Github user dusenberrymw commented on a diff in the pull request:
https://github.com/apache/spark/pull/7554#discussion_r35251990
--- Diff: python/pyspark/mllib/linalg.py ---
@@ -1152,9 +1156,385 @@ def sparse(numRows, numCols, colPtrs, rowIndices,
values):
return SparseMatrix(numRows, numCols, colPtrs, rowIndices, values)
+class DistributedMatrix(object):
+ """Represents a distributively stored matrix backed by one or more
RDDs."""
+ def numRows(self):
+ """Get or compute the number of rows."""
+ raise NotImplementedError
+
+ def numCols(self):
+ """Get or compute the number of cols."""
+ raise NotImplementedError
+
+
+class DistributedMatrices(object):
+ """Factory methods for distributed matrices."""
+ @staticmethod
+ def rowMatrix(rows, numRows=0, numCols=0):
+ """
+ Create a RowMatrix.
+
+ :param rows: An RDD of Vectors.
+ """
+ javaRowMatrix = callMLlibFunc("createRowMatrix", rows,
long(numRows), int(numCols))
+ jrm = JavaModelWrapper(javaRowMatrix)
+ return RowMatrix(jrm)
+
+ @staticmethod
+ def indexedRowMatrix(rows, numRows=0, numCols=0):
+ """
+ Create an IndexedRowMatrix.
+
+ :param rows: An RDD of IndexedRows or (long, Vector) tuples.
+ """
+ # We use DataFrames for serialization of IndexedRows from Python,
so convert the RDD to a
+ # DataFrame. This will convert each IndexedRow to a Row containing
the 'index' and 'vector'
+ # values, which can both be easily serialized. We will convert
back to IndexedRows on the
+ # Scala side.
+ javaIndexedRowMatrix = callMLlibFunc("createIndexedRowMatrix",
rows.toDF(),
--- End diff --
The idea is that the `IndexedRow` Python type isn't directly serializable
to a format that can be read by the Java side, and vice-versa. Any of these
custom data structures require explicit serialization, which can be found in
the `SerDe` Scala class at the bottom of `PythonMLlibAPI.scala`, as well as a
bit of representation logic on the Python side. Prior to `DataFrame`s, this
was done for any structure that needed to be passed between the languages, so
we already have explicit conversions for `DenseVector`, `DenseMatrix`,
`LabeledPoint`, `Rating`, etc. The `DataFrame` structure lets us pull the
contents out of these custom data structures and store them in a a generic
`Row`. The contents are often simpler types, and `DataFrame`s can
serialization "primitives" such as `Int`, `Long`, `Boolean`, etc., as well as a
few custom types such as `Vector` and `Matrix` types. So the preferred way of
serialization moving forward is to use `DataFrame`s for serialization when
possible.
For the `IndexedRow` type, it just contains a `long` and a `Vector`, both of
which can be easily serialized with `DataFrame`s, even though the `IndexedRow`
itself isn't cross-language serializable. Converting the `RDD` of
`IndexedRow`s in Python to a `DataFrame` creates a `DataFrame` with `Row`s
containing just the `long` and `Vector` from the `IndexedRow`.
---
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 [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
