Skip to main content

Convert ArangoDB graphs to PyG & vice-versa.

Project description

ArangoDB-PyG Adapter

build CodeQL Coverage Status Last commit

PyPI version badge Python versions badge

License Code style: black Downloads

The ArangoDB-PyG Adapter exports Graphs from ArangoDB, the multi-model database for graph & beyond, into PyTorch Geometric (PyG), a PyTorch-based Graph Neural Network library, and vice-versa.

About PyG

PyG (PyTorch Geometric) is a library built upon PyTorch to easily write and train Graph Neural Networks (GNNs) for a wide range of applications related to structured data.

It consists of various methods for deep learning on graphs and other irregular structures, also known as geometric deep learning, from a variety of published papers. In addition, it consists of easy-to-use mini-batch loaders for operating on many small and single giant graphs, multi GPU-support, DataPipe support, distributed graph learning via Quiver, a large number of common benchmark datasets (based on simple interfaces to create your own), the GraphGym experiment manager, and helpful transforms, both for learning on arbitrary graphs as well as on 3D meshes or point clouds.


Latest Release

pip install torch
pip install adbpyg-adapter

Current State

pip install torch
pip install git+


Open In Collab

Also available as an ArangoDB Lunch & Learn session on YouTube: Graph & Beyond Course: ArangoDB-PyG Adapter

import torch
import pandas
from torch_geometric.datasets import FakeHeteroDataset

from arango import ArangoClient  # Python-Arango driver

from adbpyg_adapter import ADBPyG_Adapter, ADBPyG_Controller
from adbpyg_adapter.encoders import IdentityEncoder, CategoricalEncoder

# Load some fake PyG data for demo purposes
data = FakeHeteroDataset(
    avg_num_channels=3,  # avg number of features per node
    edge_dim=2,  # number of features per edge
    num_classes=3,  # number of unique label values

# Let's assume that the ArangoDB "IMDB" dataset is imported to this endpoint
db = ArangoClient(hosts="http://localhost:8529").db("_system", username="root", password="")

adbpyg_adapter = ADBPyG_Adapter(db)

PyG to ArangoDB

Note: If the PyG graph contains _key, _v_key, or _e_key properties for any node / edge types, the adapter will assume to persist those values as ArangoDB document keys. See the Full Cycle (ArangoDB -> PyG -> ArangoDB) section below for an example.

# 1.1: PyG to ArangoDB
adb_g = adbpyg_adapter.pyg_to_arangodb("FakeData", data)

# 1.2: PyG to ArangoDB with a (completely optional) metagraph for customized adapter behaviour
def y_tensor_to_2_column_dataframe(pyg_tensor, adb_df):
    A user-defined function to create two
    ArangoDB attributes out of the 'user' label tensor

    :param dgl_tensor: The DGL Tensor containing the data
    :type dgl_tensor: torch.Tensor
    :param adb_df: The ArangoDB DataFrame to populate, whose
        size is preset to the length of **dgl_tensor**.
    :type adb_df: pandas.DataFrame

    NOTE: user-defined functions must return the modified **adb_df**
    label_map = {0: "Kiwi", 1: "Blueberry", 2: "Avocado"}

    adb_df["label_num"] = pyg_tensor.tolist()
    adb_df["label_str"] = adb_df["label_num"].map(label_map)

    return adb_df

metagraph = {
    "nodeTypes": {
        "v0": {
            "x": "features",  # 1) You can specify a string value if you want to rename your PyG data when stored in ArangoDB
            "y": y_tensor_to_2_column_dataframe,  # 2) you can specify a function for user-defined handling, as long as the function returns a Pandas DataFrame
        # 3) You can specify set of strings if you want to preserve the same PyG attribute names for the node/edge type
        "v1": {"x"} # this is equivalent to {"x": "x"}
    "edgeTypes": {
        ("v0", "e0", "v0"): {
            # 4) You can specify a list of strings for tensor dissasembly (if you know the number of node/edge features in advance)
            "edge_attr": [ "a", "b"]  

adb_g = adbpyg_adapter.pyg_to_arangodb("FakeData", data, metagraph, explicit_metagraph=False)

# 1.3: PyG to ArangoDB with the same (optional) metagraph, but with `explicit_metagraph=True`
# With `explicit_metagraph=True`, the node & edge types omitted from the metagraph will NOT be converted to ArangoDB.
# Only 'v0', 'v1' and ('v0', 'e0', 'v0') will be brought over (i.e 'v2', ('v0', 'e0', 'v1'), ... are ignored)
adb_g = adbpyg_adapter.pyg_to_arangodb("FakeData", data, metagraph, explicit_metagraph=True)

# 1.4: PyG to ArangoDB with a Custom Controller  (more user-defined behavior)
class Custom_ADBPyG_Controller(ADBPyG_Controller):
    def _prepare_pyg_node(self, pyg_node: dict, node_type: str) -> dict:
        """Optionally modify a PyG node object before it gets inserted into its designated ArangoDB collection.

        :param pyg_node: The PyG node object to (optionally) modify.
        :param node_type: The PyG Node Type of the node.
        :return: The PyG Node object
        pyg_node["foo"] = "bar"
        return pyg_node

    def _prepare_pyg_edge(self, pyg_edge: dict, edge_type: tuple) -> dict:
        """Optionally modify a PyG edge object before it gets inserted into its designated ArangoDB collection.

