Colorizing Moby-Dick

[altsoph]

This is an old idea of mine, I wanted to try it for several years, maybe this time I'll do it

[altsoph]

Gonna post details tomorrow

[altsoph]

5 years ago, I made an pet-project for a color2color_name generation.

As a side artifact I had a 10K+ colors with names dataset, so I used it for a reversed task -- taken a fasttext vector for the word as an input, predict the RGB (or HSL, to be precise) components of it's color.

[altsoph]

Suddenly, it worked pretty well, so the majority of the color-words were colored right.

Also, I've tried to train another classifier head to predict, if the word have a color information at all. But still it thinks some non-color words have some significant colorization, for example:

• words like ocean, sky, pool, sea, water tend to have blue or cyan color;
• night, dark, sleep are dark gray or dark blue;
• burnt, wine, flame, fire are red.

It colors the random words in some strange ways as well :) For example, frost and frozen have red bias somewhy.

[altsoph]

So now I want to use this classifier to colorize each possible word in Moby-Dick.

[altsoph]

Here is an example snippet of text:

[altsoph]

It is also possible to calculate the flow of intensity of different colors through the chapters.

[altsoph]

I used the python-notebook to create the colorized html, then the wkhtmltopdf tool to convert it to PDF.
The full result is here.

The repo: https://github.com/altsoph/MDC_nanogenmo2022

[lizadaly]

Such an original idea, I love it!