pymimir-rgnn

Mimir-RGNN

Mimir-RGNN is a Python library that implements Relational Graph Neural Networks (R-GNN) using PyTorch and Mimir, offering a streamlined interface for building and using these networks.

API

One notable feature of Mimir-RGNN is its ability to define both input and output specifications directly during model construction. This allows for tailored configurations, as demonstrated below:

import pymimir_rgnn as rgnn

domain = ...
embedding_size = ...
num_layers = ...

config = rgnn.RelationalGraphNeuralNetworkConfig(
    domain=domain,
    input_specification=(rgnn.InputType.State, rgnn.InputType.Goal),
    output_specification=[('value', rgnn.OutputNodeType.Objects, rgnn.OutputValueType.Scalar)],
    embedding_size=embedding_size,
    num_layers=num_layers,
    message_aggregation=rgnn.AggregationFunction.HardMaximum
)

model = rgnn.RelationalGraphNeuralNetwork(config)

In this example, the configuration (RelationalGraphNeuralNetworkConfig) allows specifying input types (e.g., state, goal) and output requirements (e.g., scalar values from object embeddings). This flexibility supports various applications, including reinforcement learning (RL) and auxiliary loss functions.

Getting Started

Installation

Mimir-RGNN is available on PyPi and can be installed via pip:

pip install pymimir-rgnn

Examples

For an example, refer to the GitHub repository:

R-GNN Example Project