Added support for extraction of patches from images

6 files changed, 587 insertions, 0 deletions

diff --git a/unsupported/Eigen/CXX11/Tensor b/unsupported/Eigen/CXX11/Tensor
index 44d5a4d82..aa26e5283 100644
--- a/unsupported/Eigen/CXX11/Tensor
+++ b/unsupported/Eigen/CXX11/Tensor
@@ -59,6 +59,7 @@
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h"
+#include "unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h"
 #include "unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h"
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index 6018ecc66..f451a3c99 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -255,6 +255,19 @@ class TensorBase<Derived, ReadOnlyAccessors>
       return TensorPatchOp<const PatchDims, const Derived>(derived(), patch_dims);
     }
 
+    template <Index Rows, Index Cols> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorImagePatchOp<Rows, Cols, const Derived>
+    extract_image_patches() const {
+      return TensorImagePatchOp<Rows, Cols, const Derived>(derived(), Rows, Cols, 1, 1);
+    }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const TensorImagePatchOp<Dynamic, Dynamic, const Derived>
+    extract_image_patches(const Index patch_rows, const Index patch_cols,
+                          const Index row_stride = 1, const Index col_stride = 1) const {
+      return TensorImagePatchOp<Dynamic, Dynamic, const Derived>(derived(), patch_rows, patch_cols, row_stride, col_stride);
+    }
+
     // Morphing operators.
     template <typename NewDimensions> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
     const TensorReshapingOp<const NewDimensions, const Derived>
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index a72e11215..85599ccfd 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -27,6 +27,7 @@ template<typename Axis, typename LeftXprType, typename RightXprType> class Tenso
 template<typename Dimensions, typename LeftXprType, typename RightXprType> class TensorContractionOp;
 template<typename Dimensions, typename InputXprType, typename KernelXprType> class TensorConvolutionOp;
 template<typename PatchDim, typename XprType> class TensorPatchOp;
+template<DenseIndex Rows, DenseIndex Cols, typename XprType> class TensorImagePatchOp;
 template<typename Broadcast, typename XprType> class TensorBroadcastingOp;
 template<std::size_t DimId, typename XprType> class TensorChippingOp;
 template<typename NewDimensions, typename XprType> class TensorReshapingOp;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
new file mode 100644
index 000000000..ce916fdfd
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -0,0 +1,291 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_IMAGE_PATCH_H
+#define EIGEN_CXX11_TENSOR_TENSOR_IMAGE_PATCH_H
+
+namespace Eigen {
+
+/** \class TensorImagePatch
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Patch extraction specialized for image processing.
+  * This assumes that the input has a least 3 dimensions ordered as follow:
+  *  1st dimension: channels (of size d)
+  *  2nd dimension: rows (of size r)
+  *  3rd dimension: columns (of size c)
+  *  There can be additional dimensions such as time (for video) or batch (for
+  * bulk processing after the first 3.
+  * Calling the image patch code with patch_rows and patch_cols is equivalent
+  * to calling the regular patch extraction code with parameters d, patch_rows,
+  * patch_cols, and 1 for all the additional dimensions.
