Playing with Google's open-source image classifier (TensorFlow for Poets)

Google has wrapped their deep learning image recognizer into a package you can play with fairly easily if you can handle console access to a computer ("TensorFlow for Poets"). You simply create a folder for each thing you want to recognize, and fill it with examples (~600 pics per flower in the flower classification example they provide), and run a command for 5 mins to a day (2012 laptop). Then you run another command to get the classification for a new image. The deep parts of the net remain fixed, and were trained by Google on a big database of pics - it just fine tunes the final stage to your pics.

Here are my first results - pretty shaky for my own purposes, but not bad for a start. I have different requirements than Google, e.g. if I say the photo has red or white, I mean it jumps out at you, not simply that it is there, so to get the most out of this technology I might need to get a GPU or 4 and retrain the entire network for a week or two, and manually classify way more than the 15K images I have now as input to the algorithm.