033-conv_bn_fusion_pass-5

mindspore/lite/tools/converter/anf_transform.cc

@@ -111,8 +111,12 @@ FuncGraphPtr AnfTransform::TransformSingleFuncGraph(const FuncGraphPtr &old_grap
     // remove quantdtype when awaretraining
     fusion_pm->AddPass(std::make_shared<opt::RemoveIdentityOpPass>());
     fusion_pm->AddPass(std::make_shared<opt::ConvBiasaddFusion>());
-    fusion_pm->AddPass(std::make_shared<opt::ConvBatchNormFusion>());
-    fusion_pm->AddPass(std::make_shared<opt::ConvScaleFusion>());
+    auto conv_bn_pass = std::make_shared<opt::ConvBatchNormFusion>();
+    conv_bn_pass->SetFmkType(config->fmk);
+    fusion_pm->AddPass(conv_bn_pass);
+    auto conv_scale_pass = std::make_shared<opt::ConvScaleFusion>();
+    conv_scale_pass->SetFmkType(config->fmk);
+    fusion_pm->AddPass(conv_scale_pass);
     fusion_pm->AddPass(std::make_shared<opt::LayerNormFusion>());
     fusion_pm->AddPass(std::make_shared<opt::BatchMatMulFusion>());
     fusion_pm->AddPass(std::make_shared<opt::SigmoidMulFusion>());

mindspore/lite/tools/converter/parser/tf/tf_batchnorm_parser.cc

@@ -36,7 +36,7 @@ STATUS TFBatchNormParser::Parse(const tensorflow::NodeDef &tf_op,
     MS_LOG(ERROR) << "primitive is nullptr";
     return RET_NULL_PTR;
   }
-  auto attr = std::make_unique<schema::BatchNormT>();
+  auto attr = std::make_unique<schema::FusedBatchNormT>();
   if (attr == nullptr) {
     MS_LOG(ERROR) << "new op failed";
     return RET_NULL_PTR;
@@ -45,7 +45,7 @@ STATUS TFBatchNormParser::Parse(const tensorflow::NodeDef &tf_op,
   TensorFlowUtils::FindAttrValue(tf_op, "epsilon", &attr_value);
   attr->epsilon = attr_value.f();
 
-  primitive->value.type = schema::PrimitiveType_BatchNorm;
+  primitive->value.type = schema::PrimitiveType_FusedBatchNorm;
   primitive->value.value = attr.release();
   *primitiveC = PrimitiveC::Create(primitive.release());
   if (*primitiveC == nullptr) {

mindspore/lite/tools/optimizer/fusion/conv_transform_fusion.cc

@@ -202,10 +202,15 @@ void ConvTransformFusion::CalNewWeightTensor(float *weight_data, int kernel_num,
     lite::ReturnCode::GetSingleReturnCode()->UpdateReturnCode(lite::RET_MEMORY_FAILED);
     return;
   }
-
-  for (int i = 0; i < kernel_num; i++) {
-    for (int j = 0; j < kernel_size; j++) {
-      tmp_weight_data[i * kernel_size + j] = weight_data[i * kernel_size + j] * trans_scale[i];
+  if (this->fmk_type_ == lite::converter::FmkType_TF) {
+    for (int i = 0; i < kernel_num * kernel_size; i++) {
+      tmp_weight_data[i] = weight_data[i] * trans_scale[i % kernel_num];
+    }
+  } else {
+    for (int i = 0; i < kernel_num; i++) {
+      for (int j = 0; j < kernel_size; j++) {
+        tmp_weight_data[i * kernel_size + j] = weight_data[i * kernel_size + j] * trans_scale[i];
+      }
     }
   }
 

mindspore/lite/tools/optimizer/fusion/conv_transform_fusion.h

@@ -19,7 +19,9 @@
 
 #include <string>
 #include "backend/optimizer/common/optimizer.h"
+#include "tools/converter/converter_flags.h"
 
+using mindspore::lite::converter::FmkType;
 namespace mindspore::opt {
 class ConvTransformFusion : public PatternProcessPass {
  public:
@@ -32,6 +34,10 @@ class ConvTransformFusion : public PatternProcessPass {
   void GenNewConvTensor(const FuncGraphPtr &, const CNodePtr &, int, const float *, const float *) const;
   void CalNewWeightTensor(float *, int, int, const float *) const;
   void CalNewBiasTensor(float *, int, bool, const float *, const float *) const;
+  void SetFmkType(FmkType type) { this->fmk_type_ = type; }
+
+ private:
+  FmkType fmk_type_ = lite::converter::FmkType_TF;
 };
 }  // namespace mindspore::opt
 #endif  // MINDSPORE_LITE_SRC_PASS_FUSION_CONV_TRANSFORM_FUSION_H_