This PR is based on #3255 , fix conflicts and code style.
Closes#3255.
Author: Yandu Oppacher <yandu.oppacher@jadedpixel.com>
Author: Davies Liu <davies@databricks.com>
Closes#3901 from davies/refactor-python-profile-code and squashes the following commits:
b4a9306 [Davies Liu] fix tests
4b79ce8 [Davies Liu] add docstring for profiler_cls
2700e47 [Davies Liu] use BasicProfiler as default
349e341 [Davies Liu] more refactor
6a5d4df [Davies Liu] refactor and fix tests
31bf6b6 [Davies Liu] fix code style
0864b5d [Yandu Oppacher] Remove unused method
76a6c37 [Yandu Oppacher] Added a profile collector to accumulate the profilers per stage
9eefc36 [Yandu Oppacher] Fix doc
9ace076 [Yandu Oppacher] Refactor of profiler, and moved tests around
8739aff [Yandu Oppacher] Code review fixes
9bda3ec [Yandu Oppacher] Refactor profiler code
Since Java and Scala both have access to iterate over partitions via the "toLocalIterator" function, python should also have that same ability.
Author: Michael Nazario <mnazario@palantir.com>
Closes#4237 from mnazario/feature/toLocalIterator and squashes the following commits:
1c58526 [Michael Nazario] Fix documentation off by one error
0cdc8f8 [Michael Nazario] Add toLocalIterator to PySpark
Author: Sandy Ryza <sandy@cloudera.com>
Closes#4251 from sryza/sandy-spark-5458 and squashes the following commits:
460827a [Sandy Ryza] Python too
d2dc160 [Sandy Ryza] SPARK-5458. Refer to aggregateByKey instead of combineByKey in docs
This is found through reading RDD from `sc.newAPIHadoopRDD` and writing it back using `rdd.saveAsNewAPIHadoopFile` in pyspark.
It turns out that whenever there are multiple RDD conversions from JavaRDD to PythonRDD then back to JavaRDD, the exception below happens:
```
15/01/16 10:28:31 ERROR Executor: Exception in task 0.0 in stage 3.0 (TID 7)
java.lang.ClassCastException: [Ljava.lang.Object; cannot be cast to java.util.ArrayList
at org.apache.spark.api.python.SerDeUtil$$anonfun$pythonToJava$1$$anonfun$apply$1.apply(SerDeUtil.scala:157)
at org.apache.spark.api.python.SerDeUtil$$anonfun$pythonToJava$1$$anonfun$apply$1.apply(SerDeUtil.scala:153)
at scala.collection.Iterator$$anon$13.hasNext(Iterator.scala:371)
at scala.collection.Iterator$$anon$10.hasNext(Iterator.scala:308)
```
The test case code below reproduces it:
```
from pyspark.rdd import RDD
dl = [
(u'2', {u'director': u'David Lean'}),
(u'7', {u'director': u'Andrew Dominik'})
]
dl_rdd = sc.parallelize(dl)
tmp = dl_rdd._to_java_object_rdd()
tmp2 = sc._jvm.SerDe.javaToPython(tmp)
t = RDD(tmp2, sc)
t.count()
tmp = t._to_java_object_rdd()
tmp2 = sc._jvm.SerDe.javaToPython(tmp)
t = RDD(tmp2, sc)
t.count() # it blows up here during the 2nd time of conversion
```
Author: Winston Chen <wchen@quid.com>
Closes#4146 from wingchen/master and squashes the following commits:
903df7d [Winston Chen] SPARK-5361, update to toSeq based on the PR
5d90a83 [Winston Chen] SPARK-5361, make python pretty, so to pass PEP 8 checks
126be6b [Winston Chen] SPARK-5361, add in test case
4cf1187 [Winston Chen] SPARK-5361, add in test case
9f1a097 [Winston Chen] add in tuple handling while converting form python RDD back to JavaRDD
This patch adds more helpful error messages for invalid programs that define nested RDDs, broadcast RDDs, perform actions inside of transformations (e.g. calling `count()` from inside of `map()`), and call certain methods on stopped SparkContexts. Currently, these invalid programs lead to confusing NullPointerExceptions at runtime and have been a major source of questions on the mailing list and StackOverflow.
In a few cases, I chose to log warnings instead of throwing exceptions in order to avoid any chance that this patch breaks programs that worked "by accident" in earlier Spark releases (e.g. programs that define nested RDDs but never run any jobs with them).
In SparkContext, the new `assertNotStopped()` method is used to check whether methods are being invoked on a stopped SparkContext. In some cases, user programs will not crash in spite of calling methods on stopped SparkContexts, so I've only added `assertNotStopped()` calls to methods that always throw exceptions when called on stopped contexts (e.g. by dereferencing a null `dagScheduler` pointer).
Author: Josh Rosen <joshrosen@databricks.com>
Closes#3884 from JoshRosen/SPARK-5063 and squashes the following commits:
a38774b [Josh Rosen] Fix spelling typo
a943e00 [Josh Rosen] Convert two exceptions into warnings in order to avoid breaking user programs in some edge-cases.
2d0d7f7 [Josh Rosen] Fix test to reflect 1.2.1 compatibility
3f0ea0c [Josh Rosen] Revert two unintentional formatting changes
8e5da69 [Josh Rosen] Remove assertNotStopped() calls for methods that were sometimes safe to call on stopped SC's in Spark 1.2
8cff41a [Josh Rosen] IllegalStateException fix
6ef68d0 [Josh Rosen] Fix Python line length issues.
9f6a0b8 [Josh Rosen] Add improved error messages to PySpark.
13afd0f [Josh Rosen] SparkException -> IllegalStateException
8d404f3 [Josh Rosen] Merge remote-tracking branch 'origin/master' into SPARK-5063
b39e041 [Josh Rosen] Fix BroadcastSuite test which broadcasted an RDD
99cc09f [Josh Rosen] Guard against calling methods on stopped SparkContexts.
34833e8 [Josh Rosen] Add more descriptive error message.
57cc8a1 [Josh Rosen] Add error message when directly broadcasting RDD.
