CondConv: Conditionally Parameterized Convolutions for Efficient Inference

Convolutional layers are one of the basic building blocks of modern deep networks. One fundamental assumption is that convolutional kernels be shared for all examples in a dataset. We propose conditionally convolutions (CondConv), which learn specialized convolutional for each example. Replacing normal convolutions with CondConv enables to increase the size and capacity of a network, while maintaining efficient. We demonstrate that scaling networks with CondConv improves the and inference cost trade-off of several existing convolutional network architectures on both classification and detection tasks. On classification, our CondConv approach applied to EfficientNet-B0 state-of-the-art performance of 78. 3% accuracy with only 413M-adds. Code and checkpoints for the CondConv Tensorflow layer and-EfficientNet models are available at: : //github. com/tensorflow/tpu/tree/master/models/official/efficientnet/condconv.