Architecture Descriptions and High Frequency Accuracy and Loss Data of Random Neural Networks Trained on Image Datasets

This dataset supports early prediction augmentation of NAS by providing high frequency NN accuracy and loss data for collections of random NNs. The NNs are generated with randomized parameter values and trained on image datsets. Each NN is trained for 40 epochs, and all generated parameter values are recorded so that the NN can be easily reconstructed. Throughout training, the NN's training loss, validation loss, and validation accuracy are calculated and recorded with high frequency--every half epoch--allowing construction of accuracy and loss curves. For each NN, the loss and accuracy data and the generated parameter values for that network are recorded in a table. This dataset of trained NNs can be used to analyze NN loss and accuracy curves and compare NN performance with various properties and parameters of the NN, enabling construction and tuning of early prediction methods for NAS.本数据集通过为随机生成的神经网络（Neural Network, NN）集合提供高频精度与损失数据，为神经架构搜索（Neural Architecture Search, NAS）的早期预测增强提供支撑。所生成的神经网络均采用随机化参数初始化，并在图像数据集上完成训练。每一个神经网络均会被训练40个训练轮次（epoch），同时所有生成的参数值都会被记录，以便后续能够便捷地复现该神经网络。在整个训练过程中，系统会以每半个训练轮次的高频率，记录该神经网络的训练损失、验证损失以及验证精度，由此可构建出精度与损失曲线。针对每一个神经网络，其损失与精度数据以及该网络的生成参数值都会被存储于数据表中。该经训练的神经网络数据集可用于分析神经网络的损失与精度曲线，并对比不同神经网络的性能与其各类属性、参数之间的关联，从而能够构建并调优用于神经架构搜索的早期预测方法。