15b2e6b [Josh Rosen] [SPARK-5063] Useful error messages for nested RDDs and actions inside of transformations
This implements the functionality for SPARK-4749 and provides units tests in Scala and PySpark
Author: nate.crosswhite <nate.crosswhite@stresearch.com>
Author: nxwhite-str <nxwhite-str@users.noreply.github.com>
Author: Xiangrui Meng <meng@databricks.com>
Closes#3610 from nxwhite-str/master and squashes the following commits:
a2ebbd3 [nxwhite-str] Merge pull request #1 from mengxr/SPARK-4749-kmeans-seed
7668124 [Xiangrui Meng] minor updates
f8d5928 [nate.crosswhite] Addressing PR issues
277d367 [nate.crosswhite] Merge remote-tracking branch 'upstream/master'
9156a57 [nate.crosswhite] Merge remote-tracking branch 'upstream/master'
5d087b4 [nate.crosswhite] Adding KMeans train with seed and Scala unit test
616d111 [nate.crosswhite] Merge remote-tracking branch 'upstream/master'
35c1884 [nate.crosswhite] Add kmeans initial seed to pyspark API
Pretty minor, but submitted for consideration -- this would at least help people make this check in the most efficient way I know.
Author: Sean Owen <sowen@cloudera.com>
Closes#4074 from srowen/SPARK-5270 and squashes the following commits:
66885b8 [Sean Owen] Add note that JavaRDDLike should not be implemented by user code
2e9b490 [Sean Owen] More tests, and Mima-exclude the new isEmpty method in JavaRDDLike
28395ff [Sean Owen] Add isEmpty to Java, Python
7dd04b7 [Sean Owen] Add efficient RDD.isEmpty()
As part of SPARK-5193:
1. Removed UDFRegistration as a mixin in SQLContext and made it a field ("udf").
2. For Java UDFs, renamed dataType to returnType.
3. For Scala UDFs, added type tags.
4. Added all Java UDF registration methods to Scala's UDFRegistration.
5. Documentation
Author: Reynold Xin <rxin@databricks.com>
Closes#4056 from rxin/udf-registration and squashes the following commits:
ae9c556 [Reynold Xin] Updated example.
675a3c9 [Reynold Xin] Style fix
47c24ff [Reynold Xin] Python fix.
5f00c45 [Reynold Xin] Restore data type position in java udf and added typetags.
032f006 [Reynold Xin] [SPARK-5193][SQL] Reconcile Java and Scala UDFRegistration.
After the default batchSize changed to 0 (batched based on the size of object), but parallelize() still use BatchedSerializer with batchSize=1, this PR will use batchSize=1024 for parallelize by default.
Also, BatchedSerializer did not work well with list and numpy.ndarray, this improve BatchedSerializer by using __len__ and __getslice__.
Here is the benchmark for parallelize 1 millions int with list or ndarray:
| before | after | improvements
------- | ------------ | ------------- | -------
list | 11.7 s | 0.8 s | 14x
numpy.ndarray | 32 s | 0.7 s | 40x
Author: Davies Liu <davies@databricks.com>
Closes#4024 from davies/opt_numpy and squashes the following commits:
7618c7c [Davies Liu] improve performance of parallelize list/ndarray
Slightly different than the scala code which converts the sparsevector into a densevector and then checks the index.
I also hope I've added tests in the right place.
Author: MechCoder <manojkumarsivaraj334@gmail.com>
Closes#4025 from MechCoder/spark-2909 and squashes the following commits:
07d0f26 [MechCoder] STY: Rename item to index
f02148b [MechCoder] [SPARK-2909] [Mlib] SparseVector in pyspark now supports indexing
It will introduce problems if the object in dict/list/tuple can not support by py4j, such as Vector.
Also, pickle may have better performance for larger object (less RPC).
In some cases that the object in dict/list can not be pickled (such as JavaObject), we should still use MapConvert/ListConvert.
This PR should be ported into branch-1.2
Author: Davies Liu <davies@databricks.com>
Closes#4023 from davies/listconvert and squashes the following commits:
55d4ab2 [Davies Liu] fix MapConverter and ListConverter in MLlib
When attempting to infer the schema of an RDD that contains namedtuples, pyspark fails to identify the records as namedtuples, resulting in it raising an error.
Example:
```python
from pyspark import SparkContext
from pyspark.sql import SQLContext
from collections import namedtuple
import os
sc = SparkContext()
rdd = sc.textFile(os.path.join(os.getenv('SPARK_HOME'), 'README.md'))
TextLine = namedtuple('TextLine', 'line length')
tuple_rdd = rdd.map(lambda l: TextLine(line=l, length=len(l)))
tuple_rdd.take(5) # This works
sqlc = SQLContext(sc)
# The following line raises an error
schema_rdd = sqlc.inferSchema(tuple_rdd)
```
The error raised is:
```
File "/opt/spark-1.2.0-bin-hadoop2.4/python/pyspark/worker.py", line 107, in main
process()
File "/opt/spark-1.2.0-bin-hadoop2.4/python/pyspark/worker.py", line 98, in process
serializer.dump_stream(func(split_index, iterator), outfile)
File "/opt/spark-1.2.0-bin-hadoop2.4/python/pyspark/serializers.py", line 227, in dump_stream
vs = list(itertools.islice(iterator, batch))
File "/opt/spark-1.2.0-bin-hadoop2.4/python/pyspark/rdd.py", line 1107, in takeUpToNumLeft
yield next(iterator)
File "/opt/spark-1.2.0-bin-hadoop2.4/python/pyspark/sql.py", line 816, in convert_struct
raise ValueError("unexpected tuple: %s" % obj)
TypeError: not all arguments converted during string formatting
```
Author: Gabe Mulley <gabe@edx.org>
Closes#3978 from mulby/inferschema-namedtuple and squashes the following commits:
98c61cc [Gabe Mulley] Ensure exception message is populated correctly
375d96b [Gabe Mulley] Ensure schema can be inferred from a namedtuple
...ySpark MLlib
This is a follow up to PR3680 https://github.com/apache/spark/pull/3680 .
