configuring_model.py

#  Copyright (c) Meta Platforms, Inc. and affiliates.
#
#  This source code is licensed under the license found in the
#  LICENSE file in the root directory of this source tree.
#

import torch.nn

from benchmarl.algorithms import MappoConfig
from benchmarl.environments import VmasTask
from benchmarl.experiment import Experiment, ExperimentConfig
from benchmarl.models.mlp import MlpConfig

if __name__ == "__main__":

    # Loads from "benchmarl/conf/model/layers/mlp.yaml"
    model_config = MlpConfig.get_from_yaml()

    model_config.layer_class = torch.nn.Linear  # Set the layer class
    model_config.num_cells = [32, 32]  # 2 layers with 32 neurons each
    model_config.activation_class = (
        torch.nn.ReLU
    )  # ReLU activation in between the layers

    # Some basic other configs
    task = VmasTask.BALANCE.get_from_yaml()
    algorithm_config = MappoConfig.get_from_yaml()
    experiment_config = ExperimentConfig.get_from_yaml()
    critic_model_config = MlpConfig.get_from_yaml()

    experiment = Experiment(
        task=task,
        algorithm_config=algorithm_config,
        model_config=model_config,
        critic_model_config=critic_model_config,
        seed=0,
        config=experiment_config,
    )
    experiment.run()