### What changes were proposed in this pull request?
In this PR, we add a parameter in the python function vector_to_array(col) that allows converting to a column of arrays of Float (32bits) in scala, which would be mapped to a numpy array of dtype=float32.
### Why are the changes needed?
In the downstream ML training, using float32 instead of float64 (default) would allow a larger batch size, i.e., allow more data to fit in the memory.
### Does this PR introduce any user-facing change?
Yes.
Old: `vector_to_array()` only take one param
```
df.select(vector_to_array("colA"), ...)
```
New: `vector_to_array()` can take an additional optional param: `dtype` = "float32" (or "float64")
```
df.select(vector_to_array("colA", "float32"), ...)
```
### How was this patch tested?
Unit test in scala.
doctest in python.
Closes#27522 from liangz1/udf-float32.
Authored-by: Liang Zhang <liang.zhang@databricks.com>
Signed-off-by: WeichenXu <weichen.xu@databricks.com>
### What changes were proposed in this pull request?
Add ```HasBlockSize``` in shared Params in both Scala and Python.
Make ALS/MLP extend ```HasBlockSize```
### Why are the changes needed?
Add ```HasBlockSize ``` in ALS, so user can specify the blockSize.
Make ```HasBlockSize``` a shared param so both ALS and MLP can use it.
### Does this PR introduce any user-facing change?
Yes
```ALS.setBlockSize/getBlockSize```
```ALSModel.setBlockSize/getBlockSize```
### How was this patch tested?
Manually tested. Also added doctest.
Closes#27501 from huaxingao/spark_30662.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
Revert
#27360#27396#27374#27389
### Why are the changes needed?
BLAS need more performace tests, specially on sparse datasets.
Perfermance test of LogisticRegression (https://github.com/apache/spark/pull/27374) on sparse dataset shows that blockify vectors to matrices and use BLAS will cause performance regression.
LinearSVC and LinearRegression were also updated in the same way as LogisticRegression, so we need to revert them to make sure no regression.
### Does this PR introduce any user-facing change?
remove newly added param blockSize
### How was this patch tested?
reverted testsuites
Closes#27487 from zhengruifeng/revert_blockify_ii.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
var `negThetaSum` is always used together with `pi`, so we can add them at first
### Why are the changes needed?
only need to add one var `piMinusThetaSum`, instead of `pi` and `negThetaSum`
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#27427 from zhengruifeng/nb_predict.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
1, use blocks instead of vectors for performance improvement
2, use Level-2 BLAS
3, move standardization of input vectors outside of gradient computation
### Why are the changes needed?
1, less RAM to persist training data; (save ~40%)
2, faster than existing impl; (30% ~ 102%)
### Does this PR introduce any user-facing change?
add a new expert param `blockSize`
### How was this patch tested?
updated testsuites
Closes#27396 from zhengruifeng/blockify_lireg.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
Make ALS/MLP extend ```HasBlockSize```
### Why are the changes needed?
Currently, MLP has its own ```blockSize``` param, we should make MLP extend ```HasBlockSize``` since ```HasBlockSize``` was added in ```sharedParams.scala``` recently.
ALS doesn't have ```blockSize``` param now, we can make it extend ```HasBlockSize```, so user can specify the ```blockSize```.
### Does this PR introduce any user-facing change?
Yes
```ALS.setBlockSize``` and ```ALS.getBlockSize```
```ALSModel.setBlockSize``` and ```ALSModel.getBlockSize```
### How was this patch tested?
Manually tested. Also added doctest.
Closes#27389 from huaxingao/spark-30662.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
In the PR, I propose to replace deprecated method `computeCost` of `BisectingKMeansModel` by `summary.trainingCost`.
### Why are the changes needed?
The changes eliminate deprecation warnings:
```
BisectingKMeansSuite.scala:108: method computeCost in class BisectingKMeansModel is deprecated (since 3.0.0): This method is deprecated and will be removed in future versions. Use ClusteringEvaluator instead. You can also get the cost on the training dataset in the summary.