Author: RJ Nowling <rnowling@gmail.com>
Closes#3955 from rnowling/spark4891 and squashes the following commits:
1236a01 [RJ Nowling] Fix Python style issues
7a01a78 [RJ Nowling] Fix Python style issues
174beab [RJ Nowling] [SPARK-4891][PySpark][MLlib] Add gamma/log normal/exp dist sampling to PySpark MLlib
This is a small change addressing a potentially significant bug in how PySpark + MLlib handles non-float64 numpy arrays. The automatic conversion to `DenseVector` that occurs when passing RDDs to MLlib algorithms in PySpark should automatically upcast to float64s, but currently this wasn't actually happening. As a result, non-float64 would be silently parsed inappropriately during SerDe, yielding erroneous results when running, for example, KMeans.
The PR includes the fix, as well as a new test for the correct conversion behavior.
davies
Author: freeman <the.freeman.lab@gmail.com>
Closes#3902 from freeman-lab/fix-vector-convert and squashes the following commits:
764db47 [freeman] Add a test for proper conversion behavior
704f97e [freeman] Return array after changing type
This PR is a fixed version of the original PR #3237 by watermen and scwf.
This adds the ability to specify how many elements to print in `DStream.print`.
Author: Yadong Qi <qiyadong2010@gmail.com>
Author: q00251598 <qiyadong@huawei.com>
Author: Tathagata Das <tathagata.das1565@gmail.com>
Author: wangfei <wangfei1@huawei.com>
Closes#3865 from tdas/print-num and squashes the following commits:
cd34e9e [Tathagata Das] Fix bug
7c09f16 [Tathagata Das] Merge remote-tracking branch 'apache-github/master' into HEAD
bb35d1a [Yadong Qi] Update MimaExcludes.scala
f8098ca [Yadong Qi] Update MimaExcludes.scala
f6ac3cb [Yadong Qi] Update MimaExcludes.scala
e4ed897 [Yadong Qi] Update MimaExcludes.scala
3b9d5cf [wangfei] fix conflicts
ec8a3af [q00251598] move to Spark 1.3
26a70c0 [q00251598] extend the Python DStream's print
b589a4b [q00251598] add another print function
Creates a top level directory script (as `build/mvn`) to automatically download zinc and the specific version of scala used to easily build spark. This will also download and install maven if the user doesn't already have it and all packages are hosted under the `build/` directory. Tested on both Linux and OSX OS's and both work. All commands pass through to the maven binary so it acts exactly as a traditional maven call would.
Author: Brennon York <brennon.york@capitalone.com>
Closes#3707 from brennonyork/SPARK-4501 and squashes the following commits:
0e5a0e4 [Brennon York] minor incorrect doc verbage (with -> this)
9b79e38 [Brennon York] fixed merge conflicts with dev/run-tests, properly quoted args in sbt/sbt, fixed bug where relative paths would fail if passed in from build/mvn
d2d41b6 [Brennon York] added blurb about leverging zinc with build/mvn
b979c58 [Brennon York] updated the merge conflict
c5634de [Brennon York] updated documentation to overview build/mvn, updated all points where sbt/sbt was referenced with build/sbt
b8437ba [Brennon York] set progress bars for curl and wget when not run on jenkins, no progress bar when run on jenkins, moved sbt script to build/sbt, wrote stub and warning under sbt/sbt which calls build/sbt, modified build/sbt to use the correct directory, fixed bug in build/sbt-launch-lib.bash to correctly pull the sbt version
be11317 [Brennon York] added switch to silence download progress only if AMPLAB_JENKINS is set
28d0a99 [Brennon York] updated to remove the python dependency, uses grep instead
7e785a6 [Brennon York] added silent and quiet flags to curl and wget respectively, added single echo output to denote start of a download if download is needed
14a5da0 [Brennon York] removed unnecessary zinc output on startup
1af4a94 [Brennon York] fixed bug with uppercase vs lowercase variable
3e8b9b3 [Brennon York] updated to properly only restart zinc if it was freshly installed
a680d12 [Brennon York] Added comments to functions and tested various mvn calls
bb8cc9d [Brennon York] removed package files
ef017e6 [Brennon York] removed OS complexities, setup generic install_app call, removed extra file complexities, removed help, removed forced install (defaults now), removed double-dash from cli
07bf018 [Brennon York] Updated to specifically handle pulling down the correct scala version
f914dea [Brennon York] Beginning final portions of localized scala home
69c4e44 [Brennon York] working linux and osx installers for purely local mvn build
4a1609c [Brennon York] finalizing working linux install for maven to local ./build/apache-maven folder
cbfcc68 [Brennon York] Changed the default sbt/sbt to build/sbt and added a build/mvn which will automatically download, install, and execute maven with zinc for easier build capability
This PR modifies the python `SchemaRDD` to use `sample()` and `takeSample()` from Scala instead of the slower python implementations from `rdd.py`. This is worthwhile because the `Row`'s are already serialized as Java objects.
In order to use the faster `takeSample()`, a `takeSampleToPython()` method was implemented in `SchemaRDD.scala` following the pattern of `collectToPython()`.
Author: jbencook <jbenjamincook@gmail.com>
Author: J. Benjamin Cook <jbenjamincook@gmail.com>
Closes#3764 from jbencook/master and squashes the following commits:
6fbc769 [J. Benjamin Cook] [SPARK-4860][pyspark][sql] fixing sloppy indentation for takeSampleToPython() arguments
5170da2 [J. Benjamin Cook] [SPARK-4860][pyspark][sql] fixing typo: from RDD to SchemaRDD
de22f70 [jbencook] [SPARK-4860][pyspark][sql] using sample() method from JavaSchemaRDD
b916442 [jbencook] [SPARK-4860][pyspark][sql] adding sample() to JavaSchemaRDD
020cbdf [jbencook] [SPARK-4860][pyspark][sql] using Scala implementations of `sample()` and `takeSample()`
Modify python annotations for sphinx. There is no change to build process from.
https://github.com/apache/spark/blob/master/docs/README.md
Author: lewuathe <lewuathe@me.com>
Closes#3685 from Lewuathe/sphinx-tag-for-pydoc and squashes the following commits:
88a0fd9 [lewuathe] [SPARK-4822] Fix DevelopApi and WARN tags
3d7a398 [lewuathe] [SPARK-4822] Use sphinx tags for Python doc annotations
+ small doc edit
+ include edit to make IntelliJ happy
CC: davies mengxr
Note to davies -- this does not fix the "WARNING: Literal block expected; none found." warnings since that seems to involve spacing which IntelliJ does not like. (Those warnings occur when generating the Python docs.)
