PerChannel Quantization

mindspore/lite/tools/anf_exporter/anf_exporter.cc

@@ -134,7 +134,7 @@ int AnfExporter::ConvertQuantParam(const std::unique_ptr<schema::MetaGraphT> &me
         for (auto input_quant_param : input_quant_params[i]) {
           std::unique_ptr<schema::QuantParamT> input_quant_param_ptr =
             std::make_unique<schema::QuantParamT>(input_quant_param);
-          MS_LOG(DEBUG) << "[input]node: " << dst_node->name << " scale: " << input_quant_param_ptr->scale
+          MS_LOG(DEBUG) << "[input][" << i << "]node: " << dst_node->name << " scale: " << input_quant_param_ptr->scale
                         << " zp: " << input_quant_param_ptr->zeroPoint;
           tensor_input->quantParams.emplace_back(std::move(input_quant_param_ptr));
         }
@@ -152,7 +152,7 @@ int AnfExporter::ConvertQuantParam(const std::unique_ptr<schema::MetaGraphT> &me
         if (tensor_output->quantParams.empty()) {
           std::unique_ptr<schema::QuantParamT> output_quant_param_ptr =
             std::make_unique<schema::QuantParamT>(output_quant_param);
-          MS_LOG(DEBUG) << "[input]node: " << dst_node->name << " scale: " << output_quant_param_ptr->scale
+          MS_LOG(DEBUG) << "[output]node: " << dst_node->name << " scale: " << output_quant_param_ptr->scale
                         << " zp: " << output_quant_param_ptr->zeroPoint;
           tensor_output->quantParams.emplace_back(std::move(output_quant_param_ptr));
         }

mindspore/lite/tools/converter/quantizer/post_training_quantizer.cc

@@ -536,7 +536,7 @@ STATUS PostTrainingQuantizer::DoQuantOutput(double scale, int zeropoint, struct
 }
 
 STATUS PostTrainingQuantizer::DoWeightQuant(AnfNodePtr weight, std::shared_ptr<PrimitiveTValue> primitiveT_value,
-                                            bool depthwise) {
+                                            bool perchanel, bool depthwise) {
   // const vector<int> dims = filter->dims;
   // perlayer
   if (!weight->isa<Parameter>()) {
@@ -544,9 +544,17 @@ STATUS PostTrainingQuantizer::DoWeightQuant(AnfNodePtr weight, std::shared_ptr<P
     return RET_PARAM_INVALID;
   }
   auto parameter = std::dynamic_pointer_cast<Parameter>(weight);
+  if (parameter == nullptr) {
+    MS_LOG(ERROR) << weight->fullname_with_scope() << " can not cast to Parameter";
+    return RET_ERROR;
+  }
   ParamValueLitePtr paramValue = std::dynamic_pointer_cast<ParamValueLite>(parameter->default_param());
+  if (paramValue == nullptr) {
+    MS_LOG(ERROR) << weight->fullname_with_scope() << " can not get value";
+    return RET_ERROR;
+  }
   auto status = QuantFilter(paramValue, primitiveT_value, QuantType_PostTraining, quant_max, quant_min, bit_num,
-                            per_channel_, depthwise);
+                            perchanel, depthwise);
   if (status != RET_OK) {
     MS_LOG(ERROR) << "QuantFilter failed: " << status;
     return status;
@@ -690,11 +698,29 @@ STATUS PostTrainingQuantizer::QuantNode() {
     auto op_name = cnode->fullname_with_scope();
     auto op_type = primitiveT_value->GetPrimitiveT()->value.type;
     MS_LOG(INFO) << "OpName: " << op_name;
-    if (op_type != PrimitiveType_Conv2D && op_type != PrimitiveType_DepthwiseConv2D) {
+    if (op_type != PrimitiveType_Conv2D && op_type != PrimitiveType_DepthwiseConv2D &&
+        op_type != PrimitiveType_FullConnection) {
       for (size_t i = 1; i < cnode->inputs().size(); i++) {
         auto input_node = cnode->input(i);
         if (!input_node->isa<mindspore::CNode>()) {
-          MS_LOG(WARNING) << "node: " << cnode_name << " input " << i << " not a cnode";
+          MS_LOG(DEBUG) << "node: " << cnode_name << " input " << i << " not a cnode";
+          // get dtype
+          auto abstractBase = input_node->abstract();
+          if (abstractBase == nullptr) {
+            MS_LOG(ERROR) << "Abstract of parameter is nullptr, " << input_node->fullname_with_scope();
+            return RET_ERROR;
+          }
+          if (!utils::isa<abstract::AbstractTensorPtr>(abstractBase)) {
+            MS_LOG(ERROR) << "Abstract of parameter should be anstract tensor, " << input_node->fullname_with_scope();
+            return RET_ERROR;
+          }
+          auto abstractTensor = utils::cast<abstract::AbstractTensorPtr>(abstractBase);
+          if (abstractTensor->element()->GetTypeTrack()->type_id() == kNumberTypeFloat32) {
+            MS_LOG(DEBUG) << "this parameter do quant";
+            DoWeightQuant(input_node, primitiveT_value, false, false);
+          } else {
+            MS_LOG(DEBUG) << "this parameter no need to do quant";
+          }
           continue;
         }
         auto input_cnode = std::dynamic_pointer_cast<mindspore::CNode>(input_node);
@@ -704,8 +730,15 @@ STATUS PostTrainingQuantizer::QuantNode() {
                         << " PrimitiveTValue is null";
           continue;
         }
-        for (auto &quant_param : input_cnode_primitiveT_value->GetOutputQuantParams()) {
-          primitiveT_value->AddInputQuantParam(quant_param);
+        if (!input_cnode_primitiveT_value->GetOutputQuantParams().empty()) {
+          for (auto &quant_param : input_cnode_primitiveT_value->GetOutputQuantParams()) {
+            primitiveT_value->AddInputQuantParam(quant_param);
+          }
+        } else {
+          // do input quant
+          double scale = input_scale[cnode];
+          int32_t zp = input_zero_point[cnode];
+          DoQuantInput(scale, zp, &input_min_max[cnode], primitiveT_value);
         }
       }
     } else {
@@ -715,8 +748,12 @@ STATUS PostTrainingQuantizer::QuantNode() {
       DoQuantInput(scale, convInputzeropoint, &input_min_max[cnode], primitiveT_value);
       // do weight quant
       auto weight = cnode->input(2);
-      bool depthwise = op_type == PrimitiveType_DeDepthwiseConv2D;
-      DoWeightQuant(weight, primitiveT_value, depthwise);
+      bool depthwise = op_type == PrimitiveType_DepthwiseConv2D;
+      bool perchannel = per_channel_;
+      if (op_type == PrimitiveType_FullConnection) {
+        perchannel = false;
+      }
+      DoWeightQuant(weight, primitiveT_value, perchannel, depthwise);
       // do bias quant
       if (cnode->inputs().size() == 4) {
         auto bias = cnode->input(3);