[warn] assert(model.computeCost(dataset) < 0.1)
BisectingKMeansSuite.scala:135: method computeCost in class BisectingKMeansModel is deprecated (since 3.0.0): This method is deprecated and will be removed in future versions. Use ClusteringEvaluator instead. You can also get the cost on the training dataset in the summary.
[warn] assert(model.computeCost(dataset) == summary.trainingCost)
BisectingKMeansSuite.scala:323: method computeCost in class BisectingKMeansModel is deprecated (since 3.0.0): This method is deprecated and will be removed in future versions. Use ClusteringEvaluator instead. You can also get the cost on the training dataset in the summary.
[warn] model.computeCost(dataset)
```
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By running `BisectingKMeansSuite` via:
```
./build/sbt "test:testOnly *BisectingKMeansSuite"
```
Closes#27401 from MaxGekk/kmeans-computeCost-warning.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
1, use blocks instead of vectors
2, use Level-2 BLAS for binary, use Level-3 BLAS for multinomial
### Why are the changes needed?
1, less RAM to persist training data; (save ~40%)
2, faster than existing impl; (40% ~ 92%)
### Does this PR introduce any user-facing change?
add a new expert param `blockSize`
### How was this patch tested?
updated testsuites
Closes#27374 from zhengruifeng/blockify_lor.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
1, stack input vectors to blocks (like ALS/MLP);
2, add new param `blockSize`;
3, add a new class `InstanceBlock`
4, standardize the input outside of optimization procedure;
### Why are the changes needed?
1, reduce RAM to persist traing dataset; (save ~40% in test)
2, use Level-2 BLAS routines; (12% ~ 28% faster, without native BLAS)
### Does this PR introduce any user-facing change?
a new param `blockSize`
### How was this patch tested?
existing and updated testsuites
Closes#27360 from zhengruifeng/blockify_svc.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
Remove ```numTrees``` in GBT in 3.0.0.
### Why are the changes needed?
Currently, GBT has
```
/**
* Number of trees in ensemble
*/
Since("2.0.0")
val getNumTrees: Int = trees.length
```
and
```
/** Number of trees in ensemble */
val numTrees: Int = trees.length
```
I think we should remove one of them. We deprecated it in 2.4.5 via https://github.com/apache/spark/pull/27352.
### Does this PR introduce any user-facing change?
Yes, remove ```numTrees``` in GBT in 3.0.0
### How was this patch tested?
existing tests
Closes#27330 from huaxingao/spark-numTrees.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
add a param `bootstrap` to control whether bootstrap samples are used.
### Why are the changes needed?
Current RF with numTrees=1 will directly build a tree using the orignial dataset,
while with numTrees>1 it will use bootstrap samples to build trees.
This design is for training a DecisionTreeModel by the impl of RandomForest, however, it is somewhat strange.
In Scikit-Learn, there is a param [bootstrap](https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html#sklearn.ensemble.RandomForestClassifier) to control whether bootstrap samples are used.
### Does this PR introduce any user-facing change?
Yes, new param is added
### How was this patch tested?
existing testsuites
Closes#27254 from zhengruifeng/add_bootstrap.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
unpersist graph outside checkpointer, like what Pregel does
### Why are the changes needed?
Shown in [SPARK-30503](https://issues.apache.org/jira/browse/SPARK-30503), intermediate edges are not unpersisted
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites and manual test
Closes#27261 from zhengruifeng/lda_checkpointer.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
Change ```DecisionTreeClassifier``` to ```FMClassifier``` in ```OneVsRest``` setWeightCol test
### Why are the changes needed?
In ```OneVsRest```, if the classifier doesn't support instance weight, ```OneVsRest``` weightCol will be ignored, so unit test has tested one classifier(```LogisticRegression```) that support instance weight, and one classifier (```DecisionTreeClassifier```) that doesn't support instance weight. Since ```DecisionTreeClassifier``` now supports instance weight, we need to change it to the classifier that doesn't have weight support.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Existing test
Closes#27204 from huaxingao/spark-ovr-minor.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
add setInputCol/setOutputCol in OHEModel
### Why are the changes needed?
setInputCol/setOutputCol should be in OHEModel too.
