Add and update manual pages for slicing, indexing, and reshaping.

This commit is contained in:
Gael Guennebaud 2018-11-09 11:35:27 +01:00
parent a368848473
commit d7c644213c
10 changed files with 365 additions and 66 deletions

View File

@ -64,13 +64,15 @@ namespace Eigen {
\ingroup DenseMatrixManipulation_chapter */ \ingroup DenseMatrixManipulation_chapter */
/** \addtogroup TutorialBlockOperations /** \addtogroup TutorialBlockOperations
\ingroup DenseMatrixManipulation_chapter */ \ingroup DenseMatrixManipulation_chapter */
/** \addtogroup TutorialSlicingIndexing
\ingroup DenseMatrixManipulation_chapter */
/** \addtogroup TutorialAdvancedInitialization /** \addtogroup TutorialAdvancedInitialization
\ingroup DenseMatrixManipulation_chapter */ \ingroup DenseMatrixManipulation_chapter */
/** \addtogroup TutorialReductionsVisitorsBroadcasting /** \addtogroup TutorialReductionsVisitorsBroadcasting
\ingroup DenseMatrixManipulation_chapter */ \ingroup DenseMatrixManipulation_chapter */
/** \addtogroup TutorialMapClass /** \addtogroup TutorialMapClass
\ingroup DenseMatrixManipulation_chapter */ \ingroup DenseMatrixManipulation_chapter */
/** \addtogroup TutorialReshapeSlicing /** \addtogroup TutorialReshape
\ingroup DenseMatrixManipulation_chapter */ \ingroup DenseMatrixManipulation_chapter */
/** \addtogroup TopicAliasing /** \addtogroup TopicAliasing
\ingroup DenseMatrixManipulation_chapter */ \ingroup DenseMatrixManipulation_chapter */

82
doc/TutorialReshape.dox Normal file
View File

@ -0,0 +1,82 @@
namespace Eigen {
/** \eigenManualPage TutorialReshape Reshape
Since the version 3.4, %Eigen exposes convenient methods to reshape a matrix to another matrix of different sizes or vector.
All cases are handled via the DenseBase::reshaped(NRowsType,NColsType) and DenseBase::reshaped() functions.
Those functions do not perform in-place reshaping, but instead return a <i> view </i> on the input expression.
\eigenAutoToc
\section TutorialReshapeMat2Mat Reshaped 2D views
The more general reshaping transformation is handled via: `reshaped(nrows,ncols)`.
Here is an example reshaping a 4x4 matrix to a 2x8 one:
<table class="example">
<tr><th>Example:</th><th>Output:</th></tr>
<tr><td>
\include MatrixBase_reshaped_int_int.cpp
</td>
<td>
\verbinclude MatrixBase_reshaped_int_int.out
</td></tr></table>
By default, the input coefficients are always interpreted in column-major order regardless of the storage order of the input expression.
For more control on ordering, compile-time sizes, and automatic size deduction, please see de documentation of DenseBase::reshaped(NRowsType,NColsType) that contains all the details with many examples.
\section TutorialReshapeMat2Vec 1D linear views
A very common usage of reshaping is to create a 1D linear view over a given 2D matrix or expression.
In this case, sizes can be deduced and thus omitted as in the following example:
<table class="example">
<tr><th>Example:</th></tr>
<tr><td>
\include MatrixBase_reshaped_to_vector.cpp
</td></tr>
<tr><th>Output:</th></tr>
<tr><td>
\verbinclude MatrixBase_reshaped_to_vector.out
</td></tr></table>
This shortcut always returns a column vector and by default input coefficients are always interpreted in column-major order.
Again, see the documentation of DenseBase::reshaped() for more control on the ordering.
\section TutorialReshapeInPlace
The above examples create reshaped views, but what about reshaping inplace a given matrix?
Of course this task in only conceivable for matrix and arrays having runtime dimensions.
In many cases, this can be accomplished via PlainObjectBase::resize(Index,Index):
<table class="example">
<tr><th>Example:</th></tr>
<tr><td>
\include Tutorial_reshaped_vs_resize_1.cpp
</td></tr>
<tr><th>Output:</th></tr>
<tr><td>
\verbinclude Tutorial_reshaped_vs_resize_1.out
</td></tr></table>
However beware that unlike \c reshaped, the result of \c resize depends on the input storage order.
It thus behaves similarly to `reshaped<AutoOrder>`:
<table class="example">
<tr><th>Example:</th></tr>
<tr><td>
\include Tutorial_reshaped_vs_resize_2.cpp
</td></tr>
<tr><th>Output:</th></tr>
<tr><td>
\verbinclude Tutorial_reshaped_vs_resize_2.out
</td></tr></table>
Finally, assigning a reshaped matrix to itself is currently not supported and will result to undefined-behavior because of \link TopicAliasing aliasing \endlink.
The following is forbidden: \code A = A.reshaped(2,8); \endcode
This is OK: \code A = A.reshaped(2,8).eval(); \endcode
*/
}

