HackMerced2019
OctoCatoPus is a Tensorflow script that trains a Generative Adversarial Network (GAN) to create cat-octopus hybrid nightmares.
For our first hackathon, we were inspired by the OctoCat drawing competition so we decided to over-engineer a solution to something that should have been straightforward. Instead of actually drawing an OctoCat like normal human beings, we handed the reigns over to our future AI overlords to see what they would come up with.
We scraped the web looking for images of cats and octopus to build a usable dataset. We then trained a proof of concept using cat pictures only, and then combined the two, using our dataset consisting of over 9000 images to train our Tensorflow backend GAN on a Google Cloud instance utilizing a Tesla v100 GPU.
After roughly 12 hours of training the proof of concept (on a laptop GTX 1060), our algorithm successfully generated images resembling cats.
After about 10 hours of training with the cat and octopus picture, we generated images resembling vaguely remembered nightmares and horrifying dreamscapes. Success!
Built With
- google-cloud
- python
- tensorflow
Log in or sign up for Devpost to join the conversation.