### Does this PR introduce any user-facing change?
Yes.
```OHEModel.setInputCol```
```OHEModel.setOutputCol```
### How was this patch tested?
Manually tested.
Closes#27228 from huaxingao/spark-29565.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
1, add field `storageLevel` in `PeriodicRDDCheckpointer`
2, for ml.GBT/ml.RF set storageLevel=`StorageLevel.MEMORY_AND_DISK`
### Why are the changes needed?
Intermediate RDDs in ML are cached with storageLevel=StorageLevel.MEMORY_AND_DISK.
PeriodicRDDCheckpointer & PeriodicGraphCheckpointer now store RDD with storageLevel=StorageLevel.MEMORY_ONLY, it maybe nice to set the storageLevel of checkpointer.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#27189 from zhengruifeng/checkpointer_storage.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
1, change `convertToBaggedRDDSamplingWithReplacement` to attach instance weights
2, make RF supports weights
### Why are the changes needed?
`weightCol` is already exposed, while RF has not support weights.
### Does this PR introduce any user-facing change?
Yes, new setters
### How was this patch tested?
added testsuites
Closes#27097 from zhengruifeng/rf_support_weight.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
Fix all the failed tests when enable AQE.
### Why are the changes needed?
Run more tests with AQE to catch bugs, and make it easier to enable AQE by default in the future.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Existing unit tests
Closes#26813 from JkSelf/enableAQEDefault.
Authored-by: jiake <ke.a.jia@intel.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
### What changes were proposed in this pull request?
add weight support in BisectingKMeans
### Why are the changes needed?
BisectingKMeans should support instance weighting
### Does this PR introduce any user-facing change?
Yes. BisectingKMeans.setWeight
### How was this patch tested?
Unit test
Closes#27035 from huaxingao/spark_30351.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
Make Regressors extend abstract class Regressor:
```AFTSurvivalRegression extends Estimator => extends Regressor```
```DecisionTreeRegressor extends Predictor => extends Regressor```
```FMRegressor extends Predictor => extends Regressor```
```GBTRegressor extends Predictor => extends Regressor```
```RandomForestRegressor extends Predictor => extends Regressor```
We will not make ```IsotonicRegression``` extend ```Regressor``` because it is tricky to handle both DoubleType and VectorType.
### Why are the changes needed?
Make class hierarchy consistent for all Regressors
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing tests
Closes#27168 from huaxingao/spark-30377.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
1, del `NodeIdCache`, and use `PeriodicRDDCheckpointer` instead;
2, reuse broadcasted `Splits` in the whole training;
### Why are the changes needed?
1, The functionality of `NodeIdCache` and `PeriodicRDDCheckpointer` are highly similar, and the update process of nodeIds is simple; One goal of "Generalize PeriodicGraphCheckpointer for RDDs" in SPARK-5561 is to use checkpointer in RandomForest;
2, only need to broadcast `Splits` once;
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Existing testsuites
Closes#27145 from zhengruifeng/del_NodeIdCache.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
1, for primitive types `Array.fill(n)(0)` -> `Array.ofDim(n)`;
2, for `AnyRef` types `Array.fill(n)(null)` -> `Array.ofDim(n)`;
3, for primitive types `Array.empty[XXX]` -> `Array.emptyXXXArray`
### Why are the changes needed?
`Array.ofDim` avoid assignments;
`Array.emptyXXXArray` avoid create new object;
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#27133 from zhengruifeng/minor_fill_ofDim.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
### What changes were proposed in this pull request?
Make GBT reuse splits/treePoints for all trees:
1, reuse splits/treePoints for all trees:
existing impl will find feature splits and transform input vectors to treePoints for each tree; while other famous impls like XGBoost/lightGBM will build a global splits/binned features and reuse them for all trees;
Note: the sampling rate in existing impl to build `splits` is not the param `subsamplingRate` but the output of `RandomForest.samplesFractionForFindSplits` which depends on `maxBins` and `numExamples`.
