Speed of training proceeds

[abcsimple]

Hi, I evaluated the speed in training proceeds between VanillaNet and ResNet, vanillanet has a little bit less params and flops than resnet. However, vanillanet is slower than resnet and take much more GPU memory in the training proceed, which is strange since vanillanet has faster inference speed. Have you tried any speed comparison in the training proceeds?

[HantingChen]

The training cost of VanillaNet is higher than ResNet, since VanillaNet utilizes the deep training technique.