Home | History | Annotate | Download | only in kernels
      1 /* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
      2 
      3 Licensed under the Apache License, Version 2.0 (the "License");
      4 you may not use this file except in compliance with the License.
      5 You may obtain a copy of the License at
      6 
      7     http://www.apache.org/licenses/LICENSE-2.0
      8 
      9 Unless required by applicable law or agreed to in writing, software
     10 distributed under the License is distributed on an "AS IS" BASIS,
     11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
     12 See the License for the specific language governing permissions and
     13 limitations under the License.
     14 ==============================================================================*/
     15 
     16 #include "tensorflow/core/kernels/winograd_transform.h"
     17 #include "tensorflow/core/platform/test.h"
     18 
     19 namespace tensorflow {
     20 namespace {
     21 
     22 static void ComputeKroneckerProduct(const int rows, const int cols,
     23                                     const float* matrix, float* matrix_out) {
     24   for (int i = 0; i < rows; ++i) {
     25     for (int j = 0; j < cols; ++j) {
     26       const float v = matrix[i * cols + j];
     27       const int output_index_base = cols * (i * rows * cols + j);
     28       for (int k = 0; k < rows; ++k) {
     29         for (int l = 0; l < cols; ++l) {
     30           const int input_index = k * cols + l;
     31           const int output_index = k * cols * cols + l;
     32           matrix_out[output_index_base + output_index] =
     33               matrix[input_index] * v;
     34         }
     35       }
     36     }
     37   }
     38 }
     39 
     40 TEST(DeepConv2DTransformTest, Basic) {
     41   // Tests kronecker product of the following matrix with itself:
     42   //
     43   // [1.0 2.0]
     44   // [3.0 4.0]
     45   //
     46   const int rows = 2;
     47   const int cols = 2;
     48 
     49   float transform_matrix[] = {1, 2, 3, 4};
     50 
     51   const int kron_rows = rows * rows;
     52   const int kron_cols = cols * cols;
     53   float transform_matrix_kron[kron_rows * kron_cols];
     54 
     55   ComputeKroneckerProduct(rows, cols, &transform_matrix[0],
     56                           &transform_matrix_kron[0]);
     57 
     58   float transform_matrix_test[] = {1, 2, 2, 4, 3, 4,  6,  8,
     59                                    3, 6, 4, 8, 9, 12, 12, 16};
     60 
     61   for (int i = 0; i < kron_rows * kron_cols; ++i) {
     62     EXPECT_FLOAT_EQ(transform_matrix_kron[i], transform_matrix_test[i]);
     63   }
     64 }
     65 
     66 TEST(DeepConv2DTransformTest, WingradFilterTransformMatrix) {
     67   // Test that the filter transform matrix returned is the kronecker product of
     68   // the following matrix with itself:
     69   //
     70   //   [ 1    0   0   ]
     71   //   [ 1/2  1/2 1/2 ]
     72   //   [ 1/2 -1/2 1/2 ]
     73   //   [ 0    0   1   ]
     74   //
     75   const int rows = 4;
     76   const int cols = 3;
     77 
     78   float transform_matrix[] = {1, 0, 0, 0.5, 0.5, 0.5, 0.5, -0.5, 0.5, 0, 0, 1};
     79 
     80   const int kron_rows = rows * rows;
     81   const int kron_cols = cols * cols;
     82 
     83   float transform_matrix_kron[kron_rows * kron_cols];
     84 
     85   ComputeKroneckerProduct(rows, cols, &transform_matrix[0],
     86                           &transform_matrix_kron[0]);
     87 
     88   float transform_matrix_test[kron_rows * kron_cols];
     89   WinogradTransform<float> t;
     90   t.GetFilterTransformMatrix(kron_rows, kron_cols, &transform_matrix_test[0]);
     91 
     92   for (int i = 0; i < kron_rows * kron_cols; ++i) {
     93     EXPECT_FLOAT_EQ(transform_matrix_kron[i], transform_matrix_test[i]);
     94   }
     95 }
     96 
     97 TEST(DeepConv2DTransformTest, WingradInputTransformMatrix) {
     98   // Test that the filter transform matrix returned is the kronecker product of
     99   // the following matrix:
    100   //
    101   //   [1   0  -1   0]
    102   //   [0   1   1   0]
    103   //   [0  -1   1   0]
    104   //   [0   1   0  -1]
    105   //
    106   const int rows = 4;
    107   const int cols = 4;
    108 
    109   float transform_matrix[] = {1, 0,  -1, 0, 0, 1, 1, 0,
    110                               0, -1, 1,  0, 0, 1, 0, -1};
    111 
    112   const int kron_rows = rows * rows;
    113   const int kron_cols = cols * cols;
    114 
    115   float transform_matrix_kron[kron_rows * kron_cols];
    116 
    117   ComputeKroneckerProduct(rows, cols, &transform_matrix[0],
    118                           &transform_matrix_kron[0]);
    119 
    120   float transform_matrix_test[kron_rows * kron_cols];
    121   WinogradTransform<float> t;
    122   t.GetInputTransformMatrix(kron_rows, kron_cols, &transform_matrix_test[0]);
    123 
    124   for (int i = 0; i < kron_rows * kron_cols; ++i) {
    125     EXPECT_FLOAT_EQ(transform_matrix_kron[i], transform_matrix_test[i]);
    126   }
    127 }
    128 
    129 TEST(DeepConv2DTransformTest, WingradOutputTransformMatrix) {
    130   // Test that the filter transform matrix returned is the kronecker product of
    131   // the following matrix:
    132   //
    133   //   [1  1  1  0]
    134   //   [0  1 -1 -1]
    135   //
    136   const int rows = 2;
    137   const int cols = 4;
    138 
    139   float transform_matrix[] = {1, 1, 1, 0, 0, 1, -1, -1};
    140 
    141   const int kron_rows = rows * rows;
    142   const int kron_cols = cols * cols;
    143 
    144   float transform_matrix_kron[kron_rows * kron_cols];
    145 
    146   ComputeKroneckerProduct(rows, cols, &transform_matrix[0],
    147                           &transform_matrix_kron[0]);
    148 
    149   float transform_matrix_test[kron_rows * kron_cols];
    150   WinogradTransform<float> t;
    151   t.GetOutputTransformMatrix(kron_rows, kron_cols, &transform_matrix_test[0]);
    152 
    153   for (int i = 0; i < kron_rows * kron_cols; ++i) {
    154     EXPECT_FLOAT_EQ(transform_matrix_kron[i], transform_matrix_test[i]);
    155   }
    156 }
    157 
    158 }  // namespace
    159 }  // namespace tensorflow
    160