RQ2

Experiment Setup

We choose the classical ResNet-110, Wide-ResNet-38, and VGGNet-19 as the subject model architectures. In addition, we train the models on MNIST, FMNIST, and CIFAR-10 datasets until the model performance converges. With a total of 9 (3*3) configurations, we train each configuration (i.e., architecture-dataset combination) 10 times. There will be 90 (9*10) models. The hyperparameters' details are as follows:

batch size: 128           learning rate: 0.01        weight_decay: 0.005          epochs: 300        initializer: he_normal

Results

ResNet-110 on CIFAR-10

ResNet-110 on FMNIST/MNIST

Wide-ResNet-38 on CIFAR-10

Wide-ResNet-38 on FMNIST/MNIST

VGGNet-19 on CIFAR-10

VGGNet-19 on FMNIST/MNIST

Code Demo

https://github.com/hnurxn/Deep-Arc/RQ2