Ask HN: Who uses Azure ML?

#1

Post by Avalaxy » Fri, Feb 03, 2017, 11:47 AM UTC

I was checking out this video (https://www.youtube.com/watch?v=kZ04LnSjWek) on Azure Machine Learning the other day and I was very impressed. Basically you can click-and-drag all sorts of components together to create a model, test it, compare algorithms, visualize data sets and results and then publish the model as a working web service. With this tool you can set up a fairly complex machine learning model in just one hour, including 'implementing' different algorithms and comparing their results. Especially for building a proof of concept and for experimenting, I think it's extremely powerful.

I never heard anyone here about using it, so I was wondering why not? Bad experiences? Nobody knows about it? Can anyone tell me about their experiences?

Re: Ask HN: Who uses Azure ML?

#2

Post by hcrisp » Fri, Feb 03, 2017, 12:35 PM UTC

To my knowledge, it only works with dataframes, has limited algorithms, and results in non-portable models due to vendor lock-in. That said, it makes ML easier for the non-programmer, is seeing fast growth due to MS investment, and can be well integrated to custom R and Python code.

Re: Ask HN: Who uses Azure ML?

#3

Post by Avalaxy » Fri, Feb 03, 2017, 1:06 PM UTC

hcrisp wrote: post #2

To my knowledge, it only works with dataframes, has limited algorithms, and results in non-portable models due to vendor lock-in. That said, it makes ML easier for the non-programmer, is seeing fast growth due to MS investment, and can be well integrated to custom R and Python code.

> and results in non-portable models due to vendor lock-in

Yea, I was wondering about this. The free tier (https://azure.microsoft.com/en-us/pricing/details/machine-le...) says that publishing it as a production API is not available. I wonder how you can get data out of it without the API.

I'm also wondering how the costs in a real life scenario would compare to running your own software on bare metal. Maybe it would be a good trade off to experiment with Azure ML to find the winning algo combination and then implement it yourself and host it on your own server.

Re: Ask HN: Who uses Azure ML?

#4

Post by BIackSwan » Fri, Feb 03, 2017, 1:26 PM UTC

Its very good for quick prototyping. I have used it for experimentation and it speeds up your iterations by a significant amount.

From my blog post [1] - "I explored a whole bunch of options on what platform to use for training ML - Scikit Learn, Tensorflow, AWS Machine Learning and Microsoft Azure. Azure’s ease of use puts it in another league altogether - especially for people who want to just apply existing ML algorithms. Drag and drop, kicking off experiments parallely, easy to construct flows etc. allow you to move really really fast as compared to writing code.

This might become a bottleneck later as we scale or it might become more expensive to run - but the tools they provide save tons of time, especially for prototyping and experimentation. I highly recommend trying out Azure - their free tier gives a good taste of what’s possible on their platform."

[1] http://blog.doctorc.in/post/156115218607/predicting-medical-...

Re: Ask HN: Who uses Azure ML?

#5

Post by Avalaxy » Fri, Feb 03, 2017, 1:33 PM UTC

BIackSwan wrote: post #4

Its very good for quick prototyping. I have used it for experimentation and it speeds up your iterations by a significant amount. From my blog post [1] - "I explored a whole bunch of options on what platform to use for training ML - Scikit Learn, Tensorflow, AWS Machine Learning and Microsoft Azure. Azure’s ease of use puts it in another league altogether - especially for people who want to just apply existing ML alg…

Very interesting! Bookmarking your blog post for later. I'm still deciding whether I should focus on Scikit/TF/AzureML, or multiples. Any comparisons between them are welcome. Would be nice to see a cost breakdown for each of them either (as in development time and $$ to run it).

Re: Ask HN: Who uses Azure ML?

#6

Post by oduis » Fri, Feb 03, 2017, 1:40 PM UTC

I used it for analyzing some organisational customer data. It worked very well if you can convert your (offline) data to one of the supported formats.

Yes, price is an issue if you are a heavy user of web services with masses of calls, but especially for corporate environments where it is difficult to set up a machine but the amount of data to analyze is reasonable I can really recommend it.

Re: Ask HN: Who uses Azure ML?

#7

Post by joakleaf » Fri, Feb 03, 2017, 1:57 PM UTC

I have used several times. Uploaded a CSV file, thrown neural networks or random decision forests at it to see how well it worked.