Note II: Existing impl do not guarantee that splits among iteration are the same, so this may cause a little difference in convergence.
2, do not cache input vectors:
existing impl will cached the input twice: 1,`input: RDD[Instance]` is used to compute/update prediction and errors; 2, at each iteration, input is transformed to bagged points, the bagged points will be cached during this iteration;
In this PR,`input: RDD[Instance]` is no longer cached, since it is only used three times: 1, compute metadata; 2, find splits; 3, converted to treePoints;
Instead, the treePoints `RDD[TreePoint]` is cached, at each iter, it is convert to bagged points by attaching extra `labelWithCounts: RDD[(Double, Int)]` containing residuals/sampleCount information, this rdd is relative small (like cached `norms` in KMeans);
To compute/update prediction and errors, new prediction method based on binned features are added in `Node`
### Why are the changes needed?
for perfermance improvement:
1,40%~50% faster than existing impl
2,save 30%~50% RAM
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites & several manual tests in REPL
Closes#27103 from zhengruifeng/gbt_reuse_bagged.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
PySpark UDF to convert MLlib vectors to dense arrays.
Example:
```
from pyspark.ml.functions import vector_to_array
df.select(vector_to_array(col("features"))
```
### Why are the changes needed?
If a PySpark user wants to convert MLlib sparse/dense vectors in a DataFrame into dense arrays, an efficient approach is to do that in JVM. However, it requires PySpark user to write Scala code and register it as a UDF. Often this is infeasible for a pure python project.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
UT.
Closes#26910 from WeichenXu123/vector_to_array.
Authored-by: WeichenXu <weichen.xu@databricks.com>
Signed-off-by: Xiangrui Meng <meng@databricks.com>
### What changes were proposed in this pull request?
1, fix `BaggedPoint.convertToBaggedRDD` when `subsamplingRate < 1.0`
2, reorg `RandomForest.runWithMetadata` btw
### Why are the changes needed?
In GBT, Instance weights will be discarded if subsamplingRate<1
1, `baggedPoint: BaggedPoint[TreePoint]` is used in the tree growth to find best split;
2, `BaggedPoint[TreePoint]` contains two weights:
```scala
class BaggedPoint[Datum](val datum: Datum, val subsampleCounts: Array[Int], val sampleWeight: Double = 1.0)
class TreePoint(val label: Double, val binnedFeatures: Array[Int], val weight: Double)
```
3, only the var `sampleWeight` in `BaggedPoint` is used, the var `weight` in `TreePoint` is never used in finding splits;
4, The method `BaggedPoint.convertToBaggedRDD` was changed in https://github.com/apache/spark/pull/21632, it was only for decisiontree, so only the following code path was changed;
```
if (numSubsamples == 1 && subsamplingRate == 1.0) {
convertToBaggedRDDWithoutSampling(input, extractSampleWeight)
}
```
5, In https://github.com/apache/spark/pull/25926, I made GBT support weights, but only test it with default `subsamplingRate==1`.
GBT with `subsamplingRate<1` will convert treePoints to baggedPoints via
```scala
convertToBaggedRDDSamplingWithoutReplacement(input, subsamplingRate, numSubsamples, seed)
```
in which the orignial weights from `weightCol` will be discarded and all `sampleWeight` are assigned default 1.0;
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
updated testsuites
Closes#27070 from zhengruifeng/gbt_sampling.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
make FMClassifier/Regressor call super class method extractLabeledPoints
### Why are the changes needed?
code reuse
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing tests
Closes#27093 from huaxingao/spark-FM.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
use `.ml.Summarizer` instead of `.mllib.MultivariateOnlineSummarizer` to avoid computation of unused metrics
### Why are the changes needed?
to avoid computation of unused metrics
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#27059 from zhengruifeng/pac_summarizer.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
Check before caching zippedData (as suggested in https://github.com/apache/spark/pull/26483#issuecomment-569702482).
