**

Abstract

In this case study, a ready dataset of architectural floor plans is tested for accentuating architectural floor plan features in terms of space configuration, openings locations such as windows, and room partitions. The training architecture used in this experiment is Generative Adversarial Networks or GANs. It is an architecture for training generative models as Conv2D for generating images. Exploring the structure of the latent space for a Style GAN model of Conv2D training architecture offers through its hyper-parameter a model with multiple hyper-parameter keys to observe the learning rate of the model as well as a measured evaluation in results and outcome of the trained images.

Workflow

Dataset

Model Training

Iteration 01: model training architecture  

Iteration 01: training results

Iteration 02: model training architecture

Iteration 02: training results

Iteration 03: model training architecture

Iteration 03: training results

Conclusion

Through this exploration with GAN Conv2D architecture training, few hyper-parameters play a significant impact on the image’s resolution and their features, and increasing certain hyper-parameters don not necessarily mean better training and higher resolution for images outcomes. GAN can potentially play an instrumental generative tool in architectural design. It would be interesting to carry on this exploration to develop the training architecture which in turn yields a better and higher definition of features for this dataset. In addition, speculate the application of this tool to identify floor plans not only through their spatial composition and configuration but also to investigate an integration of other machine learning architectures to correlate floor plan spatial composition to users’ behavioral patterns.

**
Autoencoders — Introduction and Implementation in TF.

 Floor Plan Dataset compiled by Stanislas Chaillou

Credits

‘COMPOSITION ’ is a project of IAAC, Institute for Advanced Architecture of Catalonia developed in the Masters of Advanced Computation for Architecture & Design 2020/21

Students: Yara Gadah &  Amal Algamdey 

Faculty: Stanislas Chaillou & Oana Taut