Transposed convolutions

In the convolutional operations we've discussed so far, the output dimensions are either equal or smaller than the input dimensions. In contrast, transposed convolutions (first proposed in Deconvolutional Networks by Matthew D. Zeiler, Dilip Krishnan, Graham W. Taylor, and Rob Fergus: https://www.matthewzeiler.com/mattzeiler/deconvolutionalnetworks.pdf) allow us to upsample the input data (their output is larger than the input). This operation is also known as deconvolution, fractionally strided convolution, or sub-pixel convolution. These names can sometimes lead to confusion. To clarify things, note that the transposed convolution is, in fact, a regular convolution with a slightly modified input slice or convolutional ...

Get Advanced Deep Learning with Python now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.