View File

@ -1,65 +0,0 @@
namespace Eigen {
/** \eigenManualPage TutorialReshapeSlicing Reshape and Slicing
%Eigen does not expose convenient methods to take slices or to reshape a matrix yet.
Nonetheless, such features can easily be emulated using the Map class.
\eigenAutoToc
\section TutorialReshape Reshape
A reshape operation consists in modifying the sizes of a matrix while keeping the same coefficients.
Instead of modifying the input matrix itself, which is not possible for compile-time sizes, the approach consist in creating a different \em view on the storage using class Map.
Here is a typical example creating a 1D linear view of a matrix:
<table class="example">
<tr><th>Example:</th><th>Output:</th></tr>
<tr><td>
\include Tutorial_ReshapeMat2Vec.cpp
</td>
<td>
\verbinclude Tutorial_ReshapeMat2Vec.out
</td></tr></table>
Remark how the storage order of the input matrix modifies the order of the coefficients in the linear view.
Here is another example reshaping a 2x6 matrix to a 6x2 one:
<table class="example">
<tr><th>Example:</th><th>Output:</th></tr>
<tr><td>
\include Tutorial_ReshapeMat2Mat.cpp
</td>
<td>
\verbinclude Tutorial_ReshapeMat2Mat.out
</td></tr></table>
\section TutorialSlicing Slicing
Slicing consists in taking a set of rows, columns, or elements, uniformly spaced within a matrix.
Again, the class Map allows to easily mimic this feature.
For instance, one can skip every P elements in a vector:
<table class="example">
<tr><th>Example:</th><th>Output:</th></tr>
<tr><td>
\include Tutorial_SlicingVec.cpp
</td>
<td>
\verbinclude Tutorial_SlicingVec.out
</td></tr></table>
One can also take one column over three using an adequate outer-stride or inner-stride depending on the actual storage order:
<table class="example">
<tr><th>Example:</th><th>Output:</th></tr>
<tr><td>
\include Tutorial_SlicingCol.cpp
</td>
<td>
\verbinclude Tutorial_SlicingCol.out
</td></tr></table>
*/
}