Author: Joseph K. Bradley <joseph@databricks.com>
Closes#3669 from jkbradley/python-warnings and squashes the following commits:
4587868 [Joseph K. Bradley] fixed warning
8cb073c [Joseph K. Bradley] Updated based on davies recommendation
c51eca4 [Joseph K. Bradley] Updated rst file for pyspark.mllib.rand doc. Small doc edit. Small include edit to make IntelliJ happy.
This PR brings support of using StructType(and other hashable types) as key in MapType.
Author: Davies Liu <davies@databricks.com>
Closes#3714 from davies/fix_struct_in_map and squashes the following commits:
68585d7 [Davies Liu] fix primitive types in MapType
9601534 [Davies Liu] support StructType as key in MapType
This PR tests the pyspark Chi-squared hypothesis test from this commit: c8abddc516 and moves some of the error messaging in to python.
It is a port of the Scala tests here: [HypothesisTestSuite.scala](https://github.com/apache/spark/blob/master/mllib/src/test/scala/org/apache/spark/mllib/stat/HypothesisTestSuite.scala)
Hopefully, SPARK-2980 can be closed.
Author: jbencook <jbenjamincook@gmail.com>
Closes#3679 from jbencook/master and squashes the following commits:
44078e0 [jbencook] checking that bad input throws the correct exceptions
f12ee10 [jbencook] removing checks for ValueError since input tests are on the Scala side
7536cf1 [jbencook] removing python checks for invalid input
a17ee84 [jbencook] [SPARK-2980][mllib] adding unit tests for the pyspark chi-squared test
3aeb0d9 [jbencook] [SPARK-2980][mllib] bringing Chi-squared error messages to the python side
UTF8Deserializer can not be used in BatchedSerializer, so always use PickleSerializer() when change batchSize in zip().
Also, if two RDD have the same batch size already, they did not need re-serialize any more.
Author: Davies Liu <davies@databricks.com>
Closes#3706 from davies/fix_4841 and squashes the following commits:
20ce3a3 [Davies Liu] fix bug in _reserialize()
e3ebf7c [Davies Liu] add comment
379d2c8 [Davies Liu] fix zip with textFile()
I improved `IDFModel.transform` to allow using a single vector.
[[SPARK-4494] IDFModel.transform() add support for single vector - ASF JIRA](https://issues.apache.org/jira/browse/SPARK-4494)
Author: Yuu ISHIKAWA <yuu.ishikawa@gmail.com>
Closes#3603 from yu-iskw/idf and squashes the following commits:
256ff3d [Yuu ISHIKAWA] Fix typo
a3bf566 [Yuu ISHIKAWA] - Fix typo - Optimize import order - Aggregate the assertion tests - Modify `IDFModel.transform` API for pyspark
d25e49b [Yuu ISHIKAWA] Add the implementation of `IDFModel.transform` for a term frequency vector
Major changes:
* Added programming guide sections for tree ensembles
* Added examples for tree ensembles
* Updated DecisionTree programming guide with more info on parameters
* **API change**: Standardized the tree parameter for the number of classes (for classification)
Minor changes:
* Updated decision tree documentation
* Updated existing tree and tree ensemble examples
* Use train/test split, and compute test error instead of training error.
* Fixed decision_tree_runner.py to actually use the number of classes it computes from data. (small bug fix)
Note: I know this is a lot of lines, but most is covered by:
* Programming guide sections for gradient boosting and random forests. (The changes are probably best viewed by generating the docs locally.)
* New examples (which were copied from the programming guide)
* The "numClasses" renaming
I have run all examples and relevant unit tests.
CC: mengxr manishamde codedeft
Author: Joseph K. Bradley <joseph@databricks.com>
Author: Joseph K. Bradley <joseph.kurata.bradley@gmail.com>
Closes#3461 from jkbradley/ensemble-docs and squashes the following commits:
70a75f3 [Joseph K. Bradley] updated forest vs boosting comparison
d1de753 [Joseph K. Bradley] Added note about toString and toDebugString for DecisionTree to migration guide
8e87f8f [Joseph K. Bradley] Combined GBT and RandomForest guides into one ensembles guide
6fab846 [Joseph K. Bradley] small fixes based on review
b9f8576 [Joseph K. Bradley] updated decision tree doc
375204c [Joseph K. Bradley] fixed python style
2b60b6e [Joseph K. Bradley] merged Java RandomForest examples into 1 file. added header. Fixed small bug in same example in the programming guide.
706d332 [Joseph K. Bradley] updated python DT runner to print full model if it is small
c76c823 [Joseph K. Bradley] added migration guide for mllib
abe5ed7 [Joseph K. Bradley] added examples for random forest in Java and Python to examples folder
07fc11d [Joseph K. Bradley] Renamed numClassesForClassification to numClasses everywhere in trees and ensembles. This is a breaking API change, but it was necessary to correct an API inconsistency in Spark 1.1 (where Python DecisionTree used numClasses but Scala used numClassesForClassification).
cdfdfbc [Joseph K. Bradley] added examples for GBT
6372a2b [Joseph K. Bradley] updated decision tree examples to use random split. tested all of them.
ad3e695 [Joseph K. Bradley] added gbt and random forest to programming guide. still need to update their examples
Re-implement the Python broadcast using file:
1) serialize the python object using cPickle, write into disks.
2) Create a wrapper in JVM (for the dumped file), it read data from during serialization
3) Using TorrentBroadcast or HttpBroadcast to transfer the data (compressed) into executors
4) During deserialization, writing the data into disk.
