CheesoidBot — InceptionV3 Applied To Cheese

So I created a Twitter bot (of sorts). Meet @cheesoidBot, a friendly bot that will helpfully identify whether a picture you send it is ‘CHEESE’ or ‘PET-RILL’.

It’s just a silly bot! But, it’s a silly bot that is backed by an InceptionV3-based ConvNet that has been fine-tuned (via transfer learning) to recognize cheese or petrol. That’s right, this bot is powered by a close-to-state-of-the-art neural net simply because I thought it would be funny.

It also, it turns out, seems to be pretty good at recognizing cheese.

Give it a try! You’ll actually have to attach the image to your tweet rather than simply shoot over a URL at poor little CheesoidBot. Then, you’ll just need to wait a few minutes and he’ll get back to you.1


Whilst setting this up on Amazon, I threw a few pictures at the model just to make sure it was running. Obviously, I used pictures of cheese, but I also gave it a picture of a person. ‘PET-RILL!’ it shrieked back at me, and I smirked with all the confidence of that guy from Mallrats:

Then I looked at the picture again to lord the human brain over the silly little machines. At which point I noticed that the person was standing between two cars…so the model had ignored the figure and made the determination based on the cars…

clever girl

  1. Yes, he’s powered by the amazing scheduling of cron. I did have a real-time version plannes with a bunch of microservices to handle aspects of interaction, all backed with Kafka, but as I sat down to implement them, I suddenly realized that I Was Trying Far Too Hard. Instead, I bashed out a simple Ruby script. It shouldn’t be too hard to switch it over to a more real-time affair in the future if necessary, but I’m not expecting crazy levels of traffic to it… [return]