[
https://issues.apache.org/jira/browse/SPARK-12944?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15119976#comment-15119976
]
Seth Hendrickson commented on SPARK-12944:
------------------------------------------
Based on what I have seen, I think the fix is more involved... could you
elaborate on your proposed solution?
We could "fix" it in addGrid by checking that all elements of {{values}} are
instances of {{param.expectedType}}, but I don't think this really gets at the
root of the problem. First, this would still throw an error with ints that are
expected to be floats, which is unsatisfactory since [PR
9581|https://github.com/apache/spark/pull/9581] adds a type conversion
mechanism specifically for this purpose. Second, the real problem is that
params passed through the {{Estimator.fit}} and {{Transformer.transform}}
methods use the {{Params.copy}} method to circumvent the set methods and type
checking entirely! By modifying the {{_paramMap}} dictionary directly, the copy
method allows ml pipeline elements to contain params that have no context (like
a HashingTF containing a Naive Bayes smoothing parameter in its param map as in
the example above).
I think the correct way to do this is to change the {{copy}} method to call the
{{_set}} function instead of directly modifying {{_paramMap}}. That way we
ensure that the {{_paramMap}} only contains parameters that belong to that
class and that type checking is performed. I am working on a PR with the method
I described, but I am open to feedback. Appreciate any thoughts on this.
> CrossValidator doesn't accept a Pipeline as an estimator
> --------------------------------------------------------
>
> Key: SPARK-12944
> URL: https://issues.apache.org/jira/browse/SPARK-12944
> Project: Spark
> Issue Type: Bug
> Components: ML, PySpark
> Affects Versions: 1.6.0
> Environment: spark-1.6.0-bin-hadoop2.6
> Python 3.4.4 :: Anaconda 2.4.1
> Reporter: John Hogue
> Priority: Minor
>
> Pipeline is supposed to act as an estimator which CrossValidator currently
> throws error.
> {code}
> from pyspark.ml.evaluation import MulticlassClassificationEvaluator
> from pyspark.ml.tuning import ParamGridBuilder
> from pyspark.ml.tuning import CrossValidator
> # Configure an ML pipeline, which consists of tree stages: tokenizer,
> hashingTF, and nb.
> tokenizer = Tokenizer(inputCol="text", outputCol="words")
> hashingTF = HashingTF(inputCol=tokenizer.getOutputCol(), outputCol="features")
> nb = NaiveBayes()
> pipeline = Pipeline(stages=[tokenizer, hashingTF, nb])
> paramGrid = ParamGridBuilder().addGrid(nb.smoothing, [0, 1]).build()
> cv = CrossValidator(estimator=pipeline,
> estimatorParamMaps=paramGrid,
> evaluator=MulticlassClassificationEvaluator(),
> numFolds=4)
> cvModel = cv.fit(training_df)
> {code}
> Sample dataset can be found here:
> https://github.com/dreyco676/nlp_spark/blob/master/data.zip
> The file can be converted to a DataFrame with:
> {code}
> # Load precleaned training set
> training_rdd = sc.textFile("data/clean_training.txt")
> parts_rdd = training_rdd.map(lambda l: l.split("\t"))
> # Filter bad rows out
> garantee_col_rdd = parts_rdd.filter(lambda l: len(l) == 3)
> typed_rdd = garantee_col_rdd.map(lambda p: (p[0], p[1], float(p[2])))
> # Create DataFrame
> training_df = sqlContext.createDataFrame(typed_rdd, ["id", "text", "label"])
> {code}
> Running the pipeline throws the following stack trace:
> {code}
> ---------------------------------------------------------------------------Py4JJavaError
> Traceback (most recent call
> last)<ipython-input-3-34e9e27acada> in <module>()
> 17 numFolds=4)
> 18
> ---> 19 cvModel = cv.fit(training_df)
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/ml/pipeline.py in
> fit(self, dataset, params)
> 67 return self.copy(params)._fit(dataset)
> 68 else:
> ---> 69 return self._fit(dataset)
> 70 else:
> 71 raise ValueError("Params must be either a param map or a
> list/tuple of param maps, "
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/ml/tuning.py in
> _fit(self, dataset)
