!8124 change shema of quant_params to reduce model size

Merge pull request !8124 from xutianchun/1102

mindspore/lite/schema/model.fbs

@@ -32,10 +32,9 @@ table QuantParam {
     narrowRange: bool = true;
     numBits: int = 8;
     inited: bool = false;
-    varCorr: double = 1;
-    meanCorr: double = 0;
+    varCorr: float = 1;
+    meanCorr: float = 0;
     dstDtype: int = 32;
-    clusters: [float];
 }
 
 table Tensor {
@@ -49,6 +48,7 @@ table Tensor {
     offset: int;
     data: [ubyte];
     quantParams: [QuantParam];
+    quantClusters: [float];
 }
 
 union PrimitiveType {

mindspore/lite/src/lite_session.cc

@@ -107,15 +107,17 @@ int LiteSession::ConvertTensors(const lite::Model *model) {
         quant_arg.var_corr = quant_params->Get(j)->varCorr();
         quant_arg.mean_corr = quant_params->Get(j)->meanCorr();
         quant_arg.inited = quant_params->Get(j)->inited();
-        auto quant_clusters = quant_params->Get(j)->clusters();
-        if (quant_clusters != nullptr) {
-          for (size_t k = 0; k < quant_clusters->size(); k++) {
-            quant_arg.clusters.emplace_back(quant_clusters->Get(k));
-          }
-        }
         dstTensor->AddQuantParam(quant_arg);
       }
     }
+    auto quant_clusters = srcTensor->quantClusters();
+    if (quant_clusters != nullptr) {
+      std::vector<float> clusters;
+      for (size_t j = 0; j < quant_clusters->size(); j++) {
+        clusters.push_back(quant_clusters->Get(j));
+      }
+      dstTensor->SetQuantClusters(clusters);
+    }
     this->tensors_.emplace_back(dstTensor);
   }
 

mindspore/lite/src/runtime/kernel/arm/base/dequant.h

@@ -79,11 +79,12 @@ class DequantUtil {
       }
     } else {
       auto quant_param = input_tensor->GetQuantParams();
+      auto quant_clusters = input_tensor->GetQuantClusters();
       auto param = quant_param.front();
       auto scale = param.scale;
       auto zero_point = param.zeroPoint;
       for (int64_t j = 0; j < input_tensor->ElementsNum(); j++) {
-        if (param.clusters.size() != 0) {
+        if (!quant_clusters.empty()) {
           int8_t index = quant_datas[j];
           if (index > INT8_MAX || index < INT8_MIN) {
             MS_LOG(ERROR) << "KMeans param quant is error.";

mindspore/lite/src/tensor.cc

@@ -367,6 +367,10 @@ void Tensor::AddQuantParam(const QuantArg &quant_arg) { this->quant_params_.push
 
 std::vector<QuantArg> Tensor::GetQuantParams() const { return this->quant_params_; }
 
+std::vector<float> Tensor::GetQuantClusters() const { return this->quant_clusters_; }
+
+void Tensor::SetQuantClusters(const std::vector<float> &clusters) { this->quant_clusters_ = clusters; }
+
 std::vector<tensor::MSTensor *> TensorVectorCast(const std::vector<Tensor *> &src) {
   std::vector<tensor::MSTensor *> target(src.size());
   std::transform(src.begin(), src.end(), target.begin(), [](Tensor *t) { return dynamic_cast<tensor::MSTensor *>(t); });

mindspore/lite/src/tensor.h

@@ -33,8 +33,8 @@ namespace lite {
 struct QuantArg {
   double scale;
   int32_t zeroPoint;
-  double var_corr{1};
-  double mean_corr{0};
+  float var_corr{1};
+  float mean_corr{0};
   bool inited;
   std::vector<float> clusters{};
 };
@@ -119,6 +119,10 @@ class Tensor : public mindspore::tensor::MSTensor {
 
   std::vector<QuantArg> GetQuantParams() const;
 
+  std::vector<float> GetQuantClusters() const;
+
+  void SetQuantClusters(const std::vector<float> &clusters);
+
   bool IsConst();
 
   bool IsScalar();
@@ -138,6 +142,7 @@ class Tensor : public mindspore::tensor::MSTensor {
   Category category_;
   size_t ref_count_ = 0;
   std::vector<QuantArg> quant_params_;
+  std::vector<float> quant_clusters_;
   mindspore::lite::Allocator *allocator_ = nullptr;
 };
 

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

@@ -449,7 +449,6 @@ std::vector<int8_t> KMeans(float *data, size_t elem_count, size_t k, size_t epoc
     error = error_cur;
   }
   // update data
-  quantParam->clusters = clusters;
   return clusters_index;
 }
 

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

@@ -130,7 +130,7 @@ T QuantizeData(float originData, const schema::QuantParamT &quantParam, int quan
 }
 template <typename T>
 STATUS QuantFilter(ParamValueLitePtr weight, std::shared_ptr<PrimitiveC> primitive_c, QuantType quantType,
-                   int quant_max, int quant_min, size_t bitNum, bool per_channel) {
+                   int quant_max, int quant_min, size_t bitNum, bool per_channel, bool k_means = false) {
   auto dims = weight->tensor_shape();
   auto op_type = (schema::PrimitiveType)primitive_c->Type();
   if (per_channel) {
@@ -208,7 +208,7 @@ STATUS QuantFilter(ParamValueLitePtr weight, std::shared_ptr<PrimitiveC> primiti
           average_raw += raw_data;
         }
       }
-      if (quantType == QuantType_WeightQuant && quant_param.clusters.size() == 0) {
+      if (quantType == QuantType_WeightQuant && !k_means) {
         // mean
         average_dequant = average_dequant / one_filter_size;
         average_raw = average_raw / one_filter_size;
@@ -256,7 +256,7 @@ STATUS QuantFilter(ParamValueLitePtr weight, std::shared_ptr<PrimitiveC> primiti
     }
 
     schema::QuantParamT quant_param;
-    if (quant_param.clusters.size() == 0) {
+    if (!k_means) {
       STATUS status = CalQuantizationParams(&quant_param, min, max, false, quant_max, quant_min, bitNum);
       if (status != RET_OK) {
         MS_LOG(ERROR) << "CalQuantizationParams failed" << status;
@@ -267,7 +267,7 @@ STATUS QuantFilter(ParamValueLitePtr weight, std::shared_ptr<PrimitiveC> primiti
     // update data and datatype
     for (uint32_t i = 0; i < elem_count; i++) {
       float raw_data = raw_datas[i];
-      if (quant_param.clusters.size() == 0) {
+      if (!k_means) {
         auto quant_data = QuantizeData<T>(raw_data, quant_param, quant_max, quant_min);
         quant_datas[i] = quant_data;
       }