5) Passing the path into Python worker, read data from disk and unpickle it into python object, until the first access.
It fixes the performance regression introduced in #2659, has similar performance as 1.1, but support object larger than 2G, also improve the memory efficiency (only one compressed copy in driver and executor).
Testing with a 500M broadcast and 4 tasks (excluding the benefit from reused worker in 1.2):
name | 1.1 | 1.2 with this patch | improvement
---------|--------|---------|--------
python-broadcast-w-bytes | 25.20 | 9.33 | 170.13% |
python-broadcast-w-set | 4.13 | 4.50 | -8.35% |
Testing with 100 tasks (16 CPUs):
name | 1.1 | 1.2 with this patch | improvement
---------|--------|---------|--------
python-broadcast-w-bytes | 38.16 | 8.40 | 353.98%
python-broadcast-w-set | 23.29 | 9.59 | 142.80%
Author: Davies Liu <davies@databricks.com>
Closes#3417 from davies/pybroadcast and squashes the following commits:
50a58e0 [Davies Liu] address comments
b98de1d [Davies Liu] disable gc while unpickle
e5ee6b9 [Davies Liu] support large string
09303b8 [Davies Liu] read all data into memory
dde02dd [Davies Liu] improve performance of python broadcast
The Row object is created on the fly once the field is accessed, so we should access them by getattr() in asDict(0
Author: Davies Liu <davies@databricks.com>
Closes#3434 from davies/fix_asDict and squashes the following commits:
b20f1e7 [Davies Liu] fix asDict() with nested Row()
This PR change the underline array of DenseVector to numpy.ndarray to avoid the conversion, because most of the users will using numpy.array.
It also improve the serialization of DenseVector.
Before this change:
trial | trainingTime | testTime
-------|--------|--------
0 | 5.126 | 1.786
1 |2.698 |1.693
After the change:
trial | trainingTime | testTime
-------|--------|--------
0 |4.692 |0.554
1 |2.307 |0.525
This could partially fix the performance regression during test.
Author: Davies Liu <davies@databricks.com>
Closes#3420 from davies/ser2 and squashes the following commits:
0e1e6f3 [Davies Liu] fix tests
426f5db [Davies Liu] impove toArray()
44707ec [Davies Liu] add name for ISO-8859-1
fa7d791 [Davies Liu] address comments
1cfb137 [Davies Liu] handle zero sparse vector
2548ee2 [Davies Liu] fix tests
9e6389d [Davies Liu] bugfix
470f702 [Davies Liu] speed up DenseMatrix
f0d3c40 [Davies Liu] speedup SparseVector
ef6ce70 [Davies Liu] speed up dense vector
The Pyrolite is pretty slow (comparing to the adhoc serializer in 1.1), it cause much performance regression in 1.2, because we cache the serialized Python object in JVM, deserialize them into Java object in each step.
This PR change to cache the deserialized JavaRDD instead of PythonRDD to avoid the deserialization of Pyrolite. It should have similar memory usage as before, but much faster.
Author: Davies Liu <davies@databricks.com>
Closes#3397 from davies/cache and squashes the following commits:
7f6e6ce [Davies Liu] Update -> Updater
4b52edd [Davies Liu] using named argument
63b984e [Davies Liu] fix
7da0332 [Davies Liu] add unpersist()
dff33e1 [Davies Liu] address comments
c2bdfc2 [Davies Liu] refactor
d572f00 [Davies Liu] Merge branch 'master' into cache
f1063e1 [Davies Liu] cache serialized java object
In RDDSampler, it try use numpy to gain better performance for possion(), but the number of call of random() is only (1+faction) * N in the pure python implementation of possion(), so there is no much performance gain from numpy.
numpy is not a dependent of pyspark, so it maybe introduce some problem, such as there is no numpy installed in slaves, but only installed master, as reported in SPARK-927.
It also complicate the code a lot, so we may should remove numpy from RDDSampler.
I also did some benchmark to verify that:
```
>>> from pyspark.mllib.random import RandomRDDs
>>> rdd = RandomRDDs.uniformRDD(sc, 1 << 20, 1).cache()
>>> rdd.count() # cache it
>>> rdd.sample(True, 0.9).count() # measure this line
```
the results:
|withReplacement | random | numpy.random |
------- | ------------ | -------
|True | 1.5 s| 1.4 s|
|False| 0.6 s | 0.8 s|
closes#2313
Note: this patch including some commits that not mirrored to github, it will be OK after it catches up.
Author: Davies Liu <davies@databricks.com>
Author: Xiangrui Meng <meng@databricks.com>
Closes#3351 from davies/numpy and squashes the following commits:
5c438d7 [Davies Liu] fix comment
c5b9252 [Davies Liu] Merge pull request #1 from mengxr/SPARK-4477
98eb31b [Xiangrui Meng] make poisson sampling slightly faster
ee17d78 [Davies Liu] remove = for float
13f7b05 [Davies Liu] Merge branch 'master' of http://git-wip-us.apache.org/repos/asf/spark into numpy
f583023 [Davies Liu] fix tests
51649f5 [Davies Liu] remove numpy in RDDSampler
78bf997 [Davies Liu] fix tests, do not use numpy in randomSplit, no performance gain
f5fdf63 [Davies Liu] fix bug with int in weights
4dfa2cd [Davies Liu] refactor
f866bcf [Davies Liu] remove unneeded change
c7a2007 [Davies Liu] switch to python implementation
95a48ac [Davies Liu] Merge branch 'master' of github.com:apache/spark into randomSplit
0d9b256 [Davies Liu] refactor
1715ee3 [Davies Liu] address comments
41fce54 [Davies Liu] randomSplit()
```
class RandomForestModel
| A model trained by RandomForest
|
| numTrees(self)
| Get number of trees in forest.
|
| predict(self, x)
| Predict values for a single data point or an RDD of points using the model trained.
|
| toDebugString(self)
| Full model
|
| totalNumNodes(self)
| Get total number of nodes, summed over all trees in the forest.
