!7513 [MSLITE] Add KMeans method for quant

Merge pull request !7513 from ghzl/kmeans-quant

mindspore/lite/schema/model.fbs

@@ -34,6 +34,7 @@ table QuantParam {
     inited: bool = false;
     var_corr: double = 1;
     mean_corr: double = 0;
+    clusters: [float];
 }
 
 table Tensor {

mindspore/lite/src/lite_session.cc

@@ -110,6 +110,12 @@ int LiteSession::ConvertTensors(const lite::Model *model) {
         quant_arg.zeroPoint = quant_params->Get(j)->zeroPoint();
         quant_arg.var_corr = quant_params->Get(j)->var_corr();
         quant_arg.mean_corr = quant_params->Get(j)->mean_corr();
+        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);
       }
     }

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

@@ -83,7 +83,17 @@ class DequantUtil {
       auto scale = param.scale;
       auto zero_point = param.zeroPoint;
       for (int64_t j = 0; j < input_tensor->ElementsNum(); j++) {
-        dequant_datas[j] = static_cast<float>((quant_datas[j] - zero_point) * scale);
+        if (param.clusters.size() != 0) {
+          int8_t index = quant_datas[j];
+          if (index > INT8_MAX || index < INT8_MIN) {
+            MS_LOG(ERROR) << "KMeans param quant is error.";
+            free(dequant_datas);
+            return nullptr;
+          }
+          dequant_datas[j] = static_cast<float>(param.clusters[index - INT8_MIN]);
+        } else {
+          dequant_datas[j] = static_cast<float>((quant_datas[j] - zero_point) * scale);
+        }
       }
     }
     return dequant_datas;

mindspore/lite/src/tensor.h

@@ -35,6 +35,7 @@ struct QuantArg {
   int32_t zeroPoint;
   double var_corr{1};
   double mean_corr{0};
+  std::vector<float> clusters{};
 };
 
 class Tensor : public mindspore::tensor::MSTensor {

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

@@ -101,7 +101,7 @@ STATUS DivergInfo::ComputeThreshold() {
   }
 
   if (method_x == kMethodOutlier) {
-    this->percent_result = PercentMethod(min_datas, max_datas);
+    this->percent_result = OutlierMethod(min_datas, max_datas);
     this->best_T = std::max(std::fabs(percent_result.first), std::fabs(percent_result.second));
     return RET_OK;
   }

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

@@ -20,6 +20,7 @@
 #include <algorithm>
 #include <memory>
 #include <vector>
+#include <set>
 #include "src/ops/primitive_c.h"
 #include "mindspore/lite/tools/converter/quantizer/general_bitpacking.h"
 #include "src/common/utils.h"
@@ -305,8 +306,8 @@ STATUS PostBitPack(float *weight, size_t shapeSize, size_t bitNum) {
   return RET_OK;
 }
 
-bool SearchLowerBound(const std::vector<float> &data, const size_t &index, const float &max_tmp, float *min_tmp,
+static bool SearchLowerBound(const std::vector<float> &data, const size_t &index, const float &max_tmp, float *min_tmp,
-                      size_t *min_idx) {
+                             size_t *min_idx) {
   size_t length = data.size();
   if (max_tmp - data.at(index) < delta) {
     return false;
@@ -320,8 +321,8 @@ bool SearchLowerBound(const std::vector<float> &data, const size_t &index, const
   return true;
 }
 
-bool SearchUpperBound(const std::vector<float> &data, const size_t &index, float *max_tmp, const float &min_tmp,
+static bool SearchUpperBound(const std::vector<float> &data, const size_t &index, float *max_tmp, const float &min_tmp,
-                      size_t *max_idx) {
+                             size_t *max_idx) {
   size_t length = data.size();
   if (data.at(index) - min_tmp < delta) {
     return false;
@@ -335,7 +336,7 @@ bool SearchUpperBound(const std::vector<float> &data, const size_t &index, float
   return true;
 }
 
-float CalPercentile(const std::vector<float> &datas, const int &outlier_percent) {
+static float CalPercentile(const std::vector<float> &datas, const int &outlier_percent) {
   const int size = datas.size();
   float val = outlier_percent / 100.0 * size;
   int index = std::ceil(val);
@@ -348,7 +349,7 @@ float CalPercentile(const std::vector<float> &datas, const int &outlier_percent)
   return result;
 }
 
