weather-model-graphs
is a package for creating, visualising and storing message-passing graphs for data-driven weather models.
The package is designed to use networkx.DiGraph
objects as the primary data structure for the graph representation right until the graph is to be stored on disk into a specific format.
This makes the graph generation process modular (every step outputs a networkx.DiGraph
), easy to debug (visualise the graph at any step) and allows output to different file-formats and file-structures to be easily implemented. More details are given in the background and design section below.
pdm install
cpu only:
PIP_INDEX_URL=https://download.pytorch.org/whl/cpu pdm install --group pytorch
gpu support (see https://pytorch.org/get-started/locally/#linux-pip for older versions of CUDA):
pdm install --group pytorch
The best way to understand how to use weather-model-graphs
is to look at the notebooks/constructing_the_graph.ipynb notebook, to have look at the tests in tests/ or simply to read through the source code.
In addition you can read the background and design section below to understand the design principles of weather-model-graphs
.
import numpy as np
import weather_model_graphs as wmg
# define your (x,y) grid coodinates
xy_grid = np.meshgrid(np.linspace(0, 1, 32), np.linspace(0, 1, 32), indexing='ij')
# create the full graph
graph = wmg.create.archetype.create_keisler_graph(xy=xy_grid)
# split the graph by component
graph_components = wmg.split_graph_by_edge_attribute(graph=graph, attribute='component')
# save the graph components to disk in pytorch-geometric format
for component, graph in graph_components.items():
wmg.save.to_pyg(graph=graph, name=component)
The only input the graph generation in weather-model-graphs
requires is the static (x,y)
grid coordinates of the atmospheric state as the state changes over time. These coordinates are used to create the grid nodes nodes of the graph, with a node for each (x,y)
coordinate.
In addition to grid nodes the graph also contains mesh nodes that represent the latent space of the model at a set of (x,y)
coordinates (this is in general a different set of coordinates to the grid nodes coordinates).
In summary, there are two sets of nodes in the graph:
-
grid nodes: representing the physical variables of the atmospheric state at a specific
(x,y)
coordinate -
mesh nodes: representing the latent space of the model at specific
(x,y)
coordinate
With these two sets of nodes, the graph is constructed by connecting the grid nodes to the mesh nodes and the mesh nodes to each other. The edges between the grid nodes and the mesh nodes represent the encoding of the physical variables into the latent space of the model, while the edges between the mesh nodes represent the processing of the latent space through the time evolution of the atmospheric state.
In summary, the complete message-passing graph consists of three components:
-
grid-to-mesh (
g2m
): the encoding compenent, where edges represent the encoding of physical variables into the latent space of the model -
mesh-to-mesh (
m2m
): the processing component, where edges represent information flow between nodes updating the latent presentation at mesh nodes through the time evolution of the atmospheric state -
mesh-to-grid (
m2g
): the decoding component, where edges represent the decoding of the latent space back into physical variables
Practically, the mesh-to-grid and grid-to-mesh updates can probably also encode some of the time evolution processing, in addition to the latent space encoding/decoding, unless the GNN is trained specifically as an auto-encoder using the same graph as input and output.
The key design principle of weather-model-graphs
is to work with networkx.DiGraph
objects as the primary data structure for the graph representation right until the graph is to be stored on disk into a specific format.
Using only networkx.DiGraph
objects as the intermediate representations makes it possible to
- easily modularise the whole generation process, with every step outputting a
networkx.DiGraph
object, - easily visualise the graph resulting from any step and
- easily connect graph nodes across graph components, combine graphs and split graphs based on node and edge attributes.
The graph generation in weather-model-graphs
is split into to the following steps:
-
Create the three graph components of the message-passing graph that constitute the auto-regressive atmospheric flow model, all represented by
networkx.DiGraph
objects:-
grid-to-mesh (
g2m
): the encoding compenent, where edges represent the encoding of physical variables into the latent space of the model -
mesh-to-mesh (
m2m
): the processing component, where edges represent information flow between nodes through the time evolution of the atmospheric state -
mesh-to-grid (
m2g
): the decoding component, where edges represent the decoding of the latent space back into physical variables
-
-
Combine all three graph components into a single
networkx.DiGraph
object and create a unique node identifier for each node in the combined graph. -
Split the combined graph into the three output graph components again (or more if the specific graph architecture requires it).
-
Store each of the output graph components in the desired format, for example:
-
networkx
.pickle
file: savenetworkx.DiGraph
objects usingpickle
to disk (weather_model_graphs.save.to_pickle(...)
