Python Deep Learning - Second Edition

Convolutional layers in deep learning libraries

PyTorch, Keras, and TensorFlow have out of the gate support for 1D, 2D, and 3D standard, and depthwise convolutions. The inputs and outputs of the convolution operation are tensors. A 1D convolution would have 3D input and output tensors. Their axes can be in either NCW or NWC order, where:

• N stands for the index of the sample in the mini-batch
• C stands for the index of the depth slice in the volume
• W stands for the content of the slice

In the same way, a 2D convolution will be represented by NCHW or NHWC ordered tensors, where H and W are the height and width of the slices. A 3D convolution will have NCDHW or NDHWC order, where D stands for the depth of the slice.