EBernhardson has uploaded a new change for review. (
https://gerrit.wikimedia.org/r/406068 )
Change subject: Fixup xgboost training
......................................................................
Fixup xgboost training
* Use tree_method = hist when training on a single worker. This is
significantly faster than the approx method used by default.
* We've always trained with dense feature matrix's, and the ltr plugin
only supports dense evaluation, but the DataWriter was writing out
sparse matrixs. This caused a degredation in ndcg.
* The txt file emitted by DataWriter has to be read by
lightgbm and xgboost. As such it starts features at idx 1 to make
lightgbm happy (which stores the label at idx 0). This broke
XGBoostModel.eval because it was not providing the empty feature at
index 0 that training sees.
* XGBoostModel.loadModelFrom* always failed because the summary
method on the jvm side throws an exception (metrics are not
serialized). Wrap in try/except and set summary to None when not
available.
Change-Id: I48d7bf96a300313b6f62e3f60742345e8bd1a83f
---
M jvm/src/main/scala/org/wikimedia/search/mjolnir/DataWriter.scala
M jvm/src/test/scala/org/wikimedia/search/mjolnir/DataWriterSuite.scala
M mjolnir/training/xgboost.py
M mjolnir/utilities/make_folds.py
4 files changed, 10 insertions(+), 4 deletions(-)
git pull ssh://gerrit.wikimedia.org:29418/search/MjoLniR
refs/changes/68/406068/1
diff --git a/jvm/src/main/scala/org/wikimedia/search/mjolnir/DataWriter.scala
b/jvm/src/main/scala/org/wikimedia/search/mjolnir/DataWriter.scala
index 8d6976b..01fcf6d 100644
--- a/jvm/src/main/scala/org/wikimedia/search/mjolnir/DataWriter.scala
+++ b/jvm/src/main/scala/org/wikimedia/search/mjolnir/DataWriter.scala
@@ -28,7 +28,7 @@
) extends Serializable {
// Accepting JavaSparkContext for py4j compatability
- def this(sc: JavaSparkContext) = this(sc.broadcast(new
SerializableConfiguration(sc.hadoopConfiguration)))
+ def this(sc: JavaSparkContext, sparse: Boolean) = this(sc.broadcast(new
SerializableConfiguration(sc.hadoopConfiguration)), sparse)
private def asHDFSPath(path: String): HDFSPath = if (path.charAt(0) == '/') {
new HDFSPath(s"file://$path")
diff --git
a/jvm/src/test/scala/org/wikimedia/search/mjolnir/DataWriterSuite.scala
b/jvm/src/test/scala/org/wikimedia/search/mjolnir/DataWriterSuite.scala
index ca85260..06004a7 100644
--- a/jvm/src/test/scala/org/wikimedia/search/mjolnir/DataWriterSuite.scala
+++ b/jvm/src/test/scala/org/wikimedia/search/mjolnir/DataWriterSuite.scala
@@ -44,7 +44,7 @@
try {
val df = makeData()
val pattern = s"$testDir/%s-fold-%s-partition-%d"
- val writer = new DataWriter(spark.sparkContext)
+ val writer = new DataWriter(spark.sparkContext, true)
val folds = writer.write(df, numWorkers, pattern, foldCol)
assert(folds.length == expectedFolds)
diff --git a/mjolnir/training/xgboost.py b/mjolnir/training/xgboost.py
index abeaabf..4d6bb9d 100644
--- a/mjolnir/training/xgboost.py
+++ b/mjolnir/training/xgboost.py
@@ -4,6 +4,7 @@
import mjolnir.training.hyperopt
from mjolnir.training.tuning import make_cv_objective, ModelSelection
import numpy as np
+import py4j
import pyspark
import pyspark.sql
from pyspark.sql import functions as F
@@ -114,6 +115,8 @@
# ints, so this gets all the types right for Java. Also makes
# a copy of params so we don't modifying the incoming dict.
params = _coerce_params(params)
+ # Histogram doesn't work with distributed training
+ params['tree_method'] = 'hist' if len(fold) == 1 else 'approx'
# TODO: Maybe num_rounds should just be external? But it's easier
# to do hyperparameter optimization with a consistent dict interface
kwargs = {
@@ -158,7 +161,10 @@
class XGBoostModel(object):
def __init__(self, j_xgb_model):
self._j_xgb_model = j_xgb_model
- self.summary = XGBoostSummary(self._j_xgb_model.summary())
+ try:
+ self.summary = XGBoostSummary(self._j_xgb_model.summary())
+ except py4j.protocol.Py4JJavaError:
+ self.summary = None
@staticmethod
def trainWithFiles(fold, train_matrix, params, num_rounds=100,
diff --git a/mjolnir/utilities/make_folds.py b/mjolnir/utilities/make_folds.py
index c7ac04d..5cbd682 100644
--- a/mjolnir/utilities/make_folds.py
+++ b/mjolnir/utilities/make_folds.py
@@ -64,7 +64,7 @@
write_xgb(local_input, local_output.name)
# Write out as text files from scala, much faster than shuffling to python
- writer = sc._jvm.org.wikimedia.search.mjolnir.DataWriter(sc._jsc)
+ writer = sc._jvm.org.wikimedia.search.mjolnir.DataWriter(sc._jsc, False)
j_paths = writer.write(df._jdf, num_workers, path_format, fold_col)
# Convert everything to python objects
--
To view, visit https://gerrit.wikimedia.org/r/406068
To unsubscribe, visit https://gerrit.wikimedia.org/r/settings
Gerrit-MessageType: newchange
Gerrit-Change-Id: I48d7bf96a300313b6f62e3f60742345e8bd1a83f
Gerrit-PatchSet: 1
Gerrit-Project: search/MjoLniR
Gerrit-Branch: master
Gerrit-Owner: EBernhardson <ebernhard...@wikimedia.org>
_______________________________________________
MediaWiki-commits mailing list
MediaWiki-commits@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/mediawiki-commits