+  */
+namespace internal {
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>
+struct traits<TensorImagePatchOp<Rows, Cols, XprType> > : public traits<XprType>
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef traits<XprType> XprTraits;
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename XprTraits::StorageKind StorageKind;
+  typedef typename XprTraits::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = XprTraits::NumDimensions + 1;
+};
+
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>
+struct eval<TensorImagePatchOp<Rows, Cols, XprType>, Eigen::Dense>
+{
+  typedef const TensorImagePatchOp<Rows, Cols, XprType>& type;
+};
+
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>
+struct nested<TensorImagePatchOp<Rows, Cols, XprType>, 1, typename eval<TensorImagePatchOp<Rows, Cols, XprType> >::type>
+{
+  typedef TensorImagePatchOp<Rows, Cols, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<DenseIndex Rows, DenseIndex Cols, typename XprType>
+class TensorImagePatchOp : public TensorBase<TensorImagePatchOp<Rows, Cols, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorImagePatchOp>::Scalar Scalar;
+  typedef typename Eigen::internal::traits<TensorImagePatchOp>::Packet Packet;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+  typedef typename Eigen::internal::nested<TensorImagePatchOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorImagePatchOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorImagePatchOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorImagePatchOp(const XprType& expr, DenseIndex patch_rows, DenseIndex patch_cols,
+                                                           DenseIndex row_strides, DenseIndex col_strides)
+      : m_xpr(expr), m_patch_rows(patch_rows), m_patch_cols(patch_cols),
+        m_row_strides(row_strides), m_col_strides(col_strides){}
+
+    EIGEN_DEVICE_FUNC
+    DenseIndex patch_rows() const { return m_patch_rows; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex patch_cols() const { return m_patch_cols; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex row_strides() const { return m_row_strides; }
+    EIGEN_DEVICE_FUNC
+    DenseIndex col_strides() const { return m_col_strides; }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+    const DenseIndex m_patch_rows;
+    const DenseIndex m_patch_cols;
+    const DenseIndex m_row_strides;
+    const DenseIndex m_col_strides;
+};
+
+
+// Eval as rvalue
+template<DenseIndex Rows, DenseIndex Cols, typename ArgType, typename Device>
+struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
+{
+  typedef TensorImagePatchOp<Rows, Cols, ArgType> XprType;
+  typedef typename XprType::Index Index;
+  static const int NumDims = internal::array_size<typename TensorEvaluator<ArgType, Device>::Dimensions>::value + 1;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device)
+  {
+    EIGEN_STATIC_ASSERT(NumDims >= 4, YOU_MADE_A_PROGRAMMING_MISTAKE);
+
+    const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
+    m_dimensions[0] = input_dims[0];
+    m_dimensions[1] = op.patch_rows();
+    m_dimensions[2] = op.patch_cols();
+    m_dimensions[3] = ceilf(static_cast<float>(input_dims[1]) / op.row_strides()) *
+                      ceilf(static_cast<float>(input_dims[2]) / op.col_strides());
+    for (int i = 4; i < NumDims; ++i) {
+      m_dimensions[i] = input_dims[i-1];
+    }
+
+    m_colStride = m_dimensions[1];
+    m_patchStride = m_colStride * m_dimensions[2] * m_dimensions[0];
+    m_otherStride = m_patchStride * m_dimensions[3];
+
+    m_inputRows = input_dims[1];
+    m_inputCols = input_dims[2];
+
+    m_rowInputStride = input_dims[0] * op.row_strides();
+    m_colInputStride = input_dims[0] * input_dims[1] * op.col_strides();
+    m_patchInputStride = input_dims[0] * input_dims[1] * input_dims[2];
+
+    m_rowPaddingTop = op.patch_rows() / 2;
+    m_colPaddingLeft = op.patch_cols() / 2;
+
+    m_fastOtherStride = internal::TensorIntDivisor<Index>(m_otherStride);
+    m_fastPatchStride = internal::TensorIntDivisor<Index>(m_patchStride);
+    m_fastColStride = internal::TensorIntDivisor<Index>(m_colStride);
+    m_fastInputRows = internal::TensorIntDivisor<Index>(m_inputRows);
+    m_fastDimZero = internal::TensorIntDivisor<Index>(m_dimensions[0]);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) {
+    m_impl.evalSubExprsIfNeeded(NULL);
+    return true;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
+    m_impl.cleanup();
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    // Find the location of the first element of the patch.
+    const Index patchIndex = index / m_fastPatchStride;
+
+    // Find the offset of the element wrt the location of the first element.
+    const Index patchOffset = (index - patchIndex * m_patchStride) / m_fastDimZero;
+
+    const Index otherIndex = (NumDims == 4) ? 0 : index / m_fastOtherStride;
+    const Index patch2DIndex = (NumDims == 4) ? patchIndex : (index - otherIndex * m_otherStride) / m_fastPatchStride;
+
+    const Index colIndex = patch2DIndex / m_fastInputRows;
+    const Index colOffset = patchOffset / m_fastColStride;
+
+    const Index inputCol = colIndex + colOffset - m_colPaddingLeft;
+    if (inputCol < 0 || inputCol >= m_inputCols) {
+      return Scalar(0);
+    }
+    const Index rowIndex = patch2DIndex - colIndex * m_inputRows;  // m_rowStride is always 1
+    const Index rowOffset = patchOffset - colOffset * m_colStride;
+
+    const Index inputRow = rowIndex + rowOffset - m_rowPaddingTop;
+    if (inputRow < 0 || inputRow >= m_inputRows) {
+      return Scalar(0);
+    }
+
+    const Index depth = index - (index / m_fastDimZero) * m_dimensions[0];
+
+    const Index inputIndex = depth + inputRow * m_rowInputStride + inputCol * m_colInputStride + otherIndex * m_patchInputStride;
+    return m_impl.coeff(inputIndex);
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    const Index packetSize = internal::unpacket_traits<PacketReturnType>::size;
+    EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
+    eigen_assert(index+packetSize-1 < dimensions().TotalSize());
+
+    const Index indices[2] = {index, index + packetSize - 1};
+    const Index patchIndex = indices[0] / m_fastPatchStride;
+    if (patchIndex != indices[1] / m_fastPatchStride) {
+      return packetWithPossibleZero(index);
+    }
+    const Index otherIndex = (NumDims == 4) ? 0 : indices[0] / m_fastOtherStride;
+    eigen_assert(otherIndex == indices[1] / m_fastOtherStride);
+
+    // Find the offset of the element wrt the location of the first element.
