In this chapter, we introduced the problem of semantic segmentation and implemented U-Net: a deep encoder-decoder architecture used to tackle this problem. A short introduction about the possible use cases and the challenges this problem poses has been presented, followed by an intuitive introduction of the deconvolution (transposed convolution) operation, used to build the decoder part of the architecture. Since, at the time of writing, there is not a dataset for semantic segmentation that's ready to use in TensorFlow Datasets, we took the advantage of this to show the architecture of TensorFlow Datasets and show how to implement a custom DatasetBuilder. Implementing it is straightforward, and it is something that's recommended to every TensorFlow user since it is a handy way of creating a high-efficiency data input pipeline (tf.data.Dataset). Moreover, by implementing...