Another baseline model, linear regression is used.

• 3a.cnn-example.ipynb: A CNN model training using Keras.

• 3b.cnn-example.ipynb: Similar to the previous CNN model, however, trains three different models, a 3-day, 5-day, and 6-hour iterative prediction model.

• 3c.cnn-example.ipynb: Pytorch implementation of the 3a.cnn-example.ipynb notebook

• 3d.cnn-example.ipynb: Pytorch implementation of the 3b.cnn-example.ipynb notebook

• 4a.evaluation-checkpoint.ipynb: Creates and saves prediction and calculates accuracy for all previous models (persistence, climatology, linear regression, and CNN)

• 4b.evaluation-checkpoint.ipynb: Displays the results produced in the previous notebook.

Note

• Depending on the code, each notebook may take from 30 minutes to 2 hours to run.
• When a notebook is finished, reset it to free the memory before running the next one. Otherwise, you may run out of memory.
• If a notebook runs out of memory, the system automatically restarts it. If you find this is happening to you, then simply start a bigger/custom ARE instance with more memory.
• All created models and predictions are stored in the following directory: /scratch/vp91/<USER>/NCI-Weatherbench/. The vp91 is a training project and is cleaned at regular intervals. If you want to save your data then copy it to a safe location. Alternatively, one can set a different location (PREDDIR variable) at the top of each notebook.