Phase Transition Model Zoo
收藏资源简介:
Phase Transition Model Zoo是一个包含约2.5k个独特神经网络的模型动物园,这些模型涵盖了11K到900M参数的范围,超过60k个检查点。该数据集涵盖了计算机视觉、自然语言处理和科学机器学习等多个领域,旨在为权重空间学习(WSL)和更广泛的应用提供资源。数据集中的模型具有不同的架构、大小和训练数据集,覆盖了所有已知的模型阶段。每个模型都计算了损失景观指标,并通过多个训练周期的检查点进行验证,以全面覆盖所有阶段。
Phase Transition Model Zoo is a model zoo containing approximately 2.5k unique neural networks, spanning a parameter scale from 11K to 900M and including over 60k checkpoints. This dataset covers multiple research fields including computer vision, natural language processing, and scientific machine learning, aiming to provide resources for weight space learning (WSL) and broader applications. The models in this dataset feature diverse architectures, sizes, and training datasets, covering all known model phases. Loss landscape metrics have been calculated for each model, and validation is conducted using checkpoints across multiple training epochs to achieve comprehensive coverage of all phases.