### Why are the changes needed?
If the `data` is already cached before calling `run` method of `KMeans` then `zippedData.persist()` will hurt the performance. Hence, persisting it conditionally.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
Manually.
Closes#27052 from amanomer/29823followup.
Authored-by: Aman Omer <amanomer1996@gmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
Make ```MultilayerPerceptronClassificationModel``` extend ```MultilayerPerceptronParams```
### Why are the changes needed?
Make ```MultilayerPerceptronClassificationModel``` extend ```MultilayerPerceptronParams``` to expose the training params, so user can see these params when calling ```extractParamMap```
### Does this PR introduce any user-facing change?
Yes. The ```MultilayerPerceptronParams``` such as ```seed```, ```maxIter``` ... are available in ```MultilayerPerceptronClassificationModel``` now
### How was this patch tested?
Manually tested ```MultilayerPerceptronClassificationModel.extractParamMap()``` to verify all the new params are there.
Closes#26838 from huaxingao/spark-30144.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
1, add new foreach-like methods: foreach/foreachNonZero
2, add iterator: iterator/activeIterator/nonZeroIterator
### Why are the changes needed?
see the [ticke](https://issues.apache.org/jira/browse/SPARK-30329) for details
foreach/foreachNonZero: for both convenience and performace (SparseVector.foreach should be faster than current traversal method)
iterator/activeIterator/nonZeroIterator: add the three iterators, so that we can futuremore add/change some impls based on those iterators for both ml and mllib sides, to avoid vector conversions.
### Does this PR introduce any user-facing change?
Yes, new methods are added
### How was this patch tested?
added testsuites
Closes#26982 from zhengruifeng/vector_iter.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
add instr.logSumOfWeights in the Algo that has weightCol support
### Why are the changes needed?
Many algorithms support weightCol now. I think weightsum is useful info to add to the log.
### Does this PR introduce any user-facing change?
no
### How was this patch tested?
manually tested
Closes#26972 from huaxingao/spark-30321.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
Refactor `RandomForest.findSplits` by applying `aggregateByKey` instead of `groupByKey`
### Why are the changes needed?
Current impl of `RandomForest.findSplits` uses `groupByKey` to collect non-zero values for each feature, so it is quite dangerous.
After looking into the following logic to find splits, I found that collecting all non-zero values is not necessary, and we only need weightSums of distinct values.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#27040 from zhengruifeng/rf_opt.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
1, expose predictRaw and predictProbability
2, add tests
### Why are the changes needed?
single instance prediction is useful out of spark, specially for online prediction.
Current `predict` is exposed, but it is not enough.
### Does this PR introduce any user-facing change?
Yes, new methods are exposed
### How was this patch tested?
added testsuites
Closes#27015 from zhengruifeng/expose_raw_prob.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
using `MetadataUtils.getNumFeatures` to extract the numFeatures
### Why are the changes needed?
may avoid `first`/`head` job if metadata has attrGroup
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#27037 from zhengruifeng/unify_numFeatures.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
compute count and quantile on one pass
### Why are the changes needed?
to avoid extra pass
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#26990 from zhengruifeng/quantile_count_lsh.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
1, remove unused imports and variables
2, remove `countAccum: LongAccumulator`, since `costAccum: DoubleAccumulator` also records the count
3, mark `clusterCentersWithNorm` in KMeansModel trasient and lazy, since it is only used in transformation and can be directly generated from the centers.
