Ask HN: Machine Learning Training as a Service?

#1

Post by fyquah95 » Fri, Jan 13, 2017, 10:33 AM UTC

Does such a service exist? I am thinking along the lines of:

- write a model like you normally would (in tensorflow, for eg) with some loose predefined format (eg: model must be returned by model.py#model and must be an instance of something) - git push - trains the model on a cloud service for a small fee - Provides useful graphs on training epoch / loss / error / accuracy - (Optional) Host machine learning APIs

I used Microsoft Azure at some point, but found it quite complex to setup and works well only for well-known models only. I.e: If i want to build a custom neural net, things gets messy and confusing.

It'd be useful to have such a service (or something that achieves a similar purpose). I have only a laptop that I carry around all the time, so it'd be infeasible for me to train a model on it.

EDIT: Update title

Re: Ask HN: Machine Learning Training as a Service?

#2

Post by gonber » Fri, Jan 13, 2017, 11:10 AM UTC

I once setup a genetic algorithm running in Heroku. See https://minme.herokuapp.com/. It is quite simple to build if you use Flask for the API and Python for the machine learning in the backend.

So, to answer your question, such a service could be built on top of Heroku. I'm just not sure if Heroku can provide the performance you're looking for.

Re: Ask HN: Machine Learning Training as a Service?

#3

Post by fyquah95 » Fri, Jan 13, 2017, 11:25 AM UTC

gonber wrote: post #2

I once setup a genetic algorithm running in Heroku. See https://minme.herokuapp.com/ . It is quite simple to build if you use Flask for the API and Python for the machine learning in the backend. So, to answer your question, such a service could be built on top of Heroku. I'm just not sure if Heroku can provide the performance you're looking for.

I rephrased the question to make my intend clearer. I was referring to some sort of service like heroku, but Optimization as a Service sounds like the thing i am looking for.

Re: Ask HN: Machine Learning Training as a Service?

#4

Post by brudgers » Fri, Jan 13, 2017, 3:18 PM UTC

To me, this seems like more of consultancy type business than a SaaS. The reason being the logistics of moving meaningful data sets to the location of the compute are much more difficult than moving compute to the location of the data as data set size becomes more interesting.

The other thing is that the competition for ML SaaS is Google and Microsoft and maybe Amazon. They're able to apply ML to the logistics of ML SaaS and are competing on pricing in a commodity model. That's probably not an ideal market for a small company to pursue.

Good luck.