Architect’s RNN / GAN

Design Process with Machine Learning and Internet

Individual Research, 2018

“We shape our tools,
and thereafter our tools shape us.” —Marshall Mcluhan

1-0 Architecture, Still a Place for Knowledge Activities?

The beginnings of CAD can be traced to the year 1957 when Dr. Patrick J. Hanratty
developed PRONTO, the first commercial numerical-control programming system.
In 1960, Ivan Sutherland MIT’s Lincoln Laboratory created SKETCHPAD, which
demonstrated the basic principles and feasibility of computer technical drawing.
As the neural network came back to a hot topic in recent years, the combination of
Artificial Intelligence and Computer Aided Design is a trend in computer-aided
techniques.
Another trend which has already deeply affected our lives is the Internet. With the Internet,
people could access numerous data and information in just a few clicks. That
provides good opportunities for designers to gather enough data to train his or her
own ML tools.
This project will implement these techniques, and adjust them to make new tools
for designers, especially architects.

1-1 Web Fetcher

This project started with a custom web fetcher aiming to collect architectural media from the Internet. The fetcher will be used to collect data for the later part of this project.
One might check the functions by following these steps. Currently, it supports ArchDaily.com
*The fetcher require bs4 and PILLOW

View the source code on GitHub.

The fetcher itself will detect the CSS and HTML tags to decompose a webpage or a whole website.
Therefore, we can reconstruct the layout of an architectural webpage, making it more suitable for
the architectural design process, such as case study among multiple similar projects.

1-2 Web Fetcher as a Designers’ Tool

The idea of char-RNN origins from Andrej Karpathy’s wonderful research in 2015. In this section, I implemented his idea with specification on architectural articles on the web.

2- Architectural Article Generator

The term“ char-rnn” is short for“ character recurrent neural network”, and is effectively a recurrent neural network trained to predict the next character given a sequence of previous characters. In this way, we can think of a char-rnn as a classification model. In the same way that we output a probability distribution over classes when doing image classification, for a char-rnn we wish to output a probability distribution over character classes, i.e., a vocabulary of characters.

2-1 Char-RNN / Preparing Data

2-2 Training

The training dataset was fetched from ArchDaily with keyword“ House”, which contains about 26MB of plain text in ASCII. Among several combinations of training parameters, the one shown on the right side is one with distinguishing convergent loss rate. The generated results of the trained model are shown on the next page.

Epoch 1 / Loss 4.6171

Wh?rCl3 1 t t)n)@ 8& = ev1YS@\ {$t@zisy? ke f”u@2 +f e oero @5 ]>% ctnH /M$e’ gwQ dognP, t44Tsate\

Epoch 300 / Loss 2.0473

Whe ring concrete dissomal a my conterifing ecpoang as cacty the to frast sutlic os and upbor. for the

Epoch 3000 / Loss 1.7543

When sleeping area with a view of a light group and a separate space, allowing a constantly monolith p

2-3 Generating Text Sequence

2-4 Char-RNN and Architectural Design Process

Initialization: Brainstorm with criteria

By selecting the content of dataset, architects may make a chatbot specialized on describing the certain topics in architectural design. This bot can then served as a partner of brainstorming at the
initial phase of building design.

Design development: BIM model continuation

BIM models can be stored in plain text format - BIM-XML. If we collect enough BIM models
in XML format, together with format checker to ensure formatting, we could build a model
continuation tools according to the existing model as input.

Presentation / press release: Article generation

Combining with object identification of photos, architects may generate an article with descriptions of a series photo by just arranging the sequence of the photo from built projects.

3-1 Architects and GAN in ML

Computer scientists have created lots of powerful NN models for various tasks. As for architects, we can combine and adjust these modules to compose a custom tool for the creative process of architectural design.

3-2 Demonstration of Custom Tool for Architects

For example, if I want a tool that can color a line art from my sketch, I’ll first use HED to generate pairs of line art and photo, then use these pairs of data to train a Pix2Pix module.

3-3 Result from Trained Model After Epoch 12

Download the pre-trained model here.

3-4 Sketch2Render for Architects

With this trained model, we can make an interface that can automatically shade the sketch from architects.

Citation
Pix2Pix | Isola, Phillip and Zhu, Jun-Yan and Zhou, Tinghui and Efros,
Alexei A. Image-to-Image Translation with Conditional Adversarial Networks. Computer Vision and Pattern Recognition (CVPR), 2017 IEEE Conference On, 2017.

HED | Saining Xie and Zhuowen Tu. Holistically-Nested Edge Detection. IEEE International Conference on Computer Vision, 2015.