There are now a bunch of off-the-shelf tools for training artistic style transfer models and thousands of open source implementations. Most use a variation of the network architecture described by Johnson et al to perform fast, feed-forward stylization.
As a result, the majority of the style transfer models you find are the same size: 7MB. That’s not an unreasonably large asset to add to your application, but it’s also not insignificant.
Research suggests that neural networks are often way larger than they need to be—that many of the millions of weights they contain are insignificant and needlessly precise. So I wondered: What’s the smallest model I can create that still reliably performs style transfer?
Continue reading “Creating a 17 KB style transfer model with layer pruning and quantization”