PyTorch Skillful Nowcasting GAN Implementation
Project description
Skillful Nowcasting with Deep Generative Model of Radar (DGMR)
Implementation of DeepMind's Skillful Nowcasting GAN Deep Generative Model of Radar (DGMR) (https://arxiv.org/abs/2104.00954) in PyTorch Lightning.
This implementation matches as much as possible the pseudocode released by DeepMind. Each of the components (Sampler, Context conditioning stack, Latent conditioning stack, Discriminator, and Generator) are normal PyTorch modules. As the model training is a bit complicated, the overall architecture is wrapped in PyTorch Lightning.
The default parameters match what is written in the paper.
Installation
Clone the repository, then run
pip install -r requirements.txt
pip install -e .
Alternatively, you can also install through pip install dgmr
Training Data
The open-sourced UK training dataset has been mirrored to HuggingFace Datasets! This should enable training the original architecture on the original data for reproducing the results from the paper. The full dataset is roughly 1TB in size, and unfortunately, streaming the data from HF Datasets doesn't seem to work, so it has to be cached locally. We have added the sample dataset as well though, which can be directly streamed from GCP without costs.
The dataset can be loaded with
from datasets import load_dataset
dataset = load_dataset("openclimatefix/nimrod-uk-1km")
For now, only the sample dataset support streaming in, as its data files are hosted on GCP, not HF, so it can be used with:
from datasets import load_dataset
dataset = load_dataset("openclimatefix/nimrod-uk-1km", "sample", streaming=True)
The authors also used MRMS US precipitation radar data as another comparison. While that dataset was not released, the MRMS data is publicly available, and we have made that data available on HuggingFace Datasets as well here. This dataset is the raw 3500x7000 contiguous US MRMS data for 2016 through May 2022, is a few hundred GBs in size, with sporadic updates to more recent data planned. This dataset is in Zarr format, and can be streamed without caching locally through
from datasets import load_dataset
dataset = load_dataset("openclimatefix/mrms", "default_sequence", streaming=True)
This steams the data with 24 timesteps per example, just like the UK DGMR dataset. To get individual MRMS frames, instead of a sequence, this can be achieved through
from datasets import load_dataset
dataset = load_dataset("openclimatefix/mrms", "default", streaming=True)
Pretrained Weights
Pretrained weights are be available through HuggingFace Hub, currently weights trained on the sample dataset. The whole DGMR model or different components can be loaded as the following:
from dgmr import DGMR, Sampler, Generator, Discriminator, LatentConditioningStack, ContextConditioningStack
model = DGMR.from_pretrained("openclimatefix/dgmr")
sampler = Sampler.from_pretrained("openclimatefix/dgmr-sampler")
discriminator = Discriminator.from_pretrained("openclimatefix/dgmr-discriminator")
latent_stack = LatentConditioningStack.from_pretrained("openclimatefix/dgmr-latent-conditioning-stack")
context_stack = ContextConditioningStack.from_pretrained("openclimatefix/dgmr-context-conditioning-stack")
generator = Generator(conditioning_stack=context_stack, latent_stack=latent_stack, sampler=sampler)
Example Usage
from dgmr import DGMR
model = DGMR(
forecast_steps=4,
input_channels=1,
output_shape=128,
latent_channels=384,
context_channels=192,
num_samples=3,
)
x = torch.rand((2, 4, 1, 128, 128))
out = model(x)
y = torch.rand((2, 4, 1, 128, 128))
loss = F.mse_loss(y, out)
loss.backward()
Citation
@article{ravuris2021skillful,
author={Suman Ravuri and Karel Lenc and Matthew Willson and Dmitry Kangin and Remi Lam and Piotr Mirowski and Megan Fitzsimons and Maria Athanassiadou and Sheleem Kashem and Sam Madge and Rachel Prudden Amol Mandhane and Aidan Clark and Andrew Brock and Karen Simonyan and Raia Hadsell and Niall Robinson Ellen Clancy and Alberto Arribas† and Shakir Mohamed},
title={Skillful Precipitation Nowcasting using Deep Generative Models of Radar},
journal={Nature},
volume={597},
pages={672--677},
year={2021}
}
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file dgmr-1.3.4.tar.gz
.
File metadata
- Download URL: dgmr-1.3.4.tar.gz
- Upload date:
- Size: 20.4 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.1 CPython/3.9.13
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | dd847f3943b2c9be6697abf4c37e5c1298e4714fec767f34160345ff5c6c49bf |
|
MD5 | ec5a32d5b146430a59fd4f156c3b4185 |
|
BLAKE2b-256 | e004ec73d31d21dee3618608f4567bcf55840c5b07366a62972ec54ce40429b0 |
File details
Details for the file dgmr-1.3.4-py3-none-any.whl
.
File metadata
- Download URL: dgmr-1.3.4-py3-none-any.whl
- Upload date:
- Size: 22.3 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.1 CPython/3.9.13
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | fff98942380de0a7bc8a81cc350eb02eb3ac2a595515935dc2b94ab82a696795 |
|
MD5 | 56539b61e3691d5c580604af705975fa |
|
BLAKE2b-256 | d1097799166cd82451ac73feab990cbd5beb1c749370973e41811ee950eb4d9d |