|
class RandomForest
| trainClassifier(cls, data, numClassesForClassification, categoricalFeaturesInfo, numTrees, featureSubsetStrategy='auto', impurity='gini', maxDepth=4, maxBins=32, seed=None):
| Method to train a decision tree model for binary or multiclass classification.
|
| :param data: Training dataset: RDD of LabeledPoint.
| Labels should take values {0, 1, ..., numClasses-1}.
| :param numClassesForClassification: number of classes for classification.
| :param categoricalFeaturesInfo: Map storing arity of categorical features.
| E.g., an entry (n -> k) indicates that feature n is categorical
| with k categories indexed from 0: {0, 1, ..., k-1}.
| :param numTrees: Number of trees in the random forest.
| :param featureSubsetStrategy: Number of features to consider for splits at each node.
| Supported: "auto" (default), "all", "sqrt", "log2", "onethird".
| If "auto" is set, this parameter is set based on numTrees:
| if numTrees == 1, set to "all";
| if numTrees > 1 (forest) set to "sqrt".
| :param impurity: Criterion used for information gain calculation.
| Supported values: "gini" (recommended) or "entropy".
| :param maxDepth: Maximum depth of the tree. E.g., depth 0 means 1 leaf node; depth 1 means
| 1 internal node + 2 leaf nodes. (default: 4)
| :param maxBins: maximum number of bins used for splitting features (default: 100)
| :param seed: Random seed for bootstrapping and choosing feature subsets.
| :return: RandomForestModel that can be used for prediction
|
| trainRegressor(cls, data, categoricalFeaturesInfo, numTrees, featureSubsetStrategy='auto', impurity='variance', maxDepth=4, maxBins=32, seed=None):
| Method to train a decision tree model for regression.
|
| :param data: Training dataset: RDD of LabeledPoint.
| Labels are real numbers.
| :param categoricalFeaturesInfo: Map storing arity of categorical features.
| E.g., an entry (n -> k) indicates that feature n is categorical
| with k categories indexed from 0: {0, 1, ..., k-1}.
| :param numTrees: Number of trees in the random forest.
| :param featureSubsetStrategy: Number of features to consider for splits at each node.
| Supported: "auto" (default), "all", "sqrt", "log2", "onethird".
| If "auto" is set, this parameter is set based on numTrees:
| if numTrees == 1, set to "all";
| if numTrees > 1 (forest) set to "onethird".
| :param impurity: Criterion used for information gain calculation.
| Supported values: "variance".
| :param maxDepth: Maximum depth of the tree. E.g., depth 0 means 1 leaf node; depth 1 means
| 1 internal node + 2 leaf nodes.(default: 4)
| :param maxBins: maximum number of bins used for splitting features (default: 100)
| :param seed: Random seed for bootstrapping and choosing feature subsets.
| :return: RandomForestModel that can be used for prediction
|
```
Author: Davies Liu <davies@databricks.com>
Closes#3320 from davies/forest and squashes the following commits:
8003dfc [Davies Liu] reorder
53cf510 [Davies Liu] fix docs
4ca593d [Davies Liu] fix docs
e0df852 [Davies Liu] fix docs
0431746 [Davies Liu] rebased
2b6f239 [Davies Liu] Merge branch 'master' of github.com:apache/spark into forest
885abee [Davies Liu] address comments
dae7fc0 [Davies Liu] address comments
89a000f [Davies Liu] fix docs
565d476 [Davies Liu] add python api for random forest
If there some big broadcasts (or other object) in Python worker, the free memory could be used for sorting will be too small, then it will keep spilling small files into disks, finally failed with too many open files.
This PR try to delay the spilling until the used memory goes over limit and start to increase since last spilling, it will increase the size of spilling files, improve the stability and performance in this cases. (We also do this in ExternalAggregator).
Author: Davies Liu <davies@databricks.com>
Closes#3252 from davies/sort and squashes the following commits:
711fb6c [Davies Liu] improve sort spilling
This commit should be merged for 1.2 release.
cc tdas
Author: Ken Takagiwa <ugw.gi.world@gmail.com>
Closes#3311 from giwa/patch-3 and squashes the following commits:
ab474a8 [Ken Takagiwa] [DOC][PySpark][Streaming] Fix docstring for sphinx
```
pyspark.RDD.randomSplit(self, weights, seed=None)
Randomly splits this RDD with the provided weights.
:param weights: weights for splits, will be normalized if they don't sum to 1
:param seed: random seed
:return: split RDDs in an list
>>> rdd = sc.parallelize(range(10), 1)
>>> rdd1, rdd2, rdd3 = rdd.randomSplit([0.4, 0.6, 1.0], 11)
>>> rdd1.collect()
[3, 6]
>>> rdd2.collect()
[0, 5, 7]
>>> rdd3.collect()
[1, 2, 4, 8, 9]
```
Author: Davies Liu <davies@databricks.com>
Closes#3193 from davies/randomSplit and squashes the following commits:
78bf997 [Davies Liu] fix tests, do not use numpy in randomSplit, no performance gain
f5fdf63 [Davies Liu] fix bug with int in weights
4dfa2cd [Davies Liu] refactor
f866bcf [Davies Liu] remove unneeded change
c7a2007 [Davies Liu] switch to python implementation
95a48ac [Davies Liu] Merge branch 'master' of github.com:apache/spark into randomSplit
0d9b256 [Davies Liu] refactor
1715ee3 [Davies Liu] address comments
41fce54 [Davies Liu] randomSplit()
This patch will bring support for broadcasting objects larger than 2G.
pickle, zlib, FrameSerializer and Array[Byte] all can not support objects larger than 2G, so this patch introduce LargeObjectSerializer to serialize broadcast objects, the object will be serialized and compressed into small chunks, it also change the type of Broadcast[Array[Byte]]] into Broadcast[Array[Array[Byte]]]].
Testing for support broadcast objects larger than 2G is slow and memory hungry, so this is tested manually, could be added into SparkPerf.
