Github user MLnick commented on a diff in the pull request: https://github.com/apache/spark/pull/15593#discussion_r87550060 --- Diff: mllib/src/main/scala/org/apache/spark/ml/classification/LogisticRegression.scala --- @@ -1486,57 +1504,65 @@ private class LogisticAggregator( var marginOfLabel = 0.0 var maxMargin = Double.NegativeInfinity - val margins = Array.tabulate(numClasses) { i => - var margin = 0.0 - features.foreachActive { (index, value) => - if (localFeaturesStd(index) != 0.0 && value != 0.0) { - margin += localCoefficients(i * numFeaturesPlusIntercept + index) * - value / localFeaturesStd(index) - } + val margins = new Array[Double](numClasses) + features.foreachActive { (index, value) => + val stdValue = value / localFeaturesStd(index) + var j = 0 + while (j < numClasses) { + margins(j) += localCoefficients(index * numClasses + j) * stdValue + j += 1 } - + } + var i = 0 + while (i < numClasses) { if (fitIntercept) { - margin += localCoefficients(i * numFeaturesPlusIntercept + numFeatures) + margins(i) += localCoefficients(numClasses * numFeatures + i) } - if (i == label.toInt) marginOfLabel = margin - if (margin > maxMargin) { - maxMargin = margin + if (i == label.toInt) marginOfLabel = margins(i) + if (margins(i) > maxMargin) { + maxMargin = margins(i) } - margin + i += 1 } /** * When maxMargin > 0, the original formula could cause overflow. * We address this by subtracting maxMargin from all the margins, so it's guaranteed * that all of the new margins will be smaller than zero to prevent arithmetic overflow. */ + val multipliers = new Array[Double](numClasses) val sum = { var temp = 0.0 - if (maxMargin > 0) { - for (i <- 0 until numClasses) { - margins(i) -= maxMargin - temp += math.exp(margins(i)) - } - } else { - for (i <- 0 until numClasses) { - temp += math.exp(margins(i)) - } + var i = 0 + while (i < numClasses) { + if (maxMargin > 0) margins(i) -= maxMargin + val exp = math.exp(margins(i)) + temp += exp + multipliers(i) = exp + i += 1 } temp } - for (i <- 0 until numClasses) { - val multiplier = math.exp(margins(i)) / sum - { - if (label == i) 1.0 else 0.0 - } - features.foreachActive { (index, value) => - if (localFeaturesStd(index) != 0.0 && value != 0.0) { - localGradientArray(i * numFeaturesPlusIntercept + index) += - weight * multiplier * value / localFeaturesStd(index) + margins.indices.foreach { i => + multipliers(i) = multipliers(i) / sum - (if (label == i) 1.0 else 0.0) + } + features.foreachActive { (index, value) => + if (localFeaturesStd(index) != 0.0 && value != 0.0) { + val stdValue = value / localFeaturesStd(index) + var j = 0 + while (j < numClasses) { + localGradientArray(index * numClasses + j) += + weight * multipliers(j) * stdValue + j += 1 } } - if (fitIntercept) { - localGradientArray(i * numFeaturesPlusIntercept + numFeatures) += weight * multiplier + } + if (fitIntercept) { + var i = 0 + while (i < numClasses) { + localGradientArray(numFeatures * numClasses + i) += weight * multipliers(i) + i += 1 } } --- End diff -- My (perhaps incorrect) understanding of what you are saying: "The aggregator should internally convert to col-major during training & in-place update and from `gradient` will return a Matrix that can be used during L2 update (cycle through using `foreachActive`) in L1678-L1716." But to me that would require swapping `coeffs.foreachActive` for `gradient.foreachActive` in L1684. Since `coeff` is itself a vector (which we can't really change due to `BreezeDiffFunction`, and we laid it out col-major for training) we would need to index back into that using the same col-major indexing anyway, no? So I don't see that we gain anything for L2 reg. Also, if we change back and forth every iteration that seems like it would be inefficient. For the L1 reg function, that is purely index-based so `val isIntercept = $(fitIntercept) && ((index + 1) % numFeaturesPlusIntercept == 0)` doesn't seem any more or less complex than `val isIntercept = $(fitIntercept) && index >= numFeatures * numCoefficientSets`. So I don't think we gain anything during L1 reg function either. Finally, during initial coefficients setup, we already loop through them if they are provided. If not, the intercept adjustments are the only things that occur (the rest are zeros) so that is just indexing into the intercepts, which whether row- or col-major doesn't seem to me to be a major issue. I tend to agree that the col-major indexing is no "more complex" than row-major. It's also commented fairly clearly why and where we lay it out col-major (with example of the form), and where we convert back to row-major. We can't also need to convert back to row-major outside of the aggregator because the coefficients are returned from `optimizer.states.x` which is Breeze-specific, not `LogisticAggregator` or even `LogisticCostFun`.
--- 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