It works well when you are prototyping and is quite accessible. It is also good to test various algorithms and the GUI enables you to get a good sense 'flow of data' (input, filters, algorithms, comparisons, output, etc.).

However, it becomes painful really quickly, and it is not for production!

For the particular project, I ended up using RDFs (because they showed the most promising results in Azure), and implemented RDFs by hand. After some tweaking, my own implementation performed better than the one on Azure, so I was satisfied with that. BTW. RDFs are quite easy to implement!

I think, I'll switch to TensorFlow for equivalent future prototyping. Mainly, because I can run it locally and the script (Python) support on Azure ML is terrible.

Re: Ask HN: Who uses Azure ML?

#8

Post by Avalaxy » Fri, Feb 03, 2017, 2:02 PM UTC

joakleaf wrote: post #7

I have used several times. Uploaded a CSV file, thrown neural networks or random decision forests at it to see how well it worked. It works well when you are prototyping and is quite accessible. It is also good to test various algorithms and the GUI enables you to get a good sense 'flow of data' (input, filters, algorithms, comparisons, output, etc.). However, it becomes painful really quickly, and it is not for prod…

Interesting experience, thanks!

> For the particular project, I ended up using RDFs (because they showed the most promising results in Azure), and implemented RDFs by hand. After some tweaking, my own implementation performed better than the one on Azure, so I was satisfied with that.

What language/libraries did you implement this in? Is this Python? Do you have any idea what makes it faster than Azure?

Re: Ask HN: Who uses Azure ML?

#9

Post by BIackSwan » Fri, Feb 03, 2017, 2:18 PM UTC

Avalaxy wrote: post #5

BIackSwan wrote: post #4

Its very good for quick prototyping. I have used it for experimentation and it speeds up your iterations by a significant amount. From my blog post [1] - "I explored a whole bunch of options on what platform to use for training ML - Scikit Learn, Tensorflow, AWS Machine Learning and Microsoft Azure. Azure’s ease of use puts it in another league altogether - especially for people who want to just apply existing ML alg…

Very interesting! Bookmarking your blog post for later. I'm still deciding whether I should focus on Scikit/TF/AzureML, or multiples. Any comparisons between them are welcome. Would be nice to see a cost breakdown for each of them either (as in development time and $$ to run it).

Its very contextual. So the answer is "it depends".

Azure ML will cost you since they have a non free tier in $$ whereas the other tools can scale without that. BUT for the other tools which are free in $$ you need more powerful hardware available as you start using larger datasets.

Also, the ease of use of Azure is a benefit in itself, you don't need to learn any coding library and how to use it - speeds up your iteration loops. So the dev time spent in Azure will be less than others because of how its designed - I am not so sure about using it on a large production level scale though - haven't hit that case yet.

I can't get more specific than this because I don't know what your requirements/constraints are and what sort of resources are available to you.

Re: Ask HN: Who uses Azure ML?

#10

Post by joakleaf » Fri, Feb 03, 2017, 2:22 PM UTC

Avalaxy wrote: post #8

joakleaf wrote: post #7

I have used several times. Uploaded a CSV file, thrown neural networks or random decision forests at it to see how well it worked. It works well when you are prototyping and is quite accessible. It is also good to test various algorithms and the GUI enables you to get a good sense 'flow of data' (input, filters, algorithms, comparisons, output, etc.). However, it becomes painful really quickly, and it is not for prod…

Interesting experience, thanks! > For the particular project, I ended up using RDFs (because they showed the most promising results in Azure), and implemented RDFs by hand. After some tweaking, my own implementation performed better than the one on Azure, so I was satisfied with that. What language/libraries did you implement this in? Is this Python? Do you have any idea what makes it faster than Azure?

Java, actually... But not by choice -- The client requested that the project was written in Java.

It is always hard to compare these algorithms because you have lots of parameters, and I must admit, that I did not tweak the settings on Azure.

I don't know why (or indeed if) it was faster/better results. You cannot really time the algorithms on Azure the same way you can locally. But because I ran my own implementation on a fast local machine it probably felt a lot faster -- This and then that my implementation was more specific to the overall problem.

In any case, speed when constructing the RDFs is not the main concern -- It is the quality of the model you create. Here I could add some specific tweaks for the particular problem (e.g. weighing different classes) that were not available in Azure.

Finally, using the RDFs for classification is extremely fast and straight-forward (basically you just branch through a large set of trees), and were nowhere near the biggest bottleneck in the specific project.

It just works a whole lot better locally.