Fold BN after depthwise conv

2 files changed, 19 insertions, 22 deletions

diff --git a/tensorflow/tools/graph_transforms/fold_batch_norms.cc b/tensorflow/tools/graph_transforms/fold_batch_norms.cc
index 975b17380f..42eebd98c9 100644
--- a/tensorflow/tools/graph_transforms/fold_batch_norms.cc
+++ b/tensorflow/tools/graph_transforms/fold_batch_norms.cc
@@ -38,7 +38,7 @@ Status FoldBatchNorms(const GraphDef& input_graph_def,
       input_graph_def,  // clang-format off
       {"Mul",                // mul_node
         {
-          {"Conv2D|MatMul",  // conv_node
+          {"Conv2D|MatMul|DepthwiseConv2dNative",  // conv_node
             {
               {"*"},         // input_node
               {"Const"},     // weights_node
@@ -73,7 +73,7 @@ Status FoldBatchNorms(const GraphDef& input_graph_def,
 
         // Make sure all the inputs really are vectors, with as many entries as
         // there are columns in the weights.
-        const int weights_cols_index = conv_node.op() == "Conv2D" ? 3 : 1;
+        const int weights_cols_index = conv_node.op() == "Conv2D" ? 3 : (conv_node.op() == "DepthwiseConv2dNative" ? 2 : 1);
         const int64 weights_cols = weights.shape().dim_size(weights_cols_index);
         if ((mul_values.shape().dims() != 1) ||
             (mul_values.shape().dim_size(0) != weights_cols)) {
@@ -83,14 +83,12 @@ Status FoldBatchNorms(const GraphDef& input_graph_def,
         }
 
         // Multiply the original weights by the scale vector.
-        auto weights_matrix = weights.flat_inner_dims<float>();
+        auto weights_vector = weights.flat<float>();
         Tensor scaled_weights(DT_FLOAT, weights.shape());
-        auto scaled_weights_matrix = scaled_weights.flat_inner_dims<float>();
-        for (int64 row = 0; row < weights_matrix.dimension(0); ++row) {
-          for (int64 col = 0; col < weights_cols; ++col) {
-            scaled_weights_matrix(row, col) =
-                weights_matrix(row, col) * mul_values.flat<float>()(col);
-          }
+        auto scaled_weights_vector = scaled_weights.flat<float>();
+        for (int64 row = 0; row < weights_vector.dimension(0); ++row) {
+          scaled_weights_vector(row) =
+                weights_vector(row) * mul_values.flat<float>()(row % weights_cols);
         }
 
         // Construct the new nodes.
diff --git a/tensorflow/tools/graph_transforms/fold_old_batch_norms.cc b/tensorflow/tools/graph_transforms/fold_old_batch_norms.cc
index f1d361e07d..1a4b141d0e 100644
--- a/tensorflow/tools/graph_transforms/fold_old_batch_norms.cc
+++ b/tensorflow/tools/graph_transforms/fold_old_batch_norms.cc
@@ -110,24 +110,23 @@ Status FuseScaleOffsetToConvWeights(const std::vector<float>& scale_values,
                                     const string& conv_output_name,
                                     std::vector<NodeDef>* new_nodes) {
   const NodeDef& conv_node = conv_node_match.node;
-  CHECK_EQ("Conv2D", conv_node.op());
+  // CHECK_EQ("Conv2D", conv_node.op());
   const NodeDef& input_node = conv_node_match.inputs[0].node;
   const NodeDef& weights_node = conv_node_match.inputs[1].node;
   CHECK_EQ("Const", weights_node.op());
 
   Tensor weights = GetNodeTensorAttr(weights_node, "value");
-  const int64 weights_cols = weights.shape().dim_size(3);
+  const int weights_cols_idx = conv_node.op() == "Conv2D" ? 3 : 2;
+  const int64 weights_cols = weights.shape().dim_size(weights_cols_idx);
   CHECK_EQ(weights_cols, scale_values.size());
 
   // Multiply the original weights by the scale vector.
-  auto weights_matrix = weights.flat_inner_dims<float>();
+  auto weights_vector = weights.flat<float>();
   Tensor scaled_weights(DT_FLOAT, weights.shape());
-  auto scaled_weights_matrix = scaled_weights.flat_inner_dims<float>();
-  for (int64 row = 0; row < weights_matrix.dimension(0); ++row) {
-    for (int64 col = 0; col < weights_cols; ++col) {
-      scaled_weights_matrix(row, col) =
-          weights_matrix(row, col) * scale_values[col];
-    }
+  auto scaled_weights_vector = scaled_weights.flat<float>();
+  for (int64 row = 0; row < weights_vector.dimension(0); ++row) {
+    scaled_weights_vector(row) =
+          weights_vector(row) * scale_values[row % weights_cols];
   }
   // Figure out the remaining bias to add on.
   Tensor bias_offset(DT_FLOAT, {weights_cols});
@@ -293,7 +292,7 @@ Status FoldOldBatchNorms(const GraphDef& input_graph_def,
         current_graph_def,  // clang-format off
       {"BatchNormWithGlobalNormalization|FusedBatchNorm",    // batch_norm_node
         {
-          {"Conv2D",                          // conv_node
+          {"Conv2D|DepthwiseConv2dNative",                          // conv_node
             {
               {"*"},                          // input_node
               {"Const"},                      // weights_node
@@ -326,7 +325,7 @@ Status FoldOldBatchNorms(const GraphDef& input_graph_def,
          {
              {"BatchToSpaceND",                  // batch_to_space_node
               {
-                  {"Conv2D",                     // conv_node
+                  {"Conv2D|DepthwiseConv2dNative",                     // conv_node
                    {
                        {"*"},                    // input_node
                        {"Const"},                // weights_node
@@ -364,13 +363,13 @@ Status FoldOldBatchNorms(const GraphDef& input_graph_def,
         {
           {"ConcatV2|Concat",                     // concat two conv2d.
             {
-              {"Conv2D",                          // conv_node
+              {"Conv2D|DepthwiseConv2dNative",                          // conv_node
                 {
                   {"*"},                          // input_node
                   {"Const"},                      // weights_node
                 }
               },
-              {"Conv2D",                          // conv_node
+              {"Conv2D|DepthwiseConv2dNative",                          // conv_node
                 {
                   {"*"},                          // input_node
                   {"Const"},                      // weights_node