FusionBench is a benchmark suite designed to evaluate the performance of various deep model fusion techniques. It aims to provide a comprehensive comparison of different methods on a variety of datasets and tasks.

Projects based on FusionBench and news from the community (descending order of date):

Li Shen, Anke Tang, Enneng Yang et al. Efficient and Effective Weight-Ensembling Mixture of Experts for Multi-Task Model Merging. Oct, 2024. https://github.com/EnnengYang/Efficient-WEMoE

Jinluan Yang et al. Mitigating the Backdoor Effect for Multi-Task Model Merging via Safety-Aware Subspace. Oct, 2024. http://arxiv.org/abs/2410.13910

install from PyPI:

pip install fusion-bench

or install the latest version in development from github repository

git clone https://github.com/tanganke/fusion_bench.git
cd fusion_bench

pip install -e . # install the package in editable mode

Deep model fusion is a technique that merges, ensemble, or fuse multiple deep neural networks to obtain a unified model. It can be used to improve the performance and robustness of model or to combine the strengths of different models, such as fuse multiple task-specific models to create a multi-task model. For a more detailed introduction to deep model fusion, you can refer to W. Li, 2023, 'Deep Model Fusion: A Survey'. We also provide a brief overview of deep model fusion in our documentation. In this benchmark, we evaluate the performance of different fusion methods on a variety of datasets and tasks.

The project is structured as follows:

• fusion_bench/: the main package of the benchmark.
  • method: contains the implementation of the fusion methods.

    naming convention: fusion_bench/method/{method_name}/{variant}.py contains the implementation of the specific method or its variants. For example, fusion_bench/method/regmean/clip_regmean.py contains the implementation of the RegMean algorithm for CLIP vision models.

  • modelpool: contains the implementation of the model pool, responsible for managing the models and dataset to be loaded.
  • taskpool: contains the implementation of the task pool, responsible for evaluating the performance of models returned by the algorithm.
• config/: configuration files for the benchmark. We use Hydra to manage the configurations.
  • method: configuration files for the fusion methods.

    naming convention: config/method/{method_name}/{variant}.yaml contains the configuration for the specific method or its variants.

  • modelpool: configuration files for the model pool.
  • taskpool: configuration files for the task pool.
  • model: configuration files for the models.
  • dataset: configuration files for the datasets.
• docs/: documentation for the benchmark. We use mkdocs to generate the documentation. Start the documentation server locally with mkdocs serve. The required packages can be installed with pip install -r mkdocs-requirements.txt.
• examples/: example scripts for running some of the experiments.

  naming convention: examples/{method_name}/ contains the files such as bash scripts and jupyter notebooks for the specific method.

• tests/: unit tests for the benchmark.

The fusion_bench command-line interface is a powerful tool for researchers and practitioners in the field of model fusion. It provides a streamlined way to experiment with various fusion algorithms, model combinations, and evaluation tasks. By leveraging Hydra's configuration management, fusion_bench offers flexibility in setting up experiments and reproducibility in results. The CLI's design allows for easy extension to new fusion methods, model types, and tasks, making it a versatile platform for advancing research in model fusion techniques.

Read the CLI documentation for more information.

First, create a new Python file for the algorithm in the fusion_bench/method directory. Following the naming convention, the file should be named {method_name_or_class}/{variant}.py.

from fusion_bench import BaseModelFusionAlgorithm, BaseModelPool

class DerivedModelFusionAlgorithm(BaseModelFusionAlgorithm):
    """
    An example of a derived model fusion algorithm.
    """

    # _config_mapping maps the attribution to the corresponding key in the configuration file.
    # this is optional and can be used to serialize the object to a configuration file.
    # `self.config.hyperparam_1` will be mapped to the attribute `hyperparam_attr_1`.
    _config_mapping = BaseModelFusionAlgorithm._config_mapping | {
        "hyperparam_attr_1": "hyperparam_1",
        "hyperparam_attr_2": "hyperparam_2",
    }

    def __init__(self, hyperparam_1, hyperparam_2, **kwargs):
        self.hyperparam_attr_1 = hyperparam_1
        self.hyperparam_attr_2 = hyperparam_2
        super().__init__(**kwargs)

    def run(self, modelpool: BaseModelPool):
        # modelpool is an object that responsible for managing the models and dataset to be loaded.
        # implement the fusion algorithm here.
        raise NotImplementedError(
            "DerivedModelFusionAlgorithm.run() is not implemented."
        )

A corresponding configuration file should be created to specify the class and hyperparameters of the algorithm. Here we assume the configuration file is placed at config/method/your_algorithm_config.yaml.

_target_: path_to_the_module.DerivedModelFusionAlgorithm

hyperparam_1: some_value
hyperparam_2: another_value

Use the algorithm in the FusionBench:

fusion_bench \
  method=your_algorithm_config \
  method.hyperparam_1=you_can_override_this \
  method.hyperparam_2=and_this \
  ... # other configurations

We provide a project template for quickly starting a new fusion algorithm implementation here: FusionBench Project Template.

FusionBench Command Generator is a user-friendly web interface for generating FusionBench commands based on configuration files. It provides an interactive way to select and customize FusionBench configurations, making it easier to run experiments with different settings. Read more here.

If you find this benchmark useful, please consider citing our work:

@misc{tangFusionBenchComprehensiveBenchmark2024,
  title = {{{FusionBench}}: {{A Comprehensive Benchmark}} of {{Deep Model Fusion}}},
  shorttitle = {{{FusionBench}}},
  author = {Tang, Anke and Shen, Li and Luo, Yong and Hu, Han and Du, Bo and Tao, Dacheng},
  year = {2024},
  month = jun,
  number = {arXiv:2406.03280},
  eprint = {2406.03280},
  publisher = {arXiv},
  url = {http://arxiv.org/abs/2406.03280},
  archiveprefix = {arxiv},
  langid = {english},
  keywords = {Computer Science - Artificial Intelligence,Computer Science - Computation and Language,Computer Science - Machine Learning}
}