-std::pair<float, float> PercentMethod(std::vector<float> min_datas, std::vector<float> max_datas) {
+std::pair<float, float> OutlierMethod(std::vector<float> min_datas, std::vector<float> max_datas) {
   std::sort(max_datas.begin(), max_datas.end());
   std::sort(min_datas.begin(), min_datas.end());
   float min_val = CalPercentile(min_datas, percent);
@@ -372,6 +373,64 @@ std::pair<float, float> PercentMethod(std::vector<float> min_datas, std::vector<
   std::pair<float, float> result{min_tmp, max_tmp};
   return result;
 }
+
+static std::vector<float> InitClusters(float *data, size_t elem_count, size_t k) {
+  std::set<float> set_unique{};
+  for (size_t i = 0; i < elem_count; i++) {
+    set_unique.emplace(data[i]);
+  }
+  std::vector<float> data_unique;
+  data_unique.assign(set_unique.begin(), set_unique.end());
+  std::vector<float> clusters{};
+  if (set_unique.size() < k) {
+    return clusters;
+  }
+  // init cluster
+  float ratio = static_cast<float>(data_unique.size()) / (k - 1);
+  std::sort(data_unique.begin(), data_unique.end());
+  for (size_t i = 0; i < k; i++) {
+    size_t index = std::floor(i * ratio);
+    if (i * ratio - index > 0) {
+      clusters.emplace_back((data_unique[index] + data_unique[index + 1]) / 2);
+    } else {
+      clusters.emplace_back(data_unique[index]);
+    }
+  }
+  return clusters;
+}
+
+std::vector<int8_t> KMeans(float *data, size_t elem_count, size_t k, size_t epochs, schema::QuantParamT *quantParam) {
+  std::vector<float> clusters = InitClusters(data, elem_count, k);
+  std::vector<int8_t> clusters_index{};
+  if (clusters.size() < k) {
+    MS_LOG(WARNING) << "K is less than the size of data so KMeans function is not executed.";
+    return clusters_index;
+  }
+  for (size_t epoch = 0; epoch < epochs; epoch++) {
+    clusters_index.clear();
+    std::vector<std::vector<float>> clusters_data(clusters.size());
+    for (size_t i = 0; i < elem_count; i++) {
+      size_t index = 0;
+      float min_distance = pow(data[i] - clusters[0], 2);
+      for (size_t j = 1; j < clusters.size(); j++) {
+        if (pow(data[i] - clusters[j], 2) < min_distance) {
+          min_distance = pow(data[i] - clusters[j], 2);
+          index = j;
+        }
+      }
+      clusters_index.emplace_back(index + INT8_MIN);
+      clusters_data[index].emplace_back(data[i]);
+    }
+    for (size_t j = 0; j < clusters.size(); j++) {
+      if (clusters_data[j].size() > 0) {
+        clusters[j] = std::accumulate(clusters_data[j].begin(), clusters_data[j].end(), 0.0) / clusters_data[j].size();
+      }
+    }
+  }
+  // update data
+  quantParam->clusters = clusters;
+  return clusters_index;
+}
 }  // namespace quant
 }  // namespace lite
 }  // namespace mindspore

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

@@ -72,15 +72,9 @@ STATUS CalQuantizationParams(schema::QuantParamT *quantParam, double mMin, doubl
 STATUS CalQuantizationParams(schema::QuantParamT *quantParam, double mMin, double mMax, bool narrowRange = false,
                              int numBits = UINT8_QUANTIZATION);
 
-bool SearchLowerBound(const std::vector<float> &data, const size_t &index, const float &max_tmp, float *min_tmp,
+std::pair<float, float> OutlierMethod(std::vector<float> min_datas, std::vector<float> max_datas);
-                      size_t *min_idx);
 
-bool SearchUpperBound(const std::vector<float> &data, const size_t &index, float *max_tmp, const float &min_tmp,
+std::vector<int8_t> KMeans(float *data, size_t elem_count, size_t k, size_t epochs, schema::QuantParamT *quantParam);
-                      size_t *max_idx);
-
-float CalPercentile(const std::vector<float> &datas, const int &percent);
-
-std::pair<float, float> PercentMethod(std::vector<float> min_datas, std::vector<float> max_datas);
 
 template <typename T>
 T QuantizeData(const float originData, const schema::QuantParamT *quantParam) {
@@ -213,7 +207,7 @@ STATUS QuantFilter(ParamValueLitePtr weight, std::shared_ptr<PrimitiveC> primiti
           average_raw += raw_data;
         }
       }
-      if (quantType == QuantType_WeightQuant) {
+      if (quantType == QuantType_WeightQuant && quant_param.clusters.size() == 0) {
         // mean
         average_dequant = average_dequant / one_filter_size;
         average_raw = average_raw / one_filter_size;
@@ -261,17 +255,21 @@ STATUS QuantFilter(ParamValueLitePtr weight, std::shared_ptr<PrimitiveC> primiti
     }
 
     schema::QuantParamT quant_param;
-    STATUS status = CalQuantizationParams(&quant_param, min, max, false, quant_max, quant_min, bitNum);
+    if (quant_param.clusters.size() == 0) {
-    if (status != RET_OK) {
+      STATUS status = CalQuantizationParams(&quant_param, min, max, false, quant_max, quant_min, bitNum);
-      MS_LOG(ERROR) << "CalQuantizationParams failed" << status;
+      if (status != RET_OK) {
-      return status;
+        MS_LOG(ERROR) << "CalQuantizationParams failed" << status;
+        return status;
+      }
     }
     quant_params.emplace_back(quant_param);
     // update data and datatype
     for (uint32_t i = 0; i < elem_count; i++) {
       float raw_data = raw_datas[i];
-      auto quant_data = QuantizeData<T>(raw_data, quant_param, quant_max, quant_min);
+      if (quant_param.clusters.size() == 0) {
-      quant_datas[i] = quant_data;
+        auto quant_data = QuantizeData<T>(raw_data, quant_param, quant_max, quant_min);
+        quant_datas[i] = quant_data;
+      }
     }
     auto ret = memcpy_s(raw_datas, weight->tensor_size(), quant_datas.data(), elem_count * sizeof(T));
     if (ret != EOK) {