Re-enabling MatrixFactorizationSimpleTrainAndPredict() #4846
Conversation
…ttps://github.com/mstfbl/machinelearning into debuggingMatrixFactorizationSimpleTrainAndPredict
|
Currently running this test 1000 times to confirm the new tolerance values of 1e-4 for Linux builds. |
|
Linux tests are passing with 1e-4. One build (Windows_x64_NetCoreApp30Release_build) is failing during clean up stage, which is not related to MatrixFactorizationSimpleTrainAndPredict(). |
| double tolerance = Math.Pow(10, -7); | ||
| double windowsTolerance = Math.Pow(10, -7); | ||
| // 1e-7 is too small for Linux, so we try 1e-4 | ||
| double linuxTolerance = Math.Pow(10, -4); | ||
| if (RuntimeInformation.IsOSPlatform(OSPlatform.Linux)) | ||
| { | ||
| // Linux case | ||
| var expectedUnixL2Error = 0.614457914950479; // Linux baseline |
There was a problem hiding this comment.
614457914950479 [](start = 44, length = 15)
I tried several run, below is my test result:
CenterOS:
Expected: 0.614457914950479 +- 1e-7
Passed: 0.6144579149504698
Failed: 0.6144409592507726
Ubuntu:
Expected: 0.614457914950479 +- 1e-7
Passed: 0.61445791495047
Failed: 0.614440959250773
builds:
https://dev.azure.com/dnceng/public/_build/results?buildId=521052&view=logs&j=887d69fe-3a72-535d-d1ef-c8cd8825499f&t=22ac88eb-9536-588d-9093-3c6614becdd0
https://dev.azure.com/dnceng/public/_build/results?buildId=521051&view=logs&j=5aa5c7df-492a-5eaf-973a-62b7b0f0ee3b&t=ffdbd604-f3e2-5332-cf61-c8dd00799b47
https://dev.azure.com/dnceng/public/_build/results?buildId=521050&view=logs&j=28859320-f5de-51e0-1fd2-7bea8c11cf7a
https://dev.azure.com/dnceng/public/_build/results?buildId=521049&view=logs&j=ac97a0ee-ee90-5a55-af17-a5a589fa545f
I'm not sure how the tolerance was set as before but seems we do have a certain failed value that like not caused by numbrical randomness, can you investigate further the root cause?
…ctorizationSimpleTrainAndPredict
|
Closing this PR as I am activating this and three other tests with a fix in another PR. |
ghost
locked as resolved and limited conversation to collaborators
Re-enabling disabled test MatrixFactorizationSimpleTrainAndPredict(). The test was failing due to tolerance value being too small for Linux builds.