Web service API output incorrect !!

Question

Hi Team,

I have created scoring experiment. Inside scoring experiment, it produces a score of 0.0066 for the second record. But if I test the experiment on R using the Request/Response API and gives the same input which was given in the scoring experiment for the second record, it gives a different probability score (0.0126). However, if I test for other cases, probability score is same as in R and scoring experiment. Is there any restriction on the input in terms of size which is given to Azure ML API as an input?

Experiment URL: 

https://studio.azureml.net/Home/ViewWorkspace/8da31a3d80794c1bbd7c8e0fa7793765?#Workspaces/Experiments/Experiment/8da31a3d80794c1bbd7c8e0fa7793765.f-id.b271e1eda1904a99bef4595f511b95c0/ViewExperiment

Also, it needs to be tested using Request/Response API and not on Test dialog as per response below:

https://social.msdn.microsoft.com/Forums/en-US/310da517-03f0-4721-a30d-bb62e350f080/web-service-input-new-line-character?forum=MachineLearning

Answer 1

What is the size of the data you are posting to API? We allow up to 4 MB in the RRS request.

If you want to score a larger size of data, you can use the BES interface which allow scoring of large files in an async fashion.

thanks,

Raymond

Answer 2

The size of the input is way below 4 MB. 

But my input has new line characters, tab characters, carriage forward etc..

Question is if the input goes via CSV file in scoring experiment, it gives one probability score, but for few of the cases if I check the web API through Request/Response R code in desktop R and call the API, the probability score changes for the same input?? Input is textual code (diff file from Github pasted into one single string with use of new line character) something like below:

paste(x, collapse="\n")

Thus the input becomes one string.

Is there any different in the way Azure ML API treats special characters than desktop R??

Answer 3

Is there a need to have all those different character types in the input? 

I would strip those and send a clean input to compare the results. Otherwise, hard to tell what is causing the problem. 

Other than that, you should expect some differences in how data is handled by an API over the wire vs. running on the desktop - especially when you throw in special characters.

Thanks,

Raymond