All the kernel size of VggNet are 3x3 which fits arbitray image size. Considered the model complexity and last conv layer size, we discard the last conv block so the model finally has 13 layers. This model achieves the state-of-the-art 98.5% accuracy with 581s/epoch. (Because of the memory limit, we set the batch size 64 while the cifar10 and alexnet with 128)