View File

@ -0,0 +1,244 @@
namespace Eigen {
/** \eigenManualPage TutorialSlicingIndexing Slicing and Indexing
This pape presents the numerous possibilities offered by `operator()` to index sub-set of rows and columns.
This API has been introduced in %Eigen 3.4.
It supports all the feature proposed by the \link TutorialBlockOperations block API \endlink, and much more.
In particular, it supports \b slicing that consists in taking a set of rows, columns, or elements, uniformly spaced within a matrix or indexed from an array of indices.
\eigenAutoToc
\section TutorialSlicingOverview Overview
All the aforementioned operations are handled through the generic DenseBase::operator()(const RowIndices&, const ColIndices&) method.
Each argument can be:
- An integer indexing a single row or column, including symbolic indices.
- The symbol Eigen::all representing the whole set of respective rows or columns in increasing order.
- An ArithmeticSequence as constructed by the Eigen::seq, Eigen::seqN, or Eigen::lastN functions.
- Any 1D vector/array of integers including %Eigen's vector/array, expressions, std::vector, std::array, as well as plain C arrays: `int[N]`.
More generally, it can accepts any object exposing the following two member functions:
\code
<integral type> operator[](<integral type>) const;
<integral type> size() const;
\endcode
where `<integral type>` stands for any integer type compatible with Eigen::Index (i.e. `std::ptrdiff_t`).
\section TutorialSlicingBasic Basic slicing
Taking a set of rows, columns, or elements, uniformly spaced within a matrix or vector is achieved through the Eigen::seq or Eigen::seqN functions where "seq" stands for arithmetic sequence. Their signatures are summarized below:
<table class="manual">
<tr>
<th>function</th>
<th>description</th>
<th>example</th>
</tr>
<tr>
<td>\code seq(firstIdx,lastIdx) \endcode</td>
<td>represents the sequence of integers ranging from \c firstIdx to \c lastIdx</td>
<td>\code seq(2,5) <=> {2,3,4,5} \endcode</td>
</tr>
<tr>
<td>\code seq(firstIdx,lastIdx,incr) \endcode</td>
<td>same but using the increment \c incr to advance from one index to the next</td>
<td>\code seq(2,8,2) <=> {2,4,6,8} \endcode</td>
</tr>
<tr>
<td>\code seqN(firstIdx,size) \endcode</td>
<td>represents the sequence of \c size integers starting from \c firstIdx</td>
<td>\code seqN(2,5) <=> {2,3,4,5,6} \endcode</td>
</tr>
<tr>
<td>\code seqN(firstIdx,size,incr) \endcode</td>
<td>same but using the increment \c incr to advance from one index to the next</td>
<td>\code seqN(2,3,3) <=> {2,5,8} \endcode</td>
</tr>
</table>
The \c firstIdx and \c lastIdx parameters can also be defined with the help of the Eigen::last symbol representing the index of the last row, column or element of the underlying matrix/vector once the arithmetic sequence is passed to it through operator().
Here are some examples for a 2D array/matrix \c A and a 1D array/vector \c v.
<table class="manual">
<tr>
<th>Intent</th>
<th>Code</th>
<th>Block-API equivalence</th>
</tr>
<tr>
<td>Bottom-left corner starting at row \c i with \c n columns</td>
<td>\code A(seq(i,last), seqN(0,n)) \endcode</td>
<td>\code A.bottomLeftCorner(A.rows()-i,n) \endcode</td>
</tr>
<tr>
<td>%Block starting at \c i,j having \c m rows, and \c n columns</td>
<td>\code A(seqN(i,m), seqN(i,n) \endcode</td>
<td>\code A.block(i,j,m,n) \endcode</td>
</tr>
<tr>
<td>%Block starting at \c i0,j0 and ending at \c i1,j1</td>
<td>\code A(seq(i0,i1), seq(j0,j1) \endcode</td>
<td>\code A.block(i0,j0,i1-i0+1,j1-j0+1) \endcode</td>
</tr>
<tr>
<td>Even columns of A</td>
<td>\code A(all, seq(0,last,2)) \endcode</td>
<td></td>
</tr>
<tr>
<td>First \c n odd rows A</td>
<td>\code A(seqN(1,n,2), all) \endcode</td>
<td></td>
</tr>
<tr>
<td>The last past one column</td>
<td>\code A(all, last-1) \endcode</td>
<td>\code A.col(A.cols()-2) \endcode</td>
</tr>
<tr>
<td>The middle row</td>
<td>\code A(last/2,all) \endcode</td>
<td>\code A.row((A.rows()-1)/2) \endcode</td>
</tr>
<tr>
<td>Last elements of v starting at i</td>
<td>\code v(seq(i,last)) \endcode</td>
<td>\code v.tail(v.size()-i) \endcode</td>
</tr>
<tr>
<td>Last \c n elements of v</td>
<td>\code v(seq(last+1-n,last)) \endcode</td>
<td>\code v.tail(n) \endcode</td>
</tr>
</table>
As seen in the last exemple, referencing the <i> last n </i> elements (or rows/columns) is a bit cumbersome to write.
This becomes even more tricky and error prone with a non-default increment.
Here comes \link Eigen::lastN(SizeType) Eigen::lastN(size) \endlink, and \link Eigen::lastN(SizeType,IncrType) Eigen::lastN(size,incr) \endlink:
<table class="manual">
<tr>
<th>Intent</th>
<th>Code</th>
<th>Block-API equivalence</th>
</tr>
<tr>
<td>Last \c n elements of v</td>
<td>\code v(lastN(n)) \endcode</td>
<td>\code v.tail(n) \endcode</td>
</tr>
<tr>
<td>Bottom-right corner of A of size \c m times \c n</td>
<td>\code v(lastN(m), lastN(n)) \endcode</td>
<td>\code A.bottomRightCorner(m,n) \endcode</td>
</tr>
<tr>
<td>Bottom-right corner of A of size \c m times \c n</td>
<td>\code v(lastN(m), lastN(n)) \endcode</td>
<td>\code A.bottomRightCorner(m,n) \endcode</td>
</tr>
<tr>
<td>Last \c n columns taking 1 column over 3</td>
<td>\code A(all, lastN(n,3)) \endcode</td>
<td></td>
</tr>
</table>
\section TutorialSlicingFixed Compile time size and increment
In terms of performance, %Eigen and the compiler can take advantage of compile-time size and increment.
To this end, you can enforce compile-time parameters using Eigen::fix<val>.
Such compile-time value can be combined with the Eigen::last symbol:
\code v(seq(last-fix<7>, last-fix<2>))
\endcode
In this example %Eigen knowns at compile-time that the returned expression has 6 elements.
It is equivalent to:
\code v(seqN(last-7, fix<6>))
\endcode
We can revisit the <i>even columns of A</i> example as follows:
\code A(all, seq(0,last,fix<2>))
\endcode
\section TutorialSlicingReverse Reverse order
Row/column indices can also be enumerated in decreasing order using a negative increment.
For instance, one over two columns of A from the column 20 to 10:
\code A(all, seq(20, 10, fix<-2>))
\endcode
The last \c n rows starting from the last one:
\code A(seqN(last, n, fix<-1>), all)
\endcode
You can also use the ArithmeticSequence::reverse() method to reverse its order.
The previous example can thus also be written as:
\code A(lastN(n).reverse(), all)
\endcode
\section TutorialSlicingArray Array of indices
The generic `operator()` can also takes as input an arbitrary list of row or column indices stored as either an `ArrayXi`, a `std::vector<int>`, `std::array<int,N>`, etc.
<table class="example">
<tr><th>Example:</th><th>Output:</th></tr>
<tr><td>
\include Slicing_stdvector_cxx11.cpp
</td>
<td>
\verbinclude Slicing_stdvector_cxx11.out
</td></tr></table>
You can also directly pass a static array:
<table class="example">
<tr><th>Example:</th><th>Output:</th></tr>
<tr><td>
\include Slicing_rawarray_cxx11.cpp
</td>
<td>
\verbinclude Slicing_rawarray_cxx11.out
</td></tr></table>
or expressions:
<table class="example">
<tr><th>Example:</th><th>Output:</th></tr>
<tr><td>
\include Slicing_arrayexpr.cpp
</td>
<td>
\verbinclude Slicing_arrayexpr.out
</td></tr></table>
When passing an object with a compile-time size such as `Array4i`, `std::array<int,N>`, or a static array, then the returned expression also exhibit compile-time dimensions.
\section TutorialSlicingCustomArray Custom index list
More generally, `operator()` can accept as inputs any object \c ind of type \c T compatible with:
\code
Index s = ind.size(); or Index s = size(ind);
Index i;
i = ind[i];
\endcode
This means you can easily build your own fancy sequence generator and pass it to `operator()`.
Here is an exemple enlarging a given matrix while padding the first rows and columns through repetition:
<table class="example">
<tr><th>Example:</th><th>Output:</th></tr>
<tr><td>
\include Slicing_custom_padding_cxx11.cpp
</td>
<td>
\verbinclude Slicing_custom_padding_cxx11.out
</td></tr></table>
<br>
*/
/*
TODO add:
so_repeat_inner.cpp
so_repeleme.cpp
*/
}

