Exploring Hidden Dimensions In Accelerating Convolutional Neural Networks