[SPARK-5652][Mllib] Use broadcasted weights in LogisticRegressionModel

[viirya]

LogisticRegressionModel's predictPoint should directly use broadcasted weights. This pr also fixes the compilation errors of two unit test suite: JavaLogisticRegressionSuite and JavaLinearRegressionSuite.

[srowen]

CC @jkbradley for the test compile issue, which looks straightforward, and @dbtsai for the question of weights vs weightMatrix

[SparkQA]

Test build #26914 has finished for PR 4429 at commit 5a797e5.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[mengxr]

@viirya Any performance gain by changing weightMatrix to weights? Could you say more about the compilation error? The changes looks good to me, but I didn't see any error before making the change.

[jkbradley]

The changes LGTM too. That compilation error doesn't happen for me either, but the changes are fine.

The switch to weightMatrix looks correct (so that it uses the broadcast variable).

[mengxr]

@viirya For the weightMatrix -> weights change. This is inside predictPoint and by definition we should use the weightMatrix from input arguments. Are we serializing weights by putting weights in the closure?

[mengxr]

Merged into master and branch-1.3. Thanks!

[viirya]

weightMatrix is just the broadcasted variable of the local copy of original weights. It is used because we don't want to serialize whole model but just the weights.

For the compilation error, when I run the test, the compiler complains something like Object can't compare to Int.