CrossedWires

CrossedWires 是一个包含模型和超参数的动态数据集，它揭示了两个流行的深度学习框架之间的语义差异：PyTorch 和 Tensorflow。 CrossedWires 数据集目前包含在 CIFAR10 图像上训练的模型，这些模型使用三种不同的计算机视觉架构：VGG16、ResNet50 和 DenseNet121，跨越大型超参数空间。使用超参数优化，三个模型中的每一个都在 HyperSpace 搜索算法建议的 400 组超参数上进行了训练。 CrossedWires 数据集包括 PyTorch 和 Tensforflow 模型，其在语法等效模型和相同超参数选择上的测试精度差异为 0.681。这里展示的 340 GB 数据集和基准包括所有 1200 个超参数选择的性能统计数据、训练曲线和模型权重，总共有 2400 个模型。 CrossedWires 数据集提供了一个机会来研究流行的深度学习框架中的句法等效模型之间的语义差异。CrossedWires is a dynamic dataset encompassing models and hyperparameters, which reveals the semantic discrepancies between two popular deep learning frameworks: PyTorch and TensorFlow. The CrossedWires dataset currently includes models trained on CIFAR10 images, built with three distinct computer vision architectures: VGG16, ResNet50, and DenseNet121, spanning a large hyperparameter space. Via hyperparameter optimization, each of the three models was trained on 400 sets of hyperparameters proposed by the HyperSpace search algorithm. The CrossedWires dataset includes both PyTorch and TensorFlow models, with a test accuracy difference of 0.681 between syntactically equivalent models under identical hyperparameter selections. The 340 GB dataset and benchmarks presented here encompass performance statistics, training curves, and model weights for all 1200 hyperparameter selections, totaling 2400 models in all. The CrossedWires dataset provides an opportunity to investigate semantic discrepancies between syntactically equivalent models across popular deep learning frameworks.

搜集汇总

背景与挑战

背景概述

CrossedWires是一个动态数据集，旨在研究PyTorch和TensorFlow两个深度学习框架在语法等效模型间的语义差异。它包含基于CIFAR10图像、使用VGG16、ResNet50和DenseNet121架构训练的2400个模型，覆盖了大规模超参数空间，并提供了性能统计数据、训练曲线和模型权重。