> 237 train = df.filter(~condition)
> 238 for j in range(numModels):
> --> 239 model = est.fit(train, epm[j])
> 240 # TODO: duplicate evaluator to take extra params from
> input
> 241 metric = eva.evaluate(model.transform(validation,
> epm[j]))
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/ml/pipeline.py in
> fit(self, dataset, params)
> 65 elif isinstance(params, dict):
> 66 if params:
> ---> 67 return self.copy(params)._fit(dataset)
> 68 else:
> 69 return self._fit(dataset)
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/ml/pipeline.py in
> _fit(self, dataset)
> 211 dataset = stage.transform(dataset)
> 212 else: # must be an Estimator
> --> 213 model = stage.fit(dataset)
> 214 transformers.append(model)
> 215 if i < indexOfLastEstimator:
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/ml/pipeline.py in
> fit(self, dataset, params)
> 67 return self.copy(params)._fit(dataset)
> 68 else:
> ---> 69 return self._fit(dataset)
> 70 else:
> 71 raise ValueError("Params must be either a param map or a
> list/tuple of param maps, "
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/ml/wrapper.py in
> _fit(self, dataset)
> 130
> 131 def _fit(self, dataset):
> --> 132 java_model = self._fit_java(dataset)
> 133 return self._create_model(java_model)
> 134
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/ml/wrapper.py in
> _fit_java(self, dataset)
> 126 :return: fitted Java model
> 127 """
> --> 128 self._transfer_params_to_java()
> 129 return self._java_obj.fit(dataset._jdf)
> 130
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/ml/wrapper.py in
> _transfer_params_to_java(self)
> 80 for param in self.params:
> 81 if param in paramMap:
> ---> 82 pair = self._make_java_param_pair(param,
> paramMap[param])
> 83 self._java_obj.set(pair)
> 84
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/ml/wrapper.py in
> _make_java_param_pair(self, param, value)
> 71 java_param = self._java_obj.getParam(param.name)
> 72 java_value = _py2java(sc, value)
> ---> 73 return java_param.w(java_value)
> 74
> 75 def _transfer_params_to_java(self):
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/lib/py4j-0.9-src.zip/py4j/java_gateway.py
> in __call__(self, *args)
> 811 answer = self.gateway_client.send_command(command)
> 812 return_value = get_return_value(
> --> 813 answer, self.gateway_client, self.target_id, self.name)
> 814
> 815 for temp_arg in temp_args:
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/pyspark/sql/utils.py in
> deco(*a, **kw)
> 43 def deco(*a, **kw):
> 44 try:
> ---> 45 return f(*a, **kw)
> 46 except py4j.protocol.Py4JJavaError as e:
> 47 s = e.java_exception.toString()
> /Users/dreyco676/spark-1.6.0-bin-hadoop2.6/python/lib/py4j-0.9-src.zip/py4j/protocol.py
> in get_return_value(answer, gateway_client, target_id, name)
> 306 raise Py4JJavaError(
> 307 "An error occurred while calling {0}{1}{2}.\n".
> --> 308 format(target_id, ".", name), value)
> 309 else:
> 310 raise Py4JError(
> Py4JJavaError: An error occurred while calling o113.w.
> : java.lang.ClassCastException: java.lang.Integer cannot be cast to
> java.lang.Double
> at scala.runtime.BoxesRunTime.unboxToDouble(BoxesRunTime.java:119)
> at org.apache.spark.ml.param.DoubleParam.w(params.scala:223)
> at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
> at
> sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
> at
> sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
> at java.lang.reflect.Method.invoke(Method.java:497)
> at py4j.reflection.MethodInvoker.invoke(MethodInvoker.java:231)
> at py4j.reflection.ReflectionEngine.invoke(ReflectionEngine.java:381)
> at py4j.Gateway.invoke(Gateway.java:259)
> at py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:133)
> at py4j.commands.CallCommand.execute(CallCommand.java:79)
> at py4j.GatewayConnection.run(GatewayConnection.java:209)
> at java.lang.Thread.run(Thread.java:745)
> {code}
> Workaround is to run Transformers outside of pipeline. This ruins the purpose
> of Pipelines.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
