Update the padding computation for PADDING_SAME to be consistent with TensorFlow.

2 files changed, 56 insertions, 0 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
index 3c6a2e091..91d4ead28 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorImagePatch.h
@@ -265,6 +265,10 @@ struct TensorEvaluator<const TensorImagePatchOp<Rows, Cols, ArgType>, Device>
           // Calculate the padding
           m_rowPaddingTop = ((m_outputRows - 1) * m_row_strides + m_patch_rows_eff - m_input_rows_eff) / 2;
           m_colPaddingLeft = ((m_outputCols - 1) * m_col_strides + m_patch_cols_eff - m_input_cols_eff) / 2;
+          // The padding size calculation for PADDING_SAME has been updated to
+          // be consistent with how TensorFlow extracts its paddings.
+          m_rowPaddingTop = numext::maxi<Index>(0, m_rowPaddingTop);
+          m_colPaddingLeft = numext::maxi<Index>(0, m_colPaddingLeft);
           break;
         default:
           eigen_assert(false && "unexpected padding");
diff --git a/unsupported/test/cxx11_tensor_image_patch.cpp b/unsupported/test/cxx11_tensor_image_patch.cpp
index 475c59651..105d32fb4 100644
--- a/unsupported/test/cxx11_tensor_image_patch.cpp
+++ b/unsupported/test/cxx11_tensor_image_patch.cpp
@@ -405,6 +405,57 @@ void test_patch_padding_same()
   }
 }
 
+// Verifies that SAME padding, when computed as negative values, will be clipped
+// to zero.
+void test_patch_padding_same_negative_padding_clip_to_zero() {
+  int input_depth = 1;
+  int input_rows = 15;
+  int input_cols = 1;
+  int input_batches = 1;
+  int ksize = 1;  // Corresponds to the Rows and Cols for
+                  // tensor.extract_image_patches<>.
+  int row_stride = 5;
+  int col_stride = 1;
+  // ColMajor
+  Tensor<float, 4> tensor(input_depth, input_rows, input_cols, input_batches);
+  // Initializes tensor with incrementing numbers.
+  for (int i = 0; i < tensor.size(); ++i) {
+    tensor.data()[i] = i + 1;
+  }
+  Tensor<float, 5> result = tensor.extract_image_patches(
+      ksize, ksize, row_stride, col_stride, 1, 1, PADDING_SAME);
+  // row padding will be computed as -2 originally and then be clipped to 0.
+  VERIFY_IS_EQUAL(result.coeff(0), 1.0f);
+  VERIFY_IS_EQUAL(result.coeff(1), 6.0f);
+  VERIFY_IS_EQUAL(result.coeff(2), 11.0f);
+
+  VERIFY_IS_EQUAL(result.dimension(0), input_depth);    // depth
+  VERIFY_IS_EQUAL(result.dimension(1), ksize);          // kernel rows
+  VERIFY_IS_EQUAL(result.dimension(2), ksize);          // kernel cols
+  VERIFY_IS_EQUAL(result.dimension(3), 3);              // number of patches
+  VERIFY_IS_EQUAL(result.dimension(4), input_batches);  // number of batches
+
+  // RowMajor
+  Tensor<float, 4, RowMajor> tensor_row_major = tensor.swap_layout();
+  VERIFY_IS_EQUAL(tensor.dimension(0), tensor_row_major.dimension(3));
+  VERIFY_IS_EQUAL(tensor.dimension(1), tensor_row_major.dimension(2));
+  VERIFY_IS_EQUAL(tensor.dimension(2), tensor_row_major.dimension(1));
+  VERIFY_IS_EQUAL(tensor.dimension(3), tensor_row_major.dimension(0));
+
+  Tensor<float, 5, RowMajor> result_row_major =
+      tensor_row_major.extract_image_patches(ksize, ksize, row_stride,
+                                             col_stride, 1, 1, PADDING_SAME);
+  VERIFY_IS_EQUAL(result_row_major.coeff(0), 1.0f);
+  VERIFY_IS_EQUAL(result_row_major.coeff(1), 6.0f);
+  VERIFY_IS_EQUAL(result_row_major.coeff(2), 11.0f);
+
+  VERIFY_IS_EQUAL(result.dimension(0), result_row_major.dimension(4));
+  VERIFY_IS_EQUAL(result.dimension(1), result_row_major.dimension(3));
+  VERIFY_IS_EQUAL(result.dimension(2), result_row_major.dimension(2));
+  VERIFY_IS_EQUAL(result.dimension(3), result_row_major.dimension(1));
+  VERIFY_IS_EQUAL(result.dimension(4), result_row_major.dimension(0));
+}
+
 void test_patch_no_extra_dim()
 {
   Tensor<float, 3> tensor(2,3,5);
@@ -754,4 +805,5 @@ void test_cxx11_tensor_image_patch()
   CALL_SUBTEST_4(test_patch_padding_valid_same_value());
   CALL_SUBTEST_5(test_patch_padding_same());
   CALL_SUBTEST_6(test_imagenet_patches());
+  CALL_SUBTEST_7(test_patch_padding_same_negative_padding_clip_to_zero());
 }