graph_hdc implements hyperdimensional computing for graph structures.
Graphs are encoded using a message passing network that produces high dimensional
hypervectors which can be decoded again to recover structural information. The
package comes with utilities for working with PyTorch Geometric Data
objects, specialised encoders for node attributes and a small experiment suite
based on PyComex.
git clone https://github.com/the16thpythonist/graph_hdc
cd graph_hdc
python3 -m pip install .After installation the command line interface exposes the package version and the available experiments:
python3 -m graph_hdc.cli --version python3 -m graph_hdc.cli exp list
The HyperNet encoder maps graphs to hypervectors. The example below encodes
and decodes a small colour graph.
from graph_hdc.special.colors import generate_random_color_nx
from graph_hdc.special.colors import graph_dict_from_color_nx, make_color_node_encoder_map
from graph_hdc.models import HyperNet
from graph_hdc.graph import data_list_from_graph_dicts
# create a toy graph and convert it to the internal dict format
g = generate_random_color_nx(num_nodes=4, num_edges=5)
graph = graph_dict_from_color_nx(g)
# build the encoder and obtain the embedding
dim = 1000
net = HyperNet(hidden_dim=dim, depth=2,
node_encoder_map=make_color_node_encoder_map(dim))
result = net.forward_graphs([graph])[0]
embedding = result['graph_embedding']
print(embedding.shape)
# decode which node types were present
constraints = net.decode_order_zero(embedding)
print(constraints)Experiments implemented with PyComex can be listed and executed via the CLI:
python3 -m graph_hdc.cli exp list python3 -m graph_hdc.cli exp run predict_molecules__hdc
The experiments folder contains setups for comparing HDC encodings with
traditional molecular fingerprints and for first steps in graph generation.
The project depends heavily on PyComex for experiment management and
torch_geometric for graph operations.