Understanding and Visualizing Generative Adversarial Network in Architectural Drawings

Generative Adversarial Network (GAN) is a model frame-work in machine learning. It’s specially used for learning and generating input and output data with similar or the same format. PIX2PIXHD is a refined version of GAN, which is designed for learning image data in pairs, and generating predicted images based on the network model. The author applied PIX2PIXHD to learning and generating architectural drawings, marking rooms with different colours automatically by computer programs. Then, to understand how this network works, the author analysed the frame of the network, and gave a detailed explanation about the three working principles of this network, convolution layer, residual network layer and deconvolution layer. Last, to visualize the network in architectural drawings, the author exported the data from each layer and each training epoch as grayscale images, finding that the features of architectural plan drawings have been learned step by step, and stored in the network as parameters. And the features in the drawings become more concise as the network goes deeper, and clearer as the training epoch increases. It might be inspiring comparing to the learning process of our human beings.