+    const Index patchOffsets[2] = {(indices[0] - patchIndex * m_patchStride) / m_fastDimZero,
+                                   (indices[1] - patchIndex * m_patchStride) / m_fastDimZero};
+
+    const Index patch2DIndex = (NumDims == 4) ? patchIndex : (indices[0] - otherIndex * m_otherStride) / m_fastPatchStride;
+    eigen_assert(patch2DIndex == (indices[1] - otherIndex * m_otherStride) / m_fastPatchStride);
+
+    const Index colIndex = patch2DIndex / m_fastInputRows;
+    const Index colOffsets[2] = {patchOffsets[0] / m_fastColStride, patchOffsets[1] / m_fastColStride};
+
+    const Index inputCols[2] = {colIndex + colOffsets[0] - m_colPaddingLeft, colIndex + colOffsets[1] - m_colPaddingLeft};
+    if (inputCols[1] < 0 || inputCols[0] >= m_inputCols) {
+      // all zeros
+      return internal::pset1<PacketReturnType>(Scalar(0));
+    }
+
+    if (inputCols[0] == inputCols[1]) {
+      const Index rowIndex = patch2DIndex - colIndex * m_inputRows;
+      const Index rowOffsets[2] = {patchOffsets[0] - colOffsets[0]*m_colStride, patchOffsets[1] - colOffsets[1]*m_colStride};
+      eigen_assert(rowOffsets[0] <= rowOffsets[1]);
+      const Index inputRows[2] = {rowIndex + rowOffsets[0] - m_rowPaddingTop, rowIndex + rowOffsets[1] - m_rowPaddingTop};
+
+      if (inputRows[1] < 0 || inputRows[0] >= m_inputRows) {
+        // all zeros
+        return internal::pset1<PacketReturnType>(Scalar(0));
+      }
+
+      if (inputRows[0] >= 0 && inputRows[1] < m_inputRows) {
+        // no padding
+        const Index depth = index - (index / m_fastDimZero) * m_dimensions[0];
+        const Index inputIndex = depth + inputRows[0] * m_rowInputStride + inputCols[0] * m_colInputStride + otherIndex * m_patchInputStride;
+        return m_impl.template packet<Unaligned>(inputIndex);
+      }
+    }
+
+    return packetWithPossibleZero(index);
+  }
+
+  Scalar* data() const { return NULL; }
+
+ protected:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packetWithPossibleZero(Index index) const
+  {
+    const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
+    EIGEN_ALIGN_DEFAULT typename internal::remove_const<CoeffReturnType>::type values[packetSize];
+    for (int i = 0; i < packetSize; ++i) {
+      values[i] = coeff(index+i);
+    }
+    PacketReturnType rslt = internal::pload<PacketReturnType>(values);
+    return rslt;
+  }
+
+  Dimensions m_dimensions;
+
+  Index m_otherStride;
+  Index m_patchStride;
+  Index m_colStride;
+  internal::TensorIntDivisor<Index> m_fastOtherStride;
+  internal::TensorIntDivisor<Index> m_fastPatchStride;
+  internal::TensorIntDivisor<Index> m_fastColStride;
+
+  Index m_rowInputStride;
+  Index m_colInputStride;
+  Index m_patchInputStride;
+
+  Index m_inputRows;
+  Index m_inputCols;
+
+  Index m_rowPaddingTop;
+  Index m_colPaddingLeft;
+
+  internal::TensorIntDivisor<Index> m_fastInputRows;
+  internal::TensorIntDivisor<Index> m_fastDimZero;
+
+  TensorEvaluator<ArgType, Device> m_impl;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_IMAGE_PATCH_H
diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index 181f06fc7..89c651804 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -122,6 +122,7 @@ if(EIGEN_TEST_CXX11)