Author: Davies Liu <davies@databricks.com>
Author: Davies Liu <davies.liu@gmail.com>
Closes#2659 from davies/huge and squashes the following commits:
7b57a14 [Davies Liu] add more tests for broadcast
28acff9 [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
a2f6a02 [Davies Liu] bug fix
4820613 [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
5875c73 [Davies Liu] address comments
10a349b [Davies Liu] address comments
0c33016 [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
6182c8f [Davies Liu] Merge branch 'master' into huge
d94b68f [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
2514848 [Davies Liu] address comments
fda395b [Davies Liu] Merge branch 'master' of github.com:apache/spark into huge
1c2d928 [Davies Liu] fix scala style
091b107 [Davies Liu] broadcast objects larger than 2G
```
class LogisticRegressionWithLBFGS
| train(cls, data, iterations=100, initialWeights=None, corrections=10, tolerance=0.0001, regParam=0.01, intercept=False)
| Train a logistic regression model on the given data.
|
| :param data: The training data, an RDD of LabeledPoint.
| :param iterations: The number of iterations (default: 100).
| :param initialWeights: The initial weights (default: None).
| :param regParam: The regularizer parameter (default: 0.01).
| :param regType: The type of regularizer used for training
| our model.
| :Allowed values:
| - "l1" for using L1 regularization
| - "l2" for using L2 regularization
| - None for no regularization
| (default: "l2")
| :param intercept: Boolean parameter which indicates the use
| or not of the augmented representation for
| training data (i.e. whether bias features
| are activated or not).
| :param corrections: The number of corrections used in the LBFGS update (default: 10).
| :param tolerance: The convergence tolerance of iterations for L-BFGS (default: 1e-4).
|
| >>> data = [
| ... LabeledPoint(0.0, [0.0, 1.0]),
| ... LabeledPoint(1.0, [1.0, 0.0]),
| ... ]
| >>> lrm = LogisticRegressionWithLBFGS.train(sc.parallelize(data))
| >>> lrm.predict([1.0, 0.0])
| 1
| >>> lrm.predict([0.0, 1.0])
| 0
| >>> lrm.predict(sc.parallelize([[1.0, 0.0], [0.0, 1.0]])).collect()
| [1, 0]
```
Author: Davies Liu <davies@databricks.com>
Closes#3307 from davies/lbfgs and squashes the following commits:
34bd986 [Davies Liu] Merge branch 'master' of http://git-wip-us.apache.org/repos/asf/spark into lbfgs
5a945a6 [Davies Liu] address comments
941061b [Davies Liu] Merge branch 'master' of github.com:apache/spark into lbfgs
03e5543 [Davies Liu] add it to docs
ed2f9a8 [Davies Liu] add regType
76cd1b6 [Davies Liu] reorder arguments
4429a74 [Davies Liu] Update classification.py
9252783 [Davies Liu] python api for LogisticRegressionWithLBFGS
In PySpark, ALS can take an RDD of (user, product, rating) tuples as input. However, model.predict outputs an RDD of Rating. So on the input side, users can use r[0], r[1], r[2], while on the output side, users have to use r.user, r.product, r.rating. We should allow lookup by index in Rating by making Rating a namedtuple.
davies
<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/3261)
<!-- Reviewable:end -->
Author: Xiangrui Meng <meng@databricks.com>
Closes#3261 from mengxr/SPARK-4396 and squashes the following commits:
543aef0 [Xiangrui Meng] use named tuple to implement ALS
0b61bae [Xiangrui Meng] Merge remote-tracking branch 'apache/master' into SPARK-4396
d3bd7d4 [Xiangrui Meng] allow lookup by index in Python's Rating
This PR add setThrehold() and clearThreshold() for LogisticRegressionModel and SVMModel, also support RDD of vector in LogisticRegressionModel.predict(), SVNModel.predict() and NaiveBayes.predict()
Author: Davies Liu <davies@databricks.com>
Closes#3305 from davies/setThreshold and squashes the following commits:
d0b835f [Davies Liu] Merge branch 'master' of github.com:apache/spark into setThreshold
e4acd76 [Davies Liu] address comments
2231a5f [Davies Liu] bugfix
7bd9009 [Davies Liu] address comments
0b0a8a7 [Davies Liu] address comments
c1e5573 [Davies Liu] improve classification
When JVM is started in a Python process, it should exit once the stdin is closed.
test: add spark.driver.memory in conf/spark-defaults.conf
```
daviesdm:~/work/spark$ cat conf/spark-defaults.conf
spark.driver.memory 8g
daviesdm:~/work/spark$ bin/pyspark
>>> quit
daviesdm:~/work/spark$ jps
4931 Jps
286
daviesdm:~/work/spark$ python wc.py
943738
0.719928026199
daviesdm:~/work/spark$ jps
286
4990 Jps
```
Author: Davies Liu <davies@databricks.com>
Closes#3274 from davies/exit and squashes the following commits:
df0e524 [Davies Liu] address comments
ce8599c [Davies Liu] address comments
050651f [Davies Liu] JVM should exit after Python exit
`sc.parallelize(range(1 << 20), 1).count()` may take 15 seconds to finish and the rdd object stores the entire list, making task size very large. This PR adds a specialized version for xrange.
JoshRosen davies
Author: Xiangrui Meng <meng@databricks.com>
Closes#3264 from mengxr/SPARK-4398 and squashes the following commits:
8953c41 [Xiangrui Meng] follow davies' suggestion
cbd58e3 [Xiangrui Meng] specialize sc.parallelize(xrange)
The current default regParam is 1.0 and regType is claimed to be none in Python (but actually it is l2), while regParam = 0.0 and regType is L2 in Scala. We should make the default values consistent. This PR sets the default regType to L2 and regParam to 0.01. Note that the default regParam value in LIBLINEAR (and hence scikit-learn) is 1.0. However, we use average loss instead of total loss in our formulation. Hence regParam=1.0 is definitely too heavy.
