locked
ML Pricing RRS feed

  • Question

  • Hello,

    I'm really new of ML, but before getting involved need to understand better the pricing model.

    Let's assume i have like the Amazon Sentiment Sample provided 1 million of reviews to train my model. And it took 10 hours, i will need to pay about 3.8 USD.

    Now let's go to prediction if i want to predict the sentiment of another 100K reviews, besides the hours prediction, will it going to cost me about  18K USD? 

    Is it correct?

    Thank you so much


    Monday, October 13, 2014 11:09 AM

Answers

  • Hi!

    Not quite, let me try to explain a bit more clearly using your example:

    Azure ML currently has two general means of execution. The Studio/Experimentation/Canvas Editor environment that is executed when you hit Run. This is the Experimentation hour pricing, and you calculated it correctly at $3.80

    Once you decide you have a graph you like and publish it as a web service, two pricing models appear. One is conceptually the same as the Experimentation one above - actually running your model takes time, and it is now charged at $0.75 / hour. Note that training a model is computationally expensive. Most published workflows do not usually take much time to complete, and is unlikely to run for 10 hours again.

    The per prediction cost is determined by # of requests to the prediction API - if you now want to predict 100,000 new reviews, you would be charged 100,000 * ($0.18 / 1000 predictions) = $18.

    I hope that helps!

    AK

    • Proposed as answer by neerajkh_MSFT Tuesday, October 14, 2014 2:42 AM
    • Marked as answer by neerajkh_MSFT Sunday, June 28, 2015 4:34 PM
    Monday, October 13, 2014 6:47 PM
  • You should use Batch mode - https://azure.microsoft.com/en-us/documentation/articles/machine-learning-consume-web-services/. Pricing for batch mode is explained in the pricing FAQ here - http://azure.microsoft.com/en-us/pricing/details/machine-learning/#

    • Marked as answer by neerajkh_MSFT Sunday, June 28, 2015 4:37 PM
    Sunday, June 28, 2015 4:37 PM

All replies

  • Hi!

    Not quite, let me try to explain a bit more clearly using your example:

    Azure ML currently has two general means of execution. The Studio/Experimentation/Canvas Editor environment that is executed when you hit Run. This is the Experimentation hour pricing, and you calculated it correctly at $3.80

    Once you decide you have a graph you like and publish it as a web service, two pricing models appear. One is conceptually the same as the Experimentation one above - actually running your model takes time, and it is now charged at $0.75 / hour. Note that training a model is computationally expensive. Most published workflows do not usually take much time to complete, and is unlikely to run for 10 hours again.

    The per prediction cost is determined by # of requests to the prediction API - if you now want to predict 100,000 new reviews, you would be charged 100,000 * ($0.18 / 1000 predictions) = $18.

    I hope that helps!

    AK

    • Proposed as answer by neerajkh_MSFT Tuesday, October 14, 2014 2:42 AM
    • Marked as answer by neerajkh_MSFT Sunday, June 28, 2015 4:34 PM
    Monday, October 13, 2014 6:47 PM
  • Hi AK

    The prediction cost is machine time? I could assume that the meter does not run when the web service is not being interrogated?

    Massimo

    Wednesday, October 15, 2014 11:14 AM
  • Hi Massimo!

    All our charges are use-based only, so that is a correct assumption. Both the Experimentation and Prediction hourly charges are metered on VM time usage during actual execution. If you don't use any compute resources, you will not be billed

    AK

    Wednesday, October 15, 2014 3:10 PM
  • Hello

    I have a similar question, but I am not sure if I understand it correctly.

    Lets say I want to train every month a model and then do every week 1.000.000 predictions (rows) with it.

    For me it is not clear if I should use the experiment mode or the webservice. It seems that a webservice is much more expensive but not necessary for my concern.

    Acoording to the pricing page http://azure.microsoft.com/en-us/pricing/details/machine-learning/

    the costs are the following (with standard tier and without support):

    experiment mode: €7.4396 (seat)+ €0.7447/Studio Experiment Hour * (hours that I predict+hours for training the model)~=30€

    webservice mode: 7.4396/ Seat/ Month + €0.7447/Studio Experiment Hour * hours for training the model + €1.4894/Production API Compute Hour * hours for predictions + €0.3724/1,000 Production API Transactions * 4.000.000 Predictions ~=1500€

    Hope these options make sense for you and you can commit these calculations.

    Best Regards

    Wednesday, April 22, 2015 8:39 AM
  • Correct
    Sunday, June 28, 2015 4:35 PM
  • You should use Batch mode - https://azure.microsoft.com/en-us/documentation/articles/machine-learning-consume-web-services/. Pricing for batch mode is explained in the pricing FAQ here - http://azure.microsoft.com/en-us/pricing/details/machine-learning/#

    • Marked as answer by neerajkh_MSFT Sunday, June 28, 2015 4:37 PM
    Sunday, June 28, 2015 4:37 PM