PyTorch Geometric (PyG) Adapter

The PyG Adapter exports Graphs from ArangoDB into PyTorch Geometric (PyG), a PyTorch-based Graph Neural Network library, and vice-versa

PyTorch Geometric (PyG) 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.

Resources

The PyG Adapter repository  is available on Github. Check it out!

Installation

To install the latest release of the PyG Adapter, run the following command:

pip install torch
pip install adbpyg-adapter

Quickstart

The following examples show how to get started with the PyG Adapter. Check also the interactive tutorial .

import torch
import pandas
from torch_geometric.datasets import FakeHeteroDataset

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

# Connect to ArangoDB
db = ArangoClient().db()

# Instantiate the adapter
adbpyg_adapter = ADBPyG_Adapter(db)

# Create a PyG Heterogeneous Graph
data = FakeHeteroDataset(
    num_node_types=2,
    num_edge_types=3,
    avg_num_nodes=20,
    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
)[0]

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: without a  Metagraph #
#############################

adb_g = adbpyg_adapter.pyg_to_arangodb("FakeData", data)

#########################
# 1.2: with a Metagraph #
#########################

# Specifying a Metagraph provides customized adapter behaviour
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"]  
        },
    },
}

def y_tensor_to_2_column_dataframe(pyg_tensor: torch.Tensor, adb_df: pandas.DataFrame) -> pandas.DataFrame:
    """A user-defined function to create two
    ArangoDB attributes out of the 'user' label tensor

    :param pyg_tensor: The PyG Tensor containing the data
    :type pyg_tensor: torch.Tensor
    :param adb_df: The ArangoDB DataFrame to populate, whose
        size is preset to the length of **pyg_tensor**.
    :type adb_df: pandas.DataFrame
    :return: The populated ArangoDB DataFrame
    :rtype: pandas.DataFrame
    """
    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


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

#######################################################
# 1.3: with a Metagraph and `explicit_metagraph=True` #
#######################################################

# With `explicit_metagraph=True`, the node & edge types omitted from the metagraph will NOT be converted to ArangoDB.
adb_g = adbpyg_adapter.pyg_to_arangodb("FakeData", data, metagraph, explicit_metagraph=True)

########################################
# 1.4: with a custom ADBPyG Controller #
########################################

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: via Graph name #
#######################

# Due to risk of ambiguity, this method does not transfer attributes
pyg_g = adbpyg_adapter.arangodb_graph_to_pyg("FakeData")

#############################
# 2.2: via Collection names #
#############################

# Due to risk of ambiguity, this method does not transfer attributes
pyg_g = adbpyg_adapter.arangodb_collections_to_pyg("FakeData", v_cols={"v0", "v1"}, e_cols={"e0"})

######################
# 2.3: via Metagraph #
######################

# Transfers 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: via Metagraph with user-defined encoders #
#################################################

# Transforms attributes via user-defined encoders
# For more info on user-defined encoders in PyG, see https://pytorch-geometric.readthedocs.io/en/latest/notes/load_csv.html
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: via Metagraph with user-defined functions #
##################################################

# Transforms attributes via user-defined functions
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()))},
    },
}

def udf_v0_x(v0_df: pandas.DataFrame) -> torch.Tensor:
    # v0_df["x"] = ...
    return torch.tensor(v0_df["x"].to_list())


def udf_v1_x(v1_df: pandas.DataFrame) -> torch.Tensor:
    # v1_df["x"] = ...
    return torch.tensor(v1_df["x"].to_list())

pyg_g = adbpyg_adapter.arangodb_to_pyg("FakeData", metagraph_v3)