   ei_add_test(cxx11_tensor_morphing "-std=c++0x")
   ei_add_test(cxx11_tensor_padding "-std=c++0x")
   ei_add_test(cxx11_tensor_patch "-std=c++0x")
+  ei_add_test(cxx11_tensor_image_patch "-std=c++0x")
   ei_add_test(cxx11_tensor_reduction "-std=c++0x")
   ei_add_test(cxx11_tensor_shuffling "-std=c++0x")
   ei_add_test(cxx11_tensor_striding "-std=c++0x")
diff --git a/unsupported/test/cxx11_tensor_image_patch.cpp b/unsupported/test/cxx11_tensor_image_patch.cpp
new file mode 100644
index 000000000..55d35eac0
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_image_patch.cpp
@@ -0,0 +1,280 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#include "main.h"
+
+#include <Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+
+static void test_simple_patch()
+{
+  Tensor<float, 4> tensor(2,3,5,7);
+  tensor.setRandom();
+
+  Tensor<float, 5> single_pixel_patch;
+  single_pixel_patch = tensor.extract_image_patches<1, 1>();
+
+  VERIFY_IS_EQUAL(single_pixel_patch.dimension(0), 2);
+  VERIFY_IS_EQUAL(single_pixel_patch.dimension(1), 1);
+  VERIFY_IS_EQUAL(single_pixel_patch.dimension(2), 1);
+  VERIFY_IS_EQUAL(single_pixel_patch.dimension(3), 3*5);
+  VERIFY_IS_EQUAL(single_pixel_patch.dimension(4), 7);
+
+  for (int i = 0; i < tensor.size(); ++i) {
+    VERIFY_IS_EQUAL(single_pixel_patch.data()[i], tensor.data()[i]);
+  }
+
+  Tensor<float, 5> entire_image_patch;
+  entire_image_patch = tensor.extract_image_patches<3, 5>();
+
+  VERIFY_IS_EQUAL(entire_image_patch.dimension(0), 2);
+  VERIFY_IS_EQUAL(entire_image_patch.dimension(1), 3);
+  VERIFY_IS_EQUAL(entire_image_patch.dimension(2), 5);
+  VERIFY_IS_EQUAL(entire_image_patch.dimension(3), 3*5);
+  VERIFY_IS_EQUAL(entire_image_patch.dimension(4), 7);
+
+  for (int i = 0; i < 3; ++i) {
+    for (int j = 0; j < 5; ++j) {
+      int patchId = i+3*j;
+      for (int r = 0; r < 3; ++r) {
+        for (int c = 0; c < 5; ++c) {
+          for (int d = 0; d < 2; ++d) {
+            for (int b = 0; b < 7; ++b) {
+              float expected = 0.0f;
+              if (r-1+i >= 0 && c-2+j >= 0 && r-1+i < 3 && c-2+j < 5) {
+                expected = tensor(d, r-1+i, c-2+j, b);
+              }
+              VERIFY_IS_EQUAL(entire_image_patch(d, r, c, patchId, b), expected);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  Tensor<float, 5> twod_patch;
+  twod_patch = tensor.extract_image_patches<2, 2>();
+
+  VERIFY_IS_EQUAL(twod_patch.dimension(0), 2);
+  VERIFY_IS_EQUAL(twod_patch.dimension(1), 2);
+  VERIFY_IS_EQUAL(twod_patch.dimension(2), 2);
+  VERIFY_IS_EQUAL(twod_patch.dimension(3), 3*5);
+  VERIFY_IS_EQUAL(twod_patch.dimension(4), 7);
+
+  for (int i = 0; i < 3; ++i) {
+    for (int j = 0; j < 5; ++j) {
+      int patchId = i+3*j;
+      for (int r = 0; r < 2; ++r) {
+        for (int c = 0; c < 2; ++c) {
+          for (int d = 0; d < 2; ++d) {
+            for (int b = 0; b < 7; ++b) {
+              float expected = 0.0f;
+              if (r-1+i >= 0 && c-1+j >= 0 && r-1+i < 3 && c-1+j < 5) {
+                expected = tensor(d, r-1+i, c-1+j, b);
