[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[asfgit]

Github user asfgit closed the pull request at:

https://github.com/apache/spark/pull/16037


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[AnthonyTruchet]

Github user AnthonyTruchet commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r92150485
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,24 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
--- End diff --

I did thanks and I merged it. I was just comming back to this contrib when 
I saw it.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[sethah]

Github user sethah commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r91975767
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,24 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
--- End diff --

I attempted to submit a PR to your branch, but combination of sketchy wifi 
and git - not sure it worked. 


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[AnthonyTruchet]

Github user AnthonyTruchet commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r91949738
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,24 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
--- End diff --

I'll try to work on this in the next two days. I will not be able to 
relaunch the actual benchmark we did weeks ago, our internal codebase has 
changed too much.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[srowen]

Github user srowen commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r91948007
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,24 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
--- End diff --

@AnthonyTruchet does that let you add a quick test case? then I think this 
can be merged.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[AnthonyTruchet]

Github user AnthonyTruchet commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r91493798
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,24 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
--- End diff --

I'm sorry, I have no clue how to generate a non-empty RDD with an empty 
partition. Can you please hint me at some entry points so that I can contribute 
the UTest you request ?


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[sethah]

Github user sethah commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r91206257
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,24 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
--- End diff --

I am still pretty strongly in favor of adding a test case explicitly for 
this. Just make an RDD with at least one empty partition, and be sure that 
LBFGS will run on it.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[sethah]

Github user sethah commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r91061168
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,27 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  // Adds two (gradient, loss) tuples
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
+  val denseGrad2 = grad2.toDense
+  axpy(1.0, denseGrad2, denseGrad1)
+  (denseGrad1, loss1 + loss2)
+   }
+
+  val zeroSparseVector = Vectors.sparse(n, Seq())
+  val (gradientSum, lossSum) = data.treeAggregate(zeroSparseVector, 
0.0)(seqOp, combOp)
--- End diff --

Yes :)


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[AnthonyTruchet]

Github user AnthonyTruchet commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r91036283
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,27 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  // Adds two (gradient, loss) tuples
--- End diff --

OK


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[srowen]

Github user srowen commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r91005791
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,27 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  // Adds two (gradient, loss) tuples
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
--- End diff --

Meaning, when would the args ever not be dense? I agree, shouldn't be 
sparse at this stage, but doing this defensively seems fine since it's a no-op 
for dense.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[sethah]

Github user sethah commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r90773986
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,27 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  // Adds two (gradient, loss) tuples
--- End diff --

I don't think this comment is necessary. The semantics of a combine op 
should be well understood.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[sethah]

Github user sethah commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r90773915
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,27 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  // Adds two (gradient, loss) tuples
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
--- End diff --

This is necessary because we may have empty partitions, right? Might be 
nice to add a test explicitly for this (it seems it failed in pyspark without 
it, but that was just a coincidence?) so someone doesn't remove these lines in 
the future.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[sethah]

Github user sethah commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r90773871
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,27 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  // Adds two (gradient, loss) tuples
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
+  val denseGrad2 = grad2.toDense
+  axpy(1.0, denseGrad2, denseGrad1)
+  (denseGrad1, loss1 + loss2)
+   }
+
+  val zeroSparseVector = Vectors.sparse(n, Seq())
+  val (gradientSum, lossSum) = data.treeAggregate(zeroSparseVector, 
0.0)(seqOp, combOp)
--- End diff --

I'm slightly in favor of keeping the parentheses for the zero values. If 
you don't know the signature of `treeAggregate` and that the scala compiler 
evidently packs these into a tuple, you may be confused about what the second 
argument is. Not a strong preference.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[AnthonyTruchet]

Github user AnthonyTruchet commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r90422375
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,27 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  // Adds two (gradient, loss) tuples
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
+  val denseGrad2 = grad2.toDense
+  axpy(1.0, denseGrad2, denseGrad1)
+  (grad1, loss1 + loss2)
--- End diff --

done, thanks for spotting.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[MLnick]

Github user MLnick commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r90421008
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +241,27 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
-axpy(1.0, grad2, grad1)
-(grad1, loss1 + loss2)
-  })
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(denseGrad, loss + l)
+}
+
+  // Adds two (gradient, loss) tuples
+  val combOp = (c1: (Vector, Double), c2: (Vector, Double)) =>
+(c1, c2) match { case ((grad1, loss1), (grad2, loss2)) =>
+  val denseGrad1 = grad1.toDense
+  val denseGrad2 = grad2.toDense
+  axpy(1.0, denseGrad2, denseGrad1)
+  (grad1, loss1 + loss2)
--- End diff --

I think `grad1` here will need to be `denseGrad1`?