### Why are the changes needed?
1,remove unused codes
2,avoid repeated computation
3,reduce broadcasted size
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#27014 from zhengruifeng/kmeans_clean_up.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
precompute the `DecisionTreeMetadata` and reuse it for all trees
### Why are the changes needed?
In existing impl, each `DecisionTreeRegressor` needs a pass on the whole dataset to calculate the same `DecisionTreeMetadata` repeatedly.
In this PR, with default depth=5, it is about 8% faster then existing impl
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#27011 from zhengruifeng/gbt_reuse_instr_meta.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
supports instance weighting in GMM
### Why are the changes needed?
ML should support instance weighting
### Does this PR introduce any user-facing change?
yes, a new param `weightCol` is exposed
### How was this patch tested?
added testsuits
Closes#26735 from zhengruifeng/gmm_support_weight.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
Implement Factorization Machines as a ml-pipeline component
1. loss function supports: logloss, mse
2. optimizer: GD, adamW
### Why are the changes needed?
Factorization Machines is widely used in advertising and recommendation system to estimate CTR(click-through rate).
Advertising and recommendation system usually has a lot of data, so we need Spark to estimate the CTR, and Factorization Machines are common ml model to estimate CTR.
References:
1. S. Rendle, “Factorization machines,” in Proceedings of IEEE International Conference on Data Mining (ICDM), pp. 995–1000, 2010.
https://www.csie.ntu.edu.tw/~b97053/paper/Rendle2010FM.pdf
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
run unit tests
Closes#27000 from mob-ai/ml/fm.
Authored-by: zhanjf <zhanjf@mob.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
LibSVMDataSource attach AttributeGroup
### Why are the changes needed?
LibSVMDataSource will attach a special metadata to indicate numFeatures:
```scala
scala> val data = spark.read.format("libsvm").load("/data0/Dev/Opensource/spark/data/mllib/sample_multiclass_classification_data.txt")
scala> data.schema("features").metadata
res0: org.apache.spark.sql.types.Metadata = {"numFeatures":4}
```
However, all ML impls will try to obtain vector size via AttributeGroup, which can not use this metadata:
```scala
scala> import org.apache.spark.ml.attribute._
import org.apache.spark.ml.attribute._
scala> AttributeGroup.fromStructField(data.schema("features")).size
res1: Int = -1
```
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
added tests
Closes#27003 from zhengruifeng/libsvm_attr_group.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
compute the medians/ranges more distributedly
### Why are the changes needed?
It is a bottleneck to collect the whole Array[QuantileSummaries] from executors,
since a QuantileSummaries is a large object, which maintains arrays of large sizes 10k(`defaultCompressThreshold`)/50k(`defaultHeadSize`).
In Spark-Shell with default params, I processed a dataset with numFeatures=69,200, and existing impl fail due to OOM.
After this PR, it will sucessfuly fit the model.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#26803 from zhengruifeng/robust_high_dim.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
1,modify the toString in SQLTransformer & VectorSizeHint
2,add toString in RegexTokenizer
### Why are the changes needed?
in SQLTransformer & VectorSizeHint , `toString` methods directly call getter of param without default values.
This will cause `java.util.NoSuchElementException` in REPL:
```scala
scala> val vs = new VectorSizeHint()
java.util.NoSuchElementException: Failed to find a default value for size
at org.apache.spark.ml.param.Params.$anonfun$getOrDefault$2(params.scala:780)
at scala.Option.getOrElse(Option.scala:189)
```
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
existing testsuites
Closes#26999 from zhengruifeng/fix_toString.
Authored-by: zhengruifeng <ruifengz@foxmail.com>
Signed-off-by: zhengruifeng <ruifengz@foxmail.com>
### What changes were proposed in this pull request?
Implement Factorization Machines as a ml-pipeline component
1. loss function supports: logloss, mse
2. optimizer: GD, adamW
### Why are the changes needed?
Factorization Machines is widely used in advertising and recommendation system to estimate CTR(click-through rate).