+              }
+              VERIFY_IS_EQUAL(twod_patch(d, r, c, patchId, b), expected);
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+
+static void test_patch_no_extra_dim()
+{
+  Tensor<float, 3> tensor(2,3,5);
+  tensor.setRandom();
+
+  Tensor<float, 4> single_pixel_patch;
+  single_pixel_patch = tensor.extract_image_patches<1, 1>();
+
+  VERIFY_IS_EQUAL(single_pixel_patch.dimension(0), 2);
+  VERIFY_IS_EQUAL(single_pixel_patch.dimension(1), 1);
+  VERIFY_IS_EQUAL(single_pixel_patch.dimension(2), 1);
+  VERIFY_IS_EQUAL(single_pixel_patch.dimension(3), 3*5);
+
+  for (int i = 0; i < tensor.size(); ++i) {
+    VERIFY_IS_EQUAL(single_pixel_patch.data()[i], tensor.data()[i]);
+  }
+
+  Tensor<float, 4> entire_image_patch;
+  entire_image_patch = tensor.extract_image_patches<3, 5>();
+
+  VERIFY_IS_EQUAL(entire_image_patch.dimension(0), 2);
+  VERIFY_IS_EQUAL(entire_image_patch.dimension(1), 3);
+  VERIFY_IS_EQUAL(entire_image_patch.dimension(2), 5);
+  VERIFY_IS_EQUAL(entire_image_patch.dimension(3), 3*5);
+
+  for (int i = 0; i < 3; ++i) {
+    for (int j = 0; j < 5; ++j) {
+      int patchId = i+3*j;
+      for (int r = 0; r < 3; ++r) {
+        for (int c = 0; c < 5; ++c) {
+          for (int d = 0; d < 2; ++d) {
+            float expected = 0.0f;
+            if (r-1+i >= 0 && c-2+j >= 0 && r-1+i < 3 && c-2+j < 5) {
+              expected = tensor(d, r-1+i, c-2+j);
+            }
+            VERIFY_IS_EQUAL(entire_image_patch(d, r, c, patchId), expected);
+          }
+        }
+      }
+    }
+  }
+
+  Tensor<float, 4> twod_patch;
+  twod_patch = tensor.extract_image_patches<2, 2>();
+
+  VERIFY_IS_EQUAL(twod_patch.dimension(0), 2);
+  VERIFY_IS_EQUAL(twod_patch.dimension(1), 2);
+  VERIFY_IS_EQUAL(twod_patch.dimension(2), 2);
+  VERIFY_IS_EQUAL(twod_patch.dimension(3), 3*5);
+
+  for (int i = 0; i < 3; ++i) {
+    for (int j = 0; j < 5; ++j) {
+      int patchId = i+3*j;
+      for (int r = 0; r < 2; ++r) {
+        for (int c = 0; c < 2; ++c) {
+          for (int d = 0; d < 2; ++d) {
+            float expected = 0.0f;
+            if (r-1+i >= 0 && c-1+j >= 0 && r-1+i < 3 && c-1+j < 5) {
+              expected = tensor(d, r-1+i, c-1+j);
+            }
+            VERIFY_IS_EQUAL(twod_patch(d, r, c, patchId), expected);
+          }
+        }
+      }
+    }
+  }
+}
+
+
+static void test_imagenet_patches()
+{
+  // Test the code on typical configurations used by the 'imagenet' benchmarks at
+  // https://github.com/soumith/convnet-benchmarks
+  Tensor<float, 4> l_in(3, 128, 128, 128);
+  l_in.setRandom();
+  Tensor<float, 5> l_out = l_in.extract_image_patches(11, 11);
+  VERIFY_IS_EQUAL(l_out.dimension(0), 3);
+  VERIFY_IS_EQUAL(l_out.dimension(1), 11);
+  VERIFY_IS_EQUAL(l_out.dimension(2), 11);
+  VERIFY_IS_EQUAL(l_out.dimension(3), 128*128);
+  VERIFY_IS_EQUAL(l_out.dimension(4), 128);
+  for (int b = 0; b < 128; ++b) {
+    for (int i = 0; i < 128; ++i) {
+      for (int j = 0; j < 128; ++j) {
+        int patchId = i+128*j;
+        for (int c = 0; c < 11; ++c) {
+          for (int r = 0; r < 11; ++r) {
+            for (int d = 0; d < 3; ++d) {