mindspore/lite/tools/converter/quantizer/post_training_quantizer.h

@@ -60,7 +60,7 @@ struct ConfigParam {
 class PostTrainingQuantizer : public Quantizer {
  public:
   PostTrainingQuantizer(FuncGraphPtr graph, std::string path, int bit_num, TypeId target_type = kNumberTypeInt8,
-                        bool per_channel = false);
+                        bool per_channel = true);
 
   STATUS DoQuantize(FuncGraphPtr funcGraph) override;
 
@@ -96,7 +96,8 @@ class PostTrainingQuantizer : public Quantizer {
   STATUS DoQuantInput(double scale, int32_t zeropoint, struct MaxMin *max_min, std::shared_ptr<PrimitiveTValue>);
   STATUS DoQuantOutput(double scale, int32_t zeropoint, struct MaxMin *max_min, std::shared_ptr<PrimitiveTValue>);
 
-  STATUS DoWeightQuant(AnfNodePtr weight, std::shared_ptr<PrimitiveTValue> primitiveT_value, bool depthwise);
+  STATUS DoWeightQuant(AnfNodePtr weight, std::shared_ptr<PrimitiveTValue> primitiveT_value, bool perchannel,
+                       bool depthwise);
 
   STATUS DoBiasQuant(AnfNodePtr bias, std::shared_ptr<PrimitiveTValue> primitiveT_value);
 };

mindspore/lite/tools/converter/quantizer/quantize_util.cc

@@ -100,7 +100,7 @@ bool QuantStrategy::CanOpPostQuantized(AnfNodePtr &node) const {
     schema::PrimitiveType_Conv2D,    schema::PrimitiveType_DepthwiseConv2D,
     schema::PrimitiveType_Add,       schema::PrimitiveType_Pooling,
     schema::PrimitiveType_Concat,    /*schema::PrimitiveType_SoftMax,*/
-    schema::PrimitiveType_Reshape,   /*schema::PrimitiveType_FullConnection,*/
+    schema::PrimitiveType_Reshape,   schema::PrimitiveType_FullConnection,
     schema::PrimitiveType_MatMul,
     schema::PrimitiveType_Activation};
   return IsContain(uint8OpList, type);