Advertising and recommendation system usually has a lot of data, so we need Spark to estimate the CTR, and Factorization Machines are common ml model to estimate CTR.
References:
1. S. Rendle, “Factorization machines,” in Proceedings of IEEE International Conference on Data Mining (ICDM), pp. 995–1000, 2010.
https://www.csie.ntu.edu.tw/~b97053/paper/Rendle2010FM.pdf
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
run unit tests
Closes#26124 from mob-ai/ml/fm.
Authored-by: zhanjf <zhanjf@mob.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
Fixed typo in `docs` directory and in other directories
1. Find typo in `docs` and apply fixes to files in all directories
2. Fix `the the` -> `the`
### Why are the changes needed?
Better readability of documents
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
No test needed
Closes#26976 from kiszk/typo_20191221.
Authored-by: Kazuaki Ishizaki <ishizaki@jp.ibm.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
1. Revert "Preparing development version 3.0.1-SNAPSHOT": 56dcd79
2. Revert "Preparing Spark release v3.0.0-preview2-rc2": c216ef1
### Why are the changes needed?
Shouldn't change master.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
manual test:
https://github.com/apache/spark/compare/5de5e46..wangyum:revert-masterCloses#26915 from wangyum/revert-master.
Authored-by: Yuming Wang <yumwang@ebay.com>
Signed-off-by: Yuming Wang <wgyumg@gmail.com>
### What changes were proposed in this pull request?
- Replace `Seq[String]` by `Seq[_]` in `StopWordsRemoverSuite` because `String` type is unchecked due erasure.
- Throw an exception for default case in `MLTest.checkNominalOnDF` because we don't expect other attribute types currently.
- Explicitly cast float to double in `BigDecimal(y)`. This is what the `apply()` method does for `float`s.
- Replace deprecated `verifyZeroInteractions` by `verifyNoInteractions`.
- Equivalent replacement of `\0` by `\u0000` in `CSVExprUtilsSuite`
- Import `scala.language.implicitConversions` in `CollectionExpressionsSuite`, `HashExpressionsSuite` and in `ExpressionParserSuite`.
### Why are the changes needed?
The changes fix compiler warnings showed in the JIRA ticket https://issues.apache.org/jira/browse/SPARK-30170 . Eliminating the warning highlights other warnings which could take more attention to real problems.
### Does this PR introduce any user-facing change?
No
### How was this patch tested?
By existing test suites `StopWordsRemoverSuite`, `AnalysisExternalCatalogSuite`, `CSVExprUtilsSuite`, `CollectionExpressionsSuite`, `HashExpressionsSuite`, `ExpressionParserSuite` and sub-tests of `MLTest`.
Closes#26799 from MaxGekk/eliminate-warning-2.
Authored-by: Maxim Gekk <max.gekk@gmail.com>
Signed-off-by: Sean Owen <srowen@gmail.com>
### What changes were proposed in this pull request?
See https://issues.apache.org/jira/browse/SPARK-30195 for the background; I won't repeat it here. This is sort of a grab-bag of related issues.
### Why are the changes needed?
To cross-compile with Scala 2.13 later.
### Does this PR introduce any user-facing change?
No.
### How was this patch tested?
Existing tests for 2.12. I've been manually checking that this actually resolves the compile problems in 2.13 separately.
Closes#26826 from srowen/SPARK-30195.
Authored-by: Sean Owen <srowen@gmail.com>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
### What changes were proposed in this pull request?
add weight support in KMeans
### Why are the changes needed?
KMeans should support weighting
### Does this PR introduce any user-facing change?
Yes. ```KMeans.setWeightCol```
### How was this patch tested?
Unit Tests
Closes#26739 from huaxingao/spark-29967.
Authored-by: Huaxin Gao <huaxing@us.ibm.com>
Signed-off-by: Sean Owen <srowen@gmail.com>