GENERATIVE ADVERSARIAL NETWORK
There is a class of machine learning framework called GAN which generates new data from data. One of the most popular use cases for GAN is generating people’s faces who don’t exist.
As I was learning more about it, I got curious about how much data and epochs(The number of times your models go through the whole data) you need to generate something that is recognizable as the image you are trying to generate.
EXPERIMENTING
I got 2000 images of a cityscape from Kaggle (https://www.kaggle.com/dansbecker/cityscapes-image-pairs), trimmed off unnecessary elements from the original image, and used that to iterate 100 epochs.
RESULTS
It took me about 3~4 hours to generate the data and sadly, it is COMPLETELY UNRECOGNIZABLE. Considering overfitting, iterating more epochs probably won’t help, so I guess I need more data. (2000 images are NOWHERE CLOSE TO ENOUGH).
If I’m able to generate new cityscapes, I thought it’ll be also possible to make an architectural design proposal based on the images the client liked on Pinterest or something. That would be SO COOL!
I was actually doing another experiment generating faces using more than 20,000 images, but this is taking too long, and my computer keeps on shutting down during the process so I’ll skip this one for now.