Hyper-Representations as Generative Models: Sampling Unseen Neural Network Weights

Code to reproduce the results from NeurIPS 2022 paper "Hyper-Representations as Generative Models: Sampling Unseen Neural Network Weights".

This repository contains code to reproduce the experiments from NeurIPS 2022 paper "Hyper-Representations as Generative Models: Sampling Unseen Neural Network Weights". The paper can be found here: NeurIPS website or arxiv.

Paper Abstract

Learning representations of neural network weights given a model zoo is an emerging and challenging area with many potential applications from model inspection, to neural architecture search or knowledge distillation. Recently, an autoencoder trained on a model zoo was able to learn a hyper-representation, which captures intrinsic and extrinsic properties of the models in the zoo. In this work, we extend hyper-representations for generative use to sample new model weights. We propose layer-wise loss normalization which we demonstrate is key to generate high-performing models and several sampling methods based on the topology of hyper-representations. The models generated using our methods are diverse, performant and capable to outperform strong baselines as evaluated on several downstream tasks: initialization, ensemble sampling and transfer learning. Our results indicate the potential of knowledge aggregation from model zoos to new models via hyper-representations thereby paving the avenue for novel research directions.

Reproducing Experiments

This repository contains code to reproduce the experiments in the paper. The package to train and sample hyper-representations can be installed with pip3 install -e .. It contains three modules:

• checkpoints_to_datasets contains dataset classes for neural network models, as well as data preprocessing and augmentation
• model_definitions contains the class definitions for all neural network architectures and trainers
• sampling_auxiliaries contains the functions to sample hyper-representations, finetune and evaluate the generated populations

Jupyter notebook examples to load pre-trained hyper-representations, train new hyper-representations or sample models are provided in examples.
To download the pretrained hyper-representations and training data data go to data and run bash download_data.sh.

Get in touch!

If there are any questions, reach out to us! We'll respond to issues here, or via email konstantin.schuerholt@unisg.ch.

Citation

If you want to cite this work, please use

@inproceedings{schurholtHyperRepresentationsGenerativeModels2022,
  title = {Hyper-Representations as Generative Models: Sampling Unseen Neural Network Weights},
  booktitle = {Thirty-Sixth Conference on Neural Information Processing Systems (NeurIPS)},
  author = {Sch{\"u}rholt, Konstantin and Knyazev, Boris and Gir{\'o}-i-Nieto, Xavier and Borth, Damian},
  year = {2022},
  month = sep,
}

Note

This project has been set up using PyScaffold 4.3.1. For details and usage information on PyScaffold see https://pyscaffold.org/.