Image classification API is now live!

Eugene Yan spent months building a product classification system, and in late 2016 he pushed out the piece that actually gets attention: an API that looks at a photo and guesses what it is. Not text, not a title—an actual image of a t-shirt or a pair of shoes goes in, a category comes out.

The numbers are modest but respectable for the era. Top-1 validation accuracy sits at 0.76, meaning the model's best guess is right about three out of four times. Top-5 accuracy climbs to 0.974, so if you let it list five possibilities, it almost always has the right one somewhere in there. Response time stays under 300 milliseconds, which is fast enough to feel instant to a user clicking submit, even though the whole thing runs on Keras and Theano atop a bare-bones AWS server with zero GPU acceleration.

That last detail matters more than it might seem. This isn't a research demo requiring a rack of Nvidia cards. It's a scrappy, single-server setup, the kind of thing an indie builder or small e-commerce shop could actually afford to run. Yan even notes batch mode would speed things up further if a GPU were added, but the point is it doesn't need one to be useful today.

The API was trained specifically on e-commerce fashion imagery, so it shines on clothing photos and stumbles outside that lane—Yan pokes fun at this himself, uploading a t-shirt with a joke printed on it and noting, dryly, that the model has no category for "smartass t-shirt." The tool also works on mobile, letting iOS users snap a photo directly or pull one from their camera roll, which was a nice touch for a project built by one person without a dedicated mobile team.

Worth flagging: the API listed here was later discontinued, according to an update tacked onto the original post, specifically to cut cloud costs. A reminder that even a lean, GPU-free setup still costs money to keep running indefinitely, and that side projects have a way of getting quietly switched off once the novelty fades.