[MAHOUT-790] Redundancy in Matrix API, view or get?

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.5
• Fix Version/s: 0.6
• Assignee: Ted Dunning
• Reporter: Ted Dunning
• Created: 20/Aug/11 21:28
• Updated: 02/May/13 02:29
• Resolved: 13/Sep/11 09:12

Description

We have a bunch of redundant methods in our matrix interface. These include things that return views of parts of the matrix:

  Matrix viewPart(int[] offset, int[] size);
  Matrix viewPart(int rowOffset, int rowsRequested, int columnOffset, int columnsRequested);
  Vector viewRow(int row);
  Vector viewColumn(int column);

and things that do the same but call refer to getting stuff

  Vector getColumn(int column);
  Vector getRow(int row);
  double getQuick(int row, int column);
  int[] getNumNondefaultElements();
  Map<String, Integer> getColumnLabelBindings();
  Map<String, Integer> getRowLabelBindings();
  double get(String rowLabel, String columnLabel);

To my mind, get implies a get-by-value whereas view implies get-by-reference. As such, I would suggest that getColumn and getRow should disappear. On the other hand, getQuick and get are both correctly named.

This raises the question of what getNumNondefaultElements really does. I certainly can't tell just from the signature. Is it too confusing to keep?

Additionally, what do people think that getColumnLabelBindings and getRowLabelBindings return? A mutable map? Or an immutable one?

Under the covers, viewRow and viewColumn (and the upcoming viewDiagonal) have default implementations that use MatrixVectorView, but AbstractMatrix doesn't have an implementation for getRow and getColumn.

In sum, I suggest that:

• getRow and getColumn go away

• the fancy fast implementations fo getRow and getColumn that exist be migrated to be over-rides of viewRow and viewColumn

• there be a constructor for AbstractMatrix that sets the internal size things correctly.

• that the internal cardinality array in AbstractMatrix goes away to be replaced by two integers.

• viewDiagonal() and viewDiagonal(length) and viewDiagonal(row, column) and viewDiagonal(int row, column, length) be added.

I will produce a patch shortly.

Issue Links

• blocks: MAHOUT-792 Add new stochastic decomposition code
• is blocked by: MAHOUT-793 Move MurmurHash to math
• is depended upon by: MAHOUT-796 Modified power iterations in existing SSVD code
• is depended upon by: MAHOUT-797 MapReduce SSVD: provide alternative B-pipeline per B=R' ^{-1} Y'A