+              float expected = 0.0f;
+              if (r-5+i >= 0 && c-5+j >= 0 && r-5+i < 128 && c-5+j < 128) {
+                expected = l_in(d, r-5+i, c-5+j, b);
+              }
+              VERIFY_IS_EQUAL(l_out(d, r, c, patchId, b), expected);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  l_in.resize(64, 64, 64, 128);
+  l_in.setRandom();
+  l_out = l_in.extract_image_patches(9, 9);
+  VERIFY_IS_EQUAL(l_out.dimension(0), 64);
+  VERIFY_IS_EQUAL(l_out.dimension(1), 9);
+  VERIFY_IS_EQUAL(l_out.dimension(2), 9);
+  VERIFY_IS_EQUAL(l_out.dimension(3), 64*64);
+  VERIFY_IS_EQUAL(l_out.dimension(4), 128);
+  for (int b = 0; b < 128; ++b) {
+    for (int i = 0; i < 64; ++i) {
+      for (int j = 0; j < 64; ++j) {
+        int patchId = i+64*j;
+        for (int c = 0; c < 9; ++c) {
+          for (int r = 0; r < 9; ++r) {
+            for (int d = 0; d < 64; ++d) {
+              float expected = 0.0f;
+              if (r-4+i >= 0 && c-4+j >= 0 && r-4+i < 64 && c-4+j < 64) {
+                expected = l_in(d, r-4+i, c-4+j, b);
+              }
+              VERIFY_IS_EQUAL(l_out(d, r, c, patchId, b), expected);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  l_in.resize(128, 16, 16, 128);
+  l_in.setRandom();
+  l_out = l_in.extract_image_patches(7, 7);
+  VERIFY_IS_EQUAL(l_out.dimension(0), 128);
+  VERIFY_IS_EQUAL(l_out.dimension(1), 7);
+  VERIFY_IS_EQUAL(l_out.dimension(2), 7);
+  VERIFY_IS_EQUAL(l_out.dimension(3), 16*16);
+  VERIFY_IS_EQUAL(l_out.dimension(4), 128);
+  for (int b = 0; b < 128; ++b) {
+    for (int i = 0; i < 16; ++i) {
+      for (int j = 0; j < 16; ++j) {
+        int patchId = i+16*j;
+        for (int c = 0; c < 7; ++c) {
+          for (int r = 0; r < 7; ++r) {
+            for (int d = 0; d < 128; ++d) {
+              float expected = 0.0f;
+              if (r-3+i >= 0 && c-3+j >= 0 && r-3+i < 16 && c-3+j < 16) {
+                expected = l_in(d, r-3+i, c-3+j, b);
+              }
+              VERIFY_IS_EQUAL(l_out(d, r, c, patchId, b), expected);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  l_in.resize(384, 13, 13, 128);
+  l_in.setRandom();
+  l_out = l_in.extract_image_patches(3, 3);
+  VERIFY_IS_EQUAL(l_out.dimension(0), 384);
+  VERIFY_IS_EQUAL(l_out.dimension(1), 3);
+  VERIFY_IS_EQUAL(l_out.dimension(2), 3);
+  VERIFY_IS_EQUAL(l_out.dimension(3), 13*13);
+  VERIFY_IS_EQUAL(l_out.dimension(4), 128);
+  for (int b = 0; b < 128; ++b) {
+    for (int i = 0; i < 13; ++i) {
+      for (int j = 0; j < 13; ++j) {
+        int patchId = i+13*j;
+        for (int c = 0; c < 3; ++c) {
+          for (int r = 0; r < 3; ++r) {
+            for (int d = 0; d < 384; ++d) {
+              float expected = 0.0f;
+              if (r-1+i >= 0 && c-1+j >= 0 && r-1+i < 13 && c-1+j < 13) {
+                expected = l_in(d, r-1+i, c-1+j, b);
+              }
+              VERIFY_IS_EQUAL(l_out(d, r, c, patchId, b), expected);
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+
+void test_cxx11_tensor_image_patch()
+{
+  CALL_SUBTEST(test_simple_patch());
+  CALL_SUBTEST(test_patch_no_extra_dim());
+  CALL_SUBTEST(test_imagenet_patches());
+}