flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aljoscha Krettek <aljos...@apache.org>
Subject Re: Iterative Algorithm
Date Mon, 29 Sep 2014 10:00:01 GMT
Yes, the "first level" iteration code is actually not executed at
runtime. It is only executed once to put together the DataSet
operations. This graph of operations is then repeatedly executed at
runtime.

Cheers,
Aljoscha

On Mon, Sep 29, 2014 at 10:30 AM, Maximilian Alber
<alber.maximilian@gmail.com> wrote:
> Ah ok, that's the trick. So I can just use inside dataset function applied
> via rich functions during the iteration but not in the "first level"
> iteration code?
> But it shouldn't be a problem for me.
>
> Thanks!
> Cheers,
> Max
>
> On Mon, Sep 29, 2014 at 9:51 AM, Aljoscha Krettek <aljoscha@apache.org>
> wrote:
>>
>> Hi,
>> yes you can, I modified the ConnectedComponents Example to print out the
>> iteration number inside one of the Join functions:
>>
>> // open a delta iteration
>>     val verticesWithComponents = vertices.iterateDelta(vertices,
>> maxIterations, Array(0)) {
>>       (s, ws) =>
>>
>>         // apply the step logic: join with the edges
>>         val allNeighbors = ws.join(edges).where(0).equalTo(0) (
>>           new RichJoinFunction[(Long, Long), (Long, Long), (Long, Long)] {
>>             override def join(vertex: (Long, Long), edge: (Long, Long)):
>> (Long, Long) = {
>>               val context = getIterationRuntimeContext
>>               println("Iteration #" + context.getSuperstepNumber)
>>               (edge._2, vertex._2)
>>
>>             }
>>           })
>>
>>         // select the minimum neighbor
>>         val minNeighbors = allNeighbors.groupBy(0).min(1)
>>
>>         // update if the component of the candidate is smaller
>>         val updatedComponents = minNeighbors.join(s).where(0).equalTo(0) {
>>           (newVertex, oldVertex, out: Collector[(Long, Long)]) =>
>>             if (newVertex._2 < oldVertex._2) out.collect(newVertex)
>>         }
>>
>>         // delta and new workset are identical
>>         (updatedComponents, updatedComponents)
>>     }
>>
>> Unfortunately for this you have to use a RichFunction instead of a lambda.
>>
>> Cheers,
>> Aljoscha
>>
>> On Sat, Sep 27, 2014 at 11:09 AM, Maximilian Alber
>> <alber.maximilian@gmail.com> wrote:
>>>
>>> Ok. I'm back at this point:
>>>
>>> In the 0.7 version is there a way to get the superstep number inside a
>>> iterateWithDeta function?
>>>
>>> Cheers,
>>> Max
>>>
>>> On Mon, Aug 18, 2014 at 12:05 PM, Aljoscha Krettek <aljoscha@apache.org>
>>> wrote:
>>>>
>>>> Yes, but they were always available. Because user code in Java was
>>>> always in "Rich Functions". There is no rich function for iterations,
>>>> though, since iterations themselves don't have user code attached.
>>>>
>>>> Aljoscha
>>>>
>>>>
>>>> On Mon, Aug 18, 2014 at 10:59 AM, Fabian Hueske <fhueske@apache.org>
>>>> wrote:
>>>>>
>>>>> RichFunctions were added to the JavaAPI recently:
>>>>>
>>>>>
>>>>> https://github.com/apache/incubator-flink/tree/72d7b86274c33d1570ffb22b1fca2081c15d753c/flink-java/src/main/java/org/apache/flink/api/java/functions
>>>>>
>>>>> Cheers, Fabian
>>>>>
>>>>>
>>>>> 2014-08-18 8:16 GMT+02:00 Aljoscha Krettek <aljoscha@apache.org>:
>>>>>
>>>>>> Hi,
>>>>>> there is no RichFunction in the Java API either. You don't have to
>>>>>> create a new DataSet. Your iteration result will be a DataSet that
results
>>>>>> from some operations based on the previous SolutionSet and/or WorkingSet.
>>>>>> For example:
>>>>>>
>>>>>> def stepFunction(s: DataSet[SolutionType], ws: DataSet[WorksetType])
=
>>>>>> {
>>>>>>   val intermediate = ws.join(somethingFromOutside) where {...}
>>>>>> isEqualTo {...} map {...}
>>>>>>   val newSolution = s.join(intermediate) where ...
>>>>>>   val newWorkset = ...
