Hello,
might be a silly question, but for running Cifar10 with a built in model, say Resnet34, isn’t it weird to have to add padding of 96 to each side to fulfill the requirement of 224X224? Will the model end up good enough?
Hi, I think that a better option is to change the first conv layer stride to 1 and the last pooling layer according to its input. You can check this project here for many examples on cifar10: