ML.NET-242: FastTreeRanking per-iteration loss metrics are empty #289
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
TomFinley
reviewed
Jun 2, 2018
| @@ -343,6 +343,9 @@ protected override void CustomizedTrainingIteration(RegressionTree tree) | |||
| /// <returns>standard test for the dataset</returns> | |||
| private Test CreateStandardTest(Dataset dataset) | |||
| { | |||
| if (dataset.MaxDcg.Length == 0) | |||
There was a problem hiding this comment.
dataset.MaxDcg.Length [](start = 16, length = 21)
dataset.MaxDcg.Length [](start = 16, length = 21)
Please replace dataset.MaxDcg.Length with Utils.Size(dataset.MaxDcg). Under certain codepaths this array may be null. :P #Resolved
When training a FastTreeRanker using the `testFrequency` parameter, it is expected that NDCG is prented every testFrequency iterations. However, instead of NDCG, only empty strings are printed. The root cause was that the MaxDCG property of the dataset was never calculated, so the NDCG calculation is aborted, leaving an empty string as a result. This PR fixes the problem by computing the MaxDCG for the dataset when the Tests are defined (so that if the tests are not defined, the MaxDCG will never be calculated). Closes dotnet#242
rogancarr
force-pushed
the
242_FastTreeRanking_Per-Iteration-Loss
branch
from
9e737a8
to
79a437e
Compare
|
Thanks for the review @TomFinley ! I updated the PR to use |
TomFinley
approved these changes
Jun 4, 2018
eerhardt
pushed a commit
that referenced
this pull request
Jun 27, 2018
* Bump master to v0.3 (#269) * RocketEngine fix for selecting top learners (#270) * Changes to RocketEngine to fix take top k logic. * Add namespace information to allow file to reference correct version of Formatting object. * small code cleanup (#271) * Preparation for syncing sources with internal repo (#275) * make class partial so I can add constuctor in separate file. add constructros for testing * formatting * Changes to use evaluator metrics names in PipelineSweeperSupportedMetrics. Made the private const strings in two classes public. (#276) * add missing subcomponents to sweepers (#278) * add missing subcomponents * right one * more cleanup * remove lotus references. (#252) * Random seed and concurrency for tests (#277) * first attempt * add comments * specify seed for random. make constructor internal. * Fix SupportedMetric.ByName() method (#280) * Fix for SupportedMetric.ByName() method. Include new unit test for function. * Fix for SupportedMetric.ByName() method. Include new unit test for function. * Fix for SupportedMetric.ByName() method. Include new unit test for function. * Removed unnecessary field filter, per review comment. * ML.NET-242: FastTreeRanking per-iteration loss metrics are empty (#289) When training a FastTreeRanker using the `testFrequency` parameter, it is expected that NDCG is prented every testFrequency iterations. However, instead of NDCG, only empty strings are printed. The root cause was that the MaxDCG property of the dataset was never calculated, so the NDCG calculation is aborted, leaving an empty string as a result. This PR fixes the problem by computing the MaxDCG for the dataset when the Tests are defined (so that if the tests are not defined, the MaxDCG will never be calculated). Closes #242 * Fixed typo in the method summary (#296) * Remove stale line of code from test. (#297) * Update release notes link to use aka.ms. (#294) Our release notes link is broken because the `Documentation` was renamed to `docs`. Fix this for the future to use a redirection link. * Add release notes for ML.NET 0.2 (#301) * Add release notes for ML.NET 0.2 * Adding release note about TextLoader changes and additional issue/PR references * Addressing comments: fixing typos, changing formatting, and adding references * Get the cross validation macro to work with non-default column names (#291) * Add label/grou/weight column name arguments to CV and train-test macros * Fix unit test. * Merge. * Update CSharp API. * Fix EntryPointCatalog test. * Address PR comments. * update sample in README.MD with 0.2 features. (#304) * update sample with new text loader API. * update with 0.2 stuff. * OVA should respect normalization in underlying learner (#310) * Respect normalization in OVA. * some cleanup * fix copypaste issues * Export to ONNX and cross-platform command-line tool to script ML.NET training and inference (#248) * Export to ONNX and Maml cross-platform executable. * Add Cluster evaluator (#316) * Add Cluster evaluator * fix copypaste * address comments * formatting * Fixes locale dependent test output comparisons (#109) The tests do not pass on systems with locale other than en-US. The error happens since the results are written to files and the contents of the files are compared to set of correct results produced under en-US locale. The fix is to imbue en-US culture to the test thread so that results will be output in format that is comparable with the test format. This patch fixes only tests, but do not guarantee calculation will be correct in production systems using a locale different than en-US. In particular, there can be problems in reading data and then conversing data from characters to numeric format. Fixes #74 * Add PartitionedFileLoader (#61) * Remove unexisting project from solution (#335) * GetSummaryDataView/Row implementation for Pca and Linear Predictors (#185) * Implement `ICanGetSummaryAsIDataView` on `PcaPredictor` class * Implement `ICanGetSummaryAsIRow` on `LinearPredictor` class * Disable ordinary least squares by removing the entry point (#286) * Disable ols by temporarily removing the entry point. It may be added again once we figure out how to ship MKL as part of this project. * add append function to pipeline (#284) Add `Append` function to pipeline for more fluent API than that allowed by `Add` * Removed field/column name checking of input type in TextLoader. (#327) * fix namespace issue in CSharpGenerator and some refactoring (#339) fix namespace issue and refactoring * Using named-tuple in OneToOneTransforms' constructor to make API more readable. (#324) * Minor formatting in CollectionDataSourceTests.cs (#348) * Create CalibratedPredictor instead of SchemaBindableCalibratedPredictor (#338) `CalibratorUtils.TrainCalibrator` and `TrainCalibratorIfNeeded` now creates `CalibratedPredictor` instead of `SchemaBindableCalibratedPredictor` whenever the predictor implements `IValueMapper`. * Remove reference and dependency on System.ValueTuple (#351) * Add link to samples (#355) * Use HideEnumValueAttribute for both manifest and C# API generation. (#356) * Use HideEnumValueAttribute for both manifest and C# API generation. * Unhide NAReplaceTransform.ReplacementKind.SpecifiedValue. This may require some other PR to resolve the corresponding issues. * Move the NuGet package build files into a TFM specific directory. (#370) When installing Microsoft.ML on an unsupported framework (like net452), it is currently getting installed successfully. However, users should be getting an error stating that net452 is not supported by this package. The cause is the build files exist for any TFM, which NuGet interprets as this package supports any TFM. Moving the build files to be consistent with the 'lib' folder support. Fix #357 * `Stream` subclasses now have `Close` call `base.Close` to ensure disposal. (#369) * Subclasses of `Stream` now have `Close` call `base.Close` to ensure disposal. * Add DeleteOnClose to File opening. * Remove explicit delete of file. * Remove explicit close of substream. * Since no longer deleting explicitly, no longer need `_overflowPath` member. * Return distinct array of ParameterSet when ProposeSweep is called (#368) * Changed List to HashSet to ensure that there are no duplicates * Update fast tree argument help text (#372) * Update fast tree argument help text * Update wording * Update API to fix test * Update core manifest JSON to update help text * Combine multiple tree ensemble models into a single tree ensemble (#364) * Add a way to create a single tree ensemble model from multiple tree ensemble models. * Address PR comments, and fix bugs in serializing/deserializing RegressionTrees. * Address PR comments. * add pipelineitem for Ova (#363) add pipelineitem for Ova * Fix CV macro to output the warnings data view properly. (#385) * Link to an example on using converting ML.NET model to ONNX. (#386) * Adding documentation about entry points, and entry points graphs: EntryPoints.md and GraphRunner.md (#295) * Adding EntryPoints.md and GraphRunner.md * addressing PR feedback * Updating the title of the GraphRunner.md file * adressing Tom's feedback * adressing feedback * code formatting for class names * Addressing Gal's comments * Adding an example of an entry point. Fixing casing on ML.NET * fixing link * Adding LDA Transform (#377) * Revert to using the native code (#413) Corrects an unintentional "typo" in FastTreeRanking.cs where there was mistakenly a USE_FASTTREENATIVE2 instead of USE_FASTTREENATIVE. This resulted in some obscure hidden ranking options (distance weighting, normalize query lambdas, and a few others) being unavailable. These are important for some applications. * LightGBM (#392) * LightGBM and test. * add test baselines and nuget source for lightGBM binaries. * Add entrypoint for lightGBM. * add unsafe flag for release build. * update nuget version. * make lightgbm test single threaded. * install gcc on OS machines to resolve dependencies on openmp thatis needed by lightgbm native code. * PR comments. Leave BREW and GCC in bash script to verify macOS tests work. * remove brew and gcc from build script. * PR feedback. * disable test on macOS. * disable test on macOS. * PR feedback. * Adding Factorization Machines (#383) * Adding Factorization Machines * ONNX API documentation. (#419) * ONNX API documentation. * Bring ensembles into codebase (#379) Introduce Ensemble codebase * enable macOS tests for LightGBM. (#422) * Create a shorter temp file name for model loading. (#397) Create a shorter temp file name for model loading, as well as remove the potential for a race condition among multiple openings by using the creation of a lock file. * removing extraneous character that broke the linux build, and with it unecessary cmake version requirement (#425) * EvaluatorUtils to handle label column of type key without text key values (#394) * Fix EvaluatorUtils to handle label column of type key without text key values. * Removing non source files from solution (#362)
eerhardt
pushed a commit
to eerhardt/machinelearning
that referenced
this pull request
Jul 27, 2018
…net#289) When training a FastTreeRanker using the `testFrequency` parameter, it is expected that NDCG is prented every testFrequency iterations. However, instead of NDCG, only empty strings are printed. The root cause was that the MaxDCG property of the dataset was never calculated, so the NDCG calculation is aborted, leaving an empty string as a result. This PR fixes the problem by computing the MaxDCG for the dataset when the Tests are defined (so that if the tests are not defined, the MaxDCG will never be calculated). Closes dotnet#242
ghost
locked as resolved and limited conversation to collaborators
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When training a FastTreeRanker using the
testFrequencyparameter, it is expected that NDCG is printed every testFrequency iterations. However, instead of NDCG, only empty strings are printed.The root cause was that the
MaxDCGproperty of the dataset was never calculated, so the NDCG calculation is aborted, leaving an empty string as a result.This PR fixes the problem by computing the
MaxDCGfor the dataset when the Tests are defined (so that if the tests are not defined, theMaxDCGwill never be calculated). Here, thetruncationLevelof theMaxDCGcalculation is hardcoded to 10, which is a nice round number, and within the range of theDiscountMap(11).Fixes #242