Convolutions

Photo by Ocramnaig_o1 on UnsplashLearn the building blocks of CNNs and stop getting size mismatch errorsI often see people who want to learn how to develop deep learning applications very quickly, they learn the basics of some library like PyTorch or Tensorflow but then they haven’t really understood what’s behind those magical functions that they use so superficially. It so happens, not infrequently, that when something doesn’t work or you need to customize some function nobody knows where to start.When one is interested…

Multi-dimensional Fourier transformations in convolutional neural networksPhoto by Edz Norton on UnsplashConvolutional neural networks (CNNs) are widely spread these days. Regardless of their success, convolutions are inefficient. The sliding window requires many computations and limits the size of the kernel. At the same time, a small kernel, typically between to , limits the perceptive field and many layers are required to capture the global context of an input tensor (e.g. 2D images). The larger the image, the worse…

Understanding how the "reverse" convolution worksPhoto by Philippe D. on UnsplashTransposed convolutions are like the "ugly duckling" of the convolutional family. They are quirky and weird, but there's more to a transposed convolution than it meets the eye.You will often find layers of transposed convolutions in the decoder part of AutoEncoders, or in the generator part of GANs.Their purpose is clear: to increase the size, height and width, of their inputs.This is the opposite of what regular convolutions do: shrink the…