        :param pyg_edge: The PyG edge object to (optionally) modify.
        :param edge_type: The Edge Type of the PyG edge. Formatted
            as (from_collection, edge_collection, to_collection)
        :return: The PyG Edge object
        pyg_edge["bar"] = "foo"
        return pyg_edge

adb_g = ADBPyG_Adapter(db, Custom_ADBPyG_Controller()).pyg_to_arangodb("FakeData", data)

ArangoDB to PyG

# Start from scratch!
db.delete_graph("FakeData", drop_collections=True, ignore_missing=True)
adbpyg_adapter.pyg_to_arangodb("FakeData", data)

# 2.1: ArangoDB to PyG via Graph name (does not transfer attributes)
pyg_g = adbpyg_adapter.arangodb_graph_to_pyg("FakeData")

# 2.2: ArangoDB to PyG via Collection names (does not transfer attributes)
pyg_g = adbpyg_adapter.arangodb_collections_to_pyg("FakeData", v_cols={"v0", "v1"}, e_cols={"e0"})

# 2.3: ArangoDB to PyG via Metagraph v1 (transfer attributes "as is", meaning they are already formatted to PyG data standards)
metagraph_v1 = {
    "vertexCollections": {
        # Move the "x" & "y" ArangoDB attributes to PyG as "x" & "y" Tensors
        "v0": {"x", "y"}, # equivalent to {"x": "x", "y": "y"}
        "v1": {"v1_x": "x"}, # store the 'x' feature matrix as 'v1_x' in PyG
    "edgeCollections": {
        "e0": {"edge_attr"},
pyg_g = adbpyg_adapter.arangodb_to_pyg("FakeData", metagraph_v1)

# 2.4: ArangoDB to PyG via Metagraph v2 (transfer attributes via user-defined encoders)
# For more info on user-defined encoders in PyG, see
metagraph_v2 = {
    "vertexCollections": {
        "Movies": {
            "x": {  # Build a feature matrix from the "Action" & "Drama" document attributes
                "Action": IdentityEncoder(dtype=torch.long),
                "Drama": IdentityEncoder(dtype=torch.long),
            "y": "Comedy",
        "Users": {
            "x": {
                "Gender": CategoricalEncoder(mapping={"M": 0, "F": 1}),
                "Age": IdentityEncoder(dtype=torch.long),
    "edgeCollections": {
        "Ratings": { "edge_weight": "Rating" } # Use the 'Rating' attribute for the PyG 'edge_weight' property
pyg_g = adbpyg_adapter.arangodb_to_pyg("IMDB", metagraph_v2)

# 2.5: ArangoDB to PyG via Metagraph v3 (transfer attributes via user-defined functions)
def udf_v0_x(v0_df):
    # process v0_df here to return v0 "x" feature matrix
    # v0_df["x"] = ...
    return torch.tensor(v0_df["x"].to_list())

def udf_v1_x(v1_df):
    # process v1_df here to return v1 "x" feature matrix
    # v1_df["x"] = ...
    return torch.tensor(v1_df["x"].to_list())

metagraph_v3 = {
    "vertexCollections": {
        "v0": {
            "x": udf_v0_x,  # supports named functions
            "y": lambda df: torch.tensor(df["y"].to_list()),  # also supports lambda functions
        "v1": {"x": udf_v1_x},
    "edgeCollections": {
        "e0": {"edge_attr": (lambda df: torch.tensor(df["edge_attr"].to_list()))},
pyg_g = adbpyg_adapter.arangodb_to_pyg("FakeData", metagraph_v3)

Experimental: preserve_adb_keys

# With `preserve_adb_keys=True`, the adapter will preserve the ArangoDB vertex & edge _key values into the (newly created) PyG graph.
# Users can then re-import their PyG graph into ArangoDB using the same _key values 
pyg_g = adbpyg_adapter.arangodb_graph_to_pyg("imdb", preserve_adb_keys=True)

# pyg_g["Movies"]["_key"] --> ["1", "2", ..., "1682"]
# pyg_g["Users"]["_key"] --> ["1", "2", ..., "943"]
# pyg_g[("Users", "Ratings", "Movies")]["_key"] --> ["2732620466", ..., "2730643624"]

# Let's add a new PyG User Node by updating the _key property

# Note: Prior to the re-import, we must manually set the number of nodes in the PyG graph, since the `arangodb_graph_to_pyg` API creates featureless node data
pyg_g["Movies"].num_nodes = len(pyg_g["Movies"]["_key"]) # 1682
pyg_g["Users"].num_nodes = len(pyg_g["Users"]["_key"]) # 944 (prev. 943)

# Re-import PyG graph into ArangoDB
adbpyg_adapter.pyg_to_arangodb("imdb", pyg_g, on_duplicate="update")

Development & Testing

Prerequisite: arangorestore

  1. git clone
  2. cd pyg-adapter
  3. (create virtual environment of choice)
  4. pip install torch
  5. pip install -e .[dev]
  6. (create an ArangoDB instance with method of choice)
  7. pytest --url <> --dbName <> --username <> --password <>

Note: A pytest parameter can be omitted if the endpoint is using its default value:

def pytest_addoption(parser):
    parser.addoption("--url", action="store", default="http://localhost:8529")
    parser.addoption("--dbName", action="store", default="_system")
    parser.addoption("--username", action="store", default="root")
    parser.addoption("--password", action="store", default="")

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

adbpyg_adapter-1.1.2.tar.gz (38.2 kB view hashes)

Uploaded source

Built Distribution

adbpyg_adapter-1.1.2-py3-none-any.whl (27.4 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page