>>>>>>   (newSolution, newWorkset)
>>>>>> }
>>>>>>
>>>>>> Aljoscha
>>>>>>
>>>>>>
>>>>>> On Sun, Aug 17, 2014 at 6:14 PM, Maximilian Alber
>>>>>> <alber.maximilian@gmail.com> wrote:
>>>>>>>
>>>>>>> Hi!
>>>>>>> Thank you!
>>>>>>>
>>>>>>> But how do I join my result to the solution set if I cannot create
a
>>>>>>> new DataSet inside the iteration?
>>>>>>> In Scala there is not yet a RichFunction for the Iterations,
am I
>>>>>>> right? So I should best use the Java class?
>>>>>>>
>>>>>>> Mit freundlichen Grüßen,
>>>>>>> Max!
>>>>>>>
>>>>>>>
>>>>>>> On Fri, Aug 15, 2014 at 3:50 PM, Aljoscha Krettek
>>>>>>> <aljoscha@apache.org> wrote:
>>>>>>>>
>>>>>>>> Hi,
>>>>>>>> right now, the only way of updating the solution set in a
delta
>>>>>>>> iteration is by joining with the solution set from the previous
iteration
>>>>>>>> and having the result of that join as the result of the step
function. I
>>>>>>>> working on simplifying iterations as well as bringing the
Scala API to
>>>>>>>> feature parity with the Java API. It should not be possible
right now to
>>>>>>>> create a new data source inside each iteration step.
>>>>>>>>
>>>>>>>> The way to get at the current iteration number is by having
a rich
>>>>>>>> function instead of a lambda function. So instead of:
>>>>>>>> val someSet = ...
>>>>>>>> val otherSet = someSet map { x => x + 1}
>>>>>>>>
>>>>>>>> you would have:
>>>>>>>> val someSet = ...
>>>>>>>> val otherSet = someSet map( new MapFunction[InType, OutType]()
{
>>>>>>>>   def apply(in: SomeType): SomeOtherType = {
>>>>>>>>     val iteration =
>>>>>>>> getIterationRuntimeContext().getSuperstepNumber()
>>>>>>>>     (iteration, x, y, ...)
>>>>>>>>   }
>>>>>>>> })
>>>>>>>>
>>>>>>>> I hope that helps.
>>>>>>>>
>>>>>>>> Aljoscha
>>>>>>>>
>>>>>>>>
>>>>>>>> On Tue, Aug 12, 2014 at 8:21 PM, Maximilian Alber
>>>>>>>> <alber.maximilian@gmail.com> wrote:
>>>>>>>>>
>>>>>>>>> Hi everybody,
>>>>>>>>>
>>>>>>>>> as already stated, I try currently to implement a Machine
Learning
>>>>>>>>> algorithm on Stratosphere for the ML group at TU Berlin.
I ran into some
>>>>>>>>> issues.
>>>>>>>>>
>>>>>>>>> The basic scheme of my algorithm is:
>>>>>>>>>
>>>>>>>>> X = input data
>>>>>>>>> Y = input data
>>>>>>>>> residuals = Y
>>>>>>>>>
>>>>>>>>> model = array[float, float, float] size n
>>>>>>>>>
>>>>>>>>> for i in 1:n
>>>>>>>>>   a = calc_a(X, residuals)
>>>>>>>>>   b = calc_b(X, a, residuals)
>>>>>>>>>   c = calc_c(X, a, b, c, residuals)
>>>>>>>>>
>>>>>>>>>   model(i) = (a, b, c)
>>>>>>>>>   residuals = update_residuals(residuals, a, b, c)
>>>>>>>>>
>>>>>>>>> output model
>>>>>>>>>
>>>>>>>>> My attempt now would be to use the delta iterations,
use the model
>>>>>>>>> as solution set, and the residuals as working sets:
>>>>>>>>>
>>>>>>>>> Code:
>>>>>>>>> -------------------------------------------------------------------------------------------------------------------------------------------------------
>>>>>>>>> val X = getInputSource
>>>>>>>>> val Y = DataSource(YFile, CsvInputFormat[Float])
>>>>>>>>>
>>>>>>>>> val model = CollectionDataSource[(Int, Float, Float,
>>>>>>>>> Float)](List())
>>>>>>>>> val residual = Y
>>>>>>>>>
>>>>>>>>> def step_function(model: DataSet[(Int, Float, Float,
Float)],
>>>>>>>>> residuals: DataSet[Float]) = {
>>>>>>>>> import util.Random
>>>>>>>>> (CollectionDataSource(Seq(new Tuple4(Random.nextInt,
1.0f, 1.0f,
>>>>>>>>> 2.0f))), residuals)
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> model.iterateWithDelta(
>>>>>>>>> residual,
>>>>>>>>> { x: (Int, Float, Float, Float) => x._1 },
>>>>>>>>> step_function,
>>>>>>>>> config.iterations
>>>>>>>>> )
>>>>>>>>>
>>>>>>>>> val output = model //map { x => println(x); x }
>>>>>>>>> val sink = output.write(outFile, CsvOutputFormat[(Int,
Float,
>>>>>>>>> Float, Float)], "Model output")
>>>>>>>>>
>>>>>>>>> Code End
>>>>>>>>> ----------------------------------------------------------------------------------------------------------------------------------------------------
>>>>>>>>>
>>>>>>>>> At the moment I try just to output a list of tuples.