View File

@ -0,0 +1,4 @@
ArrayXi ind(5); ind<<4,2,5,5,3;
MatrixXi A = MatrixXi::Random(4,6);
cout << "Initial matrix A:\n" << A << "\n\n";
cout << "A(all,ind-1):\n" << A(all,ind-1) << "\n\n";

View File

@ -0,0 +1,12 @@
struct pad {
Index size() const { return out_size; }
Index operator[] (Index i) const { return std::max<Index>(0,i-(out_size-in_size)); }
Index in_size, out_size;
};
Matrix3i A;
A.reshaped() = VectorXi::LinSpaced(9,1,9);
cout << "Initial matrix A:\n" << A << "\n\n";
MatrixXi B(5,5);
B = A(pad{3,5}, pad{3,5});
cout << "A(pad{3,N}, pad{3,N}):\n" << B << "\n\n";

View File

@ -0,0 +1,5 @@
#if EIGEN_HAS_STATIC_ARRAY_TEMPLATE
MatrixXi A = MatrixXi::Random(4,6);
cout << "Initial matrix A:\n" << A << "\n\n";
cout << "A(all,{4,2,5,5,3}):\n" << A(all,{4,2,5,5,3}) << "\n\n";
#endif

View File

@ -0,0 +1,4 @@
std::vector<int> ind{4,2,5,5,3};
MatrixXi A = MatrixXi::Random(4,6);
cout << "Initial matrix A:\n" << A << "\n\n";
cout << "A(all,ind):\n" << A(all,ind) << "\n\n";

View File

@ -0,0 +1,5 @@
MatrixXi m = Matrix4i::Random();
cout << "Here is the matrix m:" << endl << m << endl;
cout << "Here is m.reshaped(2, 8):" << endl << m.reshaped(2, 8) << endl;
m.resize(2,8);
cout << "Here is the matrix m after m.resize(2,8):" << endl << m << endl;

View File

@ -0,0 +1,6 @@
Matrix<int,Dynamic,Dynamic,RowMajor> m = Matrix4i::Random();
cout << "Here is the matrix m:" << endl << m << endl;
cout << "Here is m.reshaped(2, 8):" << endl << m.reshaped(2, 8) << endl;
cout << "Here is m.reshaped<AutoOrder>(2, 8):" << endl << m.reshaped<AutoOrder>(2, 8) << endl;
m.resize(2,8);
cout << "Here is the matrix m after m.resize(2,8):" << endl << m << endl;