) -
pytorch-geometric for neural-lam: edges indexes and features are stored in separate
torch.Tensor
objects serialised to disk that can then be loaded intotorch_geometric.data.Data
objects (weather_model_graphs.save.to_pyg(...)
)
-
Below, the graph generation process is visualised in weather-model-graphs
for the example given above:
graph TB
xy_grid["(x,y) grid coordinates [numpy.ndarray]"]
subgraph weather_model_graphs["weather-model-graphs"]
subgraph graph_components["Graph components"]
G_g2m["G_g2m[networkx.DiGraph]"]
G_m2m["G_m2m[networkx.DiGraph]"]
G_m2g["G_m2g[networkx.DiGraph]"]
end
G_g2m --merged into --> G_full
G_m2m --merged into --> G_full
G_m2g --merged into --> G_full
G_full["complete graph\nG_full[networkx.Digraph]"]
G_full --replace node labels with unique integer id --> G_full_int["G_int[networkx.Digraph]"]
G_full_int --split and converted into --> pyg_g2m["pyg_g2m[pyg.data.Data]"]
G_full_int --split and converted into --> pyg_m2m["pyg_m2m[pyg.data.Data]"]
G_full_int --split and converted into --> pyg_m2g["pyg_m2g[pyg.data.Data]"]
end
xy_grid --used to create node coordinates in --> G_g2m
xy_grid --used to create node coordinates in --> G_m2g
xy_grid --used to create node coordinates in --> G_m2m
subgraph stored_on_disk["stored on disk"]
subgraph hidden_graph1[ ]
pyg_g2m_file["g2m_edge_index.pt\ng2m_features.pt"]
pyg_m2m_file["m2m_edge_index.pt\nm2m_features.pt"]
pyg_m2g_file["m2g_edge_index.pt\nm2g_features.pt"]
end
end
pyg_g2m --stored in--> pyg_g2m_file
pyg_m2m --stored in--> pyg_m2m_file
pyg_m2g --stored in--> pyg_m2g_file
subgraph pyg_loaded["Loaded into model"]
subgraph hidden_graph2["(e.g. torch.nn.Module)"]
pyg_g2m_loaded["pyg_g2m[pyg.data.Data]"]
pyg_m2m_loaded["pyg_m2m[pyg.data.Data]"]
pyg_m2g_loaded["pyg_m2g[pyg.data.Data]"]
end
end
pyg_g2m_file --loaded into --> pyg_g2m_loaded
pyg_m2m_file --loaded into --> pyg_m2m_loaded
pyg_m2g_file --loaded into --> pyg_m2g_loaded
classDef title_left margin-right:250px,display:block
classDef hidden stroke-width:0px,margin-right:250px,display:block
class hidden_graph1 hidden
class hidden_graph2 hidden
class weather_model_graphs title_left
class graph_components title_left
class stored_on_disk title_left
class pyg_loaded title_left
There are a number of node and edge attributes with special meanings in weather-model-graphs
which enable the splitting and visualisation of the graph components.
pos
: the(x,y)
coordinates of the node in the gridtype
: the type of node, eithergrid
ormesh
component
: the component of the graph the edge belongs to, eitherg2m
,m2m
orm2g
level
: for multi-range mesh graphs this denotes the refinement level of mesh connection. For hierarchical graphs the different ranges of connections are split into different levels and so herelevel
also denotes the level in the hierarchy that the edge belongs to.len
: the length of the edge in the (x,y) coordinate space of the grid nodes, i.e. the distance between the two nodes in the gridvdiff
: the vector spanning between the (x,y) coordinates of the two nodesdirection
: for hierarchical graphs this denotes the direction of the edge, eitherup
,down
andsame
The splitting of the graph is done with by utilising the edge attributes, and thus it is easy to split the complete graph by either which component the edge belongs to, or by the level of the edge in the graph. This is done using the weather_model_graphs.split_graph_by_edge_attribute(...)
function.
The code layout of weather-model-graphs
is organised into submodules by the functionality they provide. The main submodules are:
weather_model_graphs
.create
.archetype:
for creating specific archetype graph
architectures (e.g. Keisler 2021, Lam et al 2023,
Oscarsson et al 2023)
.base
general interface for creating graph architectures
(here you define the g2m, m2m and m2g connectivity directly)
.mesh
for creating the mesh nodes and edges
.grid
for creating the grid nodes
.visualise
for plotting graphs, allowing for easy visualisation using any
edge or node attribute for colouring
.save
for saving the graph to specific formats (e.g. pytorch-geometric)