MetNet and MetNet-2

PyTorch Implementation of Google Research's MetNet for short term weather forecasting (https://arxiv.org/abs/2003.12140), inspired from https://github.com/tcapelle/metnet_pytorch/tree/master/metnet_pytorch

MetNet-2 (https://arxiv.org/pdf/2111.07470.pdf) is a further extension of MetNet that takes in a larger context image to predict up to 12 hours ahead, and is also implemented in PyTorch here.

Installation

Clone the repository, then run

pip install -r requirements.txt
pip install -e .

Alternatively, you can also install a usually older version through pip install metnet

Please ensure that you're using Python version 3.9 or above.

Data

While the exact training data used for both MetNet and MetNet-2 haven't been released, the papers do go into some detail as to the inputs, which were GOES-16 and MRMS precipitation data, as well as the time period covered. We will be making those splits available, as well as a larger dataset that covers a longer time period, with HuggingFace Datasets! Note: The dataset is not available yet, we are still processing data!

from datasets import load_dataset

dataset = load_dataset("openclimatefix/goes-mrms")

This uses the publicly avaiilable GOES-16 data and the MRMS archive to create a similar set of data to train and test on, with various other splits available as well.

Pretrained Weights

Pretrained model weights for MetNet and MetNet-2 have not been publicly released, and there is some difficulty in reproducing their training. We release weights for both MetNet and MetNet-2 trained on cloud mask and satellite imagery data with the same parameters as detailed in the papers on HuggingFace Hub for MetNet and MetNet-2. These weights can be downloaded and used using:

from metnet import MetNet, MetNet2
model = MetNet().from_pretrained("openclimatefix/metnet")
model = MetNet2().from_pretrained("openclimatefix/metnet-2")

Example Usage

MetNet can be used with:

from metnet import MetNet
import torch
import torch.nn.functional as F

model = MetNet(
        hidden_dim=32,
        forecast_steps=24,
        input_channels=16,
        output_channels=12,
        sat_channels=12,
        input_size=32,
        )
# MetNet expects original HxW to be 4x the input size
x = torch.randn((2, 12, 16, 128, 128))
out = []
for lead_time in range(24):
        out.append(model(x, lead_time))
out = torch.stack(out, dim=1)
# MetNet creates predictions for the center 1/4th
y = torch.randn((2, 24, 12, 8, 8))
F.mse_loss(out, y).backward()

And MetNet-2 with:

from metnet import MetNet2
import torch
import torch.nn.functional as F

model = MetNet2(
        forecast_steps=8,
        input_size=64,
        num_input_timesteps=6,
        upsampler_channels=128,
        lstm_channels=32,
        encoder_channels=64,
        center_crop_size=16,
        )
# MetNet expects original HxW to be 4x the input size
x = torch.randn((2, 6, 12, 256, 256))
out = []
for lead_time in range(8):
        out.append(model(x, lead_time))
out = torch.stack(out, dim=1)
y = torch.rand((2,8,12,64,64))
F.mse_loss(out, y).backward()