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[MLnick]

Github user MLnick commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r90391752
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +239,25 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(Vectors.dense(denseGrad.values), loss + l)
--- End diff --

I can compile with my example code...


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[MLnick]

Github user MLnick commented on a diff in the pull request:

https://github.com/apache/spark/pull/16037#discussion_r90388974
  
--- Diff: 
mllib/src/main/scala/org/apache/spark/mllib/optimization/LBFGS.scala ---
@@ -241,16 +239,25 @@ object LBFGS extends Logging {
   val bcW = data.context.broadcast(w)
   val localGradient = gradient
 
-  val (gradientSum, lossSum) = data.treeAggregate((Vectors.zeros(n), 
0.0))(
-  seqOp = (c, v) => (c, v) match { case ((grad, loss), (label, 
features)) =>
-val l = localGradient.compute(
-  features, label, bcW.value, grad)
-(grad, loss + l)
-  },
-  combOp = (c1, c2) => (c1, c2) match { case ((grad1, loss1), 
(grad2, loss2)) =>
+  // Given (current accumulated gradient, current loss) and (label, 
features)
+  // tuples, updates the current gradient and current loss
+  val seqOp = (c: (Vector, Double), v: (Double, Vector)) =>
+(c, v) match {
+  case ((grad, loss), (label, features)) =>
+val denseGrad = grad.toDense
+val l = localGradient.compute(features, label, bcW.value, 
denseGrad)
+(Vectors.dense(denseGrad.values), loss + l)
--- End diff --

Is this really necessary? 


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org

[GitHub] spark pull request #16037: [SPARK-18471][MLLIB] In LBFGS, avoid sending huge...

[AnthonyTruchet]

GitHub user AnthonyTruchet opened a pull request:

https://github.com/apache/spark/pull/16037

[SPARK-18471][MLLIB] In LBFGS, avoid sending huge vectors of 0

## What changes were proposed in this pull request?

CostFun used to send a dense vector of zeroes as a closure in a
treeAggregate call. To avoid that, we replace treeAggregate by
mapPartition + treeReduce, creating a zero vector inside the mapPartition
block in-place.

## How was this patch tested?

Unit test for module mllib run locally for correctness.

As for performance we run an heavy optimization on our production data (50 
iterations on 128 MB weight vectors) and have seen significant decrease in 
terms both of runtime and container being killed by lack of off-heap memory.


You can merge this pull request into a Git repository by running:

$ git pull https://github.com/criteo-forks/spark ENG-17719-lbfgs-only

Alternatively you can review and apply these changes as the patch at:

https://github.com/apache/spark/pull/16037.patch

To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:

This closes #16037


commit 0e924ec09554acf3191dea51c5a13b024a0c643f
Author: Eugene Kharitonov 
Date:   2016-10-14T11:25:34Z

[SPARK-18471][MLLIB] In LBFGS, avoid sending huge vectors of 0

CostFun used to send a dense vector of zeroes as a closure in a
treeAggregate call. To avoid that, we replace treeAggregate by
mapPartition + treeReduce, creating a zero vector inside the mapPartition
block in-place.

commit af4b1c96ec5a5759e8b66117794d2a011f6313c3
Author: Anthony Truchet 
Date:   2016-11-18T12:36:00Z

Style fix according to reviewers' feedback

commit 0a0571dc9557e94a909ab8d8ef62c7dca474a1c0
Author: Anthony Truchet 
Date:   2016-11-28T15:39:16Z

Code style according to reviewers feed-back.




---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---

-
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org