问题
So this is a slightly insane idea that I've been kicking around. I'm thinking of building a classifier for (a specialized subset of) some web content, and then supplying it to users in the form of a chrome extension. (So: user goes to web page, clicks button on chrome extension, extension applies pre-trained model to generate prediction, presents result to user.)
But I'm cheap, and I don't want to run a server to receive data from the chrome extension, generate a prediction, and send it back to the user. It occurs to me: why not just find some way to serialize the trained model or something, and do it all on the client side.
So if I end up fitting a very simple model, like a logistic regression or something, then it would be easy to transfer to the client-side. Ultimately, a trained logistic regression is just expressible as an ordinary mathematical function mapping features to predictions, it's like one line of code. And I guess that's true of some fancier models as well (linear SVM). Under those circumstances, the chrome extension could do all the data cleaning and such, and then generating predictions is just a matter of calculating a dot product. But the same isn't true if I end up using, say, random forests or something.
Tragically, there doesn't appear to be a PMML library or anything like that for javascript.
So, I turn to the wisdom of the internet. Are there any other methods for productionizing ML models that might be usable on client-side javascript? Or are the options really just (1) run a server, (2) write a PMML library for JS, (3) there is no third option?
Thanks!
回答1:
Do you have your models available in PMML data format already?
If so, then:
- Translate PMML to Java source code. This is easy using JPMML libraries.
- Translate Java source code to JavaScript source code. This is even easier using Google Web Toolkit.
As a result, you would have a self-contained JavaScript model object that exposes model schema query and scoring API similar to the JPMML-Evaluator library.
来源：https://stackoverflow.com/questions/42449075/use-productionize-trained-machine-learning-model-in-client-side-javascript