>>>>>>>>>
>>>>>>>>> My problems are:
>>>>>>>>> - instead of the random integer I would like to insert
the index of
>>>>>>>>> the iterations.
>>>>>>>>> - I get this error:
>>>>>>>>> 08/12/2014 20:14:37: Job execution switched to status
SCHEDULED
>>>>>>>>> 08/12/2014 20:14:37: DataSource(<Unnamed Collection
Data Source>)
>>>>>>>>> (1/1) switched to SCHEDULED
>>>>>>>>> 08/12/2014 20:14:37: DataSink(Model output) (1/1) switched
to
>>>>>>>>> SCHEDULED
>>>>>>>>> 08/12/2014 20:14:37: DataSource(<Unnamed Collection
Data Source>)
>>>>>>>>> (1/1) switched to ASSIGNED
>>>>>>>>> 08/12/2014 20:14:37: DataSink(Model output) (1/1) switched
to
>>>>>>>>> ASSIGNED
>>>>>>>>> 08/12/2014 20:14:37: DataSource(<Unnamed Collection
Data Source>)
>>>>>>>>> (1/1) switched to READY
>>>>>>>>> 08/12/2014 20:14:37: DataSource(<Unnamed Collection
Data Source>)
>>>>>>>>> (1/1) switched to STARTING
>>>>>>>>> 08/12/2014 20:14:37: Job execution switched to status
RUNNING
>>>>>>>>> 08/12/2014 20:14:37: DataSource(<Unnamed Collection
Data Source>)
>>>>>>>>> (1/1) switched to RUNNING
>>>>>>>>> 08/12/2014 20:14:37: DataSink(Model output) (1/1) switched
to READY
>>>>>>>>> 08/12/2014 20:14:37: DataSink(Model output) (1/1) switched
to
>>>>>>>>> STARTING
>>>>>>>>> 08/12/2014 20:14:37: DataSink(Model output) (1/1) switched
to
>>>>>>>>> RUNNING
>>>>>>>>> 08/12/2014 20:14:38: DataSource(<Unnamed Collection
Data Source>)
>>>>>>>>> (1/1) switched to FINISHING
>>>>>>>>> 08/12/2014 20:14:38: DataSource(<Unnamed Collection
Data Source>)
>>>>>>>>> (1/1) switched to CANCELING
>>>>>>>>> 08/12/2014 20:14:38: DataSink(Model output) (1/1) switched
to
>>>>>>>>> FAILED
>>>>>>>>> java.lang.RuntimeException: Cannot serialize record with
out field
>>>>>>>>> at position: 0
>>>>>>>>> at
>>>>>>>>> eu.stratosphere.api.java.record.io.CsvOutputFormat.writeRecord(CsvOutputFormat.java:295)
>>>>>>>>> at
>>>>>>>>> eu.stratosphere.api.java.record.io.CsvOutputFormat.writeRecord(CsvOutputFormat.java:50)
>>>>>>>>> at
>>>>>>>>> eu.stratosphere.pact.runtime.task.DataSinkTask.invoke(DataSinkTask.java:178)
>>>>>>>>> at
>>>>>>>>> eu.stratosphere.nephele.execution.RuntimeEnvironment.run(RuntimeEnvironment.java:284)
>>>>>>>>> at java.lang.Thread.run(Thread.java:744)
>>>>>>>>>
>>>>>>>>> I doubt there is no record inside model. Because if I
enable the
>>>>>>>>> map function in the second last line I get an IndexOutOfBounds
exception at
>>>>>>>>> index 0.
>>>>>>>>>
>>>>>>>>> Many thanks in adavance
>>>>>>>>>
>>>>>>>>> Cheers,
>>>>>>>>> Max
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>

Mime
View raw message