In LinearRegression, we set regParam=0.0 and regType=None, because we have separate classes for Lasso and Ridge, both of which use regParam=0.01 as the default.
davies atalwalkar
Author: Xiangrui Meng <meng@databricks.com>
Closes#3232 from mengxr/SPARK-4372 and squashes the following commits:
9979837 [Xiangrui Meng] update Ridge/Lasso to use default regParam 0.01 cast input arguments
d3ba096 [Xiangrui Meng] change 'none' back to None
1909a6e [Xiangrui Meng] change default regParam to 0.01 and regType to L2 in LR and SVM
This PR rename random.py to rand.py to avoid the side affects of conflict with random module, but still keep the same interface as before.
```
>>> from pyspark.mllib.random import RandomRDDs
```
```
$ pydoc pyspark.mllib.random
Help on module random in pyspark.mllib:
NAME
random - Python package for random data generation.
FILE
/Users/davies/work/spark/python/pyspark/mllib/rand.py
CLASSES
__builtin__.object
pyspark.mllib.random.RandomRDDs
class RandomRDDs(__builtin__.object)
| Generator methods for creating RDDs comprised of i.i.d samples from
| some distribution.
|
| Static methods defined here:
|
| normalRDD(sc, size, numPartitions=None, seed=None)
```
cc mengxr
reference link: http://xion.org.pl/2012/05/06/hacking-python-imports/
Author: Davies Liu <davies@databricks.com>
Closes#3216 from davies/random and squashes the following commits:
7ac4e8b [Davies Liu] rename random.py to rand.py
Fix TreeModel.predict() with RDD, added tests for it.
(Also checked that other models don't have this issue)
Author: Davies Liu <davies@databricks.com>
Closes#3230 from davies/predict and squashes the following commits:
81172aa [Davies Liu] fix predict
This PR check all of the existing Python MLlib API to make sure that numpy.array is supported as Vector (also RDD of numpy.array).
It also improve some docstring and doctest.
cc mateiz mengxr
Author: Davies Liu <davies@databricks.com>
Closes#3189 from davies/numpy and squashes the following commits:
d5057c4 [Davies Liu] fix tests
6987611 [Davies Liu] support numpy.array for all MLlib API
SPARK-1553 added alternating nonnegative least squares to MLLib, however it's not possible to access it via the python API. This pull request resolves that.
Author: Michelangelo D'Agostino <mdagostino@civisanalytics.com>
Closes#3095 from mdagost/python_nmf and squashes the following commits:
a6743ad [Michelangelo D'Agostino] Use setters instead of static methods in PythonMLLibAPI. Remove the new static methods I added. Set seed in tests. Change ratings to ratingsRDD in both train and trainImplicit for consistency.
7cffd39 [Michelangelo D'Agostino] Swapped nonnegative and seed in a few more places.
3fdc851 [Michelangelo D'Agostino] Moved seed to the end of the python parameter list.
bdcc154 [Michelangelo D'Agostino] Change seed type to java.lang.Long so that it can handle null.
cedf043 [Michelangelo D'Agostino] Added in ability to set the seed from python and made that play nice with the nonnegative changes. Also made the python ALS tests more exact.
a72fdc9 [Michelangelo D'Agostino] Expose nonnegative ALS in the python API.
This PR fix sortBy()/sortByKey() on empty RDD.
This should be back ported into 1.1/1.2
Author: Davies Liu <davies@databricks.com>
Closes#3162 from davies/fix_sort and squashes the following commits:
84f64b7 [Davies Liu] add tests
52995b5 [Davies Liu] fix sortByKey() on empty RDD
add binaryFiles() and binaryRecords() in Python
```
binaryFiles(self, path, minPartitions=None):
:: Developer API ::
Read a directory of binary files from HDFS, a local file system
(available on all nodes), or any Hadoop-supported file system URI
as a byte array. Each file is read as a single record and returned
in a key-value pair, where the key is the path of each file, the
value is the content of each file.
Note: Small files are preferred, large file is also allowable, but
may cause bad performance.
binaryRecords(self, path, recordLength):
Load data from a flat binary file, assuming each record is a set of numbers
with the specified numerical format (see ByteBuffer), and the number of
bytes per record is constant.
:param path: Directory to the input data files
:param recordLength: The length at which to split the records
```
Author: Davies Liu <davies@databricks.com>
Closes#3078 from davies/binary and squashes the following commits:
cd0bdbd [Davies Liu] Merge branch 'master' of github.com:apache/spark into binary
3aa349b [Davies Liu] add experimental notes
24e84b6 [Davies Liu] Merge branch 'master' of github.com:apache/spark into binary
5ceaa8a [Davies Liu] Merge branch 'master' of github.com:apache/spark into binary
1900085 [Davies Liu] bugfix
bb22442 [Davies Liu] add binaryFiles and binaryRecords in Python
```
pyspark.mllib.stat.StatisticschiSqTest(observed, expected=None)
:: Experimental ::
If `observed` is Vector, conduct Pearson's chi-squared goodness
of fit test of the observed data against the expected distribution,
or againt the uniform distribution (by default), with each category
having an expected frequency of `1 / len(observed)`.
(Note: `observed` cannot contain negative values)
If `observed` is matrix, conduct Pearson's independence test on the
input contingency matrix, which cannot contain negative entries or
columns or rows that sum up to 0.
If `observed` is an RDD of LabeledPoint, conduct Pearson's independence
test for every feature against the label across the input RDD.
For each feature, the (feature, label) pairs are converted into a
contingency matrix for which the chi-squared statistic is computed.
All label and feature values must be categorical.
:param observed: it could be a vector containing the observed categorical
counts/relative frequencies, or the contingency matrix
(containing either counts or relative frequencies),
or an RDD of LabeledPoint containing the labeled dataset
with categorical features. Real-valued features will be
treated as categorical for each distinct value.
:param expected: Vector containing the expected categorical counts/relative
frequencies. `expected` is rescaled if the `expected` sum
differs from the `observed` sum.
:return: ChiSquaredTest object containing the test statistic, degrees
of freedom, p-value, the method used, and the null hypothesis.
```
Author: Davies Liu <davies@databricks.com>
Closes#3091 from davies/his and squashes the following commits:
145d16c [Davies Liu] address comments
0ab0764 [Davies Liu] fix float
5097d54 [Davies Liu] add Hypothesis test Python API
