Remove calibration file path when deploy quantize model (#46283)

* remove calibration file path

* remove useless code

paddle/fluid/framework/ir/mkldnn/quant_dequant_mkldnn_pass.cc

@@ -716,10 +716,6 @@ void QuantDequantMkldnnPass::ApplyImpl(ir::Graph* graph) const {
   std::unordered_map<std::string, std::vector<float>> var_quant_scales{};
   bool onnx_format_quantize_model = false;
   auto* scope = param_scope();
-  GetInfoFromTheFirstOp(
-      graph, "has_quant_info", "var_quant_scales", &var_quant_scales);
-  VLOG(1) << "The nums of scale info from slim txt is: "
-          << var_quant_scales.size();
   MarkSkipQuantizedOps(graph, skip_ops);
   CollectInfoFromFake(graph, scope, fake_dequantize_types, &weight_thresholds);
   CollectWeightScalesInfoFromONNXFormatDequantize(graph,

paddle/fluid/inference/analysis/argument.h

@@ -177,9 +177,6 @@ struct Argument {
   DECL_ARGUMENT_FIELD(mkldnn_cache_capacity, MkldnnCacheCapacity, int);
 
 #ifdef PADDLE_WITH_MKLDNN
-  // Calibration file path of quantize model
-  DECL_ARGUMENT_FIELD(calibration_file_path, CalibrationFilePath, std::string);
-
   // A set of op types to enable their quantized kernels
   DECL_ARGUMENT_FIELD(quantize_enabled_op_types,
                       QuantizeEnabledOpTypes,

paddle/fluid/inference/analysis/passes/ir_analysis_pass.cc

@@ -36,19 +36,6 @@ void IrAnalysisPass::RunImpl(Argument* argument) {
   auto* the_graph = argument->ReleaseMainGraph();
   auto graph = std::unique_ptr<Graph>(the_graph);
 
-#ifdef PADDLE_WITH_MKLDNN
-  if (argument->Has("calibration_file_path")) {
-    VLOG(5) << "Calibration file path of quantize model: "
-            << argument->calibration_file_path();
-    std::unordered_map<std::string, std::vector<float>> var_quant_scales{};
-    ReadCalibrationInfo(argument, &var_quant_scales);
-    // save var_quant_scales in the first op's attr
-    // for quant_dequant_mkldnn_pass
-    SaveInfoInTheFirstOp(
-        the_graph, "has_quant_info", "var_quant_scales", var_quant_scales);
-  }
-#endif
-
   // Apply passes.
   IRPassManager the_ir_manager(argument);
   graph = the_ir_manager.Apply(std::move(graph));
@@ -61,40 +48,6 @@ void IrAnalysisPass::RunImpl(Argument* argument) {
   CollectFusionStatis(argument);
 }
 
-void IrAnalysisPass::ReadCalibrationInfo(
-    Argument* argument,
-    std::unordered_map<std::string, std::vector<float>>* var_quant_scales) {
-  std::string calibration_file_path;
-#ifdef PADDLE_WITH_MKLDNN
-  if (argument->Has("calibration_file_path")) {
-    calibration_file_path = argument->calibration_file_path();
-  }
-#endif
-  if (calibration_file_path.empty()) {
-    LOG(INFO) << "argument has no calibration_file_path";
-    return;
-  }
-  std::ifstream calibration_file(calibration_file_path);
-  std::string one_line;
-  while (getline(calibration_file, one_line)) {
-    if (one_line.find(" ") != one_line.npos) {
-      auto pos = one_line.find(" ");
-      std::string pre_str = one_line.substr(0, pos);
-      std::string pos_str = one_line.substr(pos);
-      if (pre_str.size() && pos_str.size()) {
-        std::string tensor_name = pre_str;
-        float scale = std::stod(pos_str);
-        scale = 1.0 / scale;
-        if (std::isinf(scale) || std::isnan(scale)) {
-          continue;
-        }
-        std::vector<float> scales = {scale};
-        (*var_quant_scales)[tensor_name] = scales;
-      }
-    }
-  }
-}
-
 void IrAnalysisPass::CollectFusionStatis(Argument* argument) {
   if (!argument->main_graph().Has(framework::ir::kFuseStatisAttr)) {
     LOG(INFO) << "argument has no fuse statis";

paddle/fluid/inference/api/analysis_config.cc

@@ -246,7 +246,6 @@ AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
   CP_MEMBER(opt_cache_dir_);
   CP_MEMBER(prog_file_);
   CP_MEMBER(params_file_);
-  CP_MEMBER(calibration_file_path_);
 
   CP_MEMBER(use_fc_padding_);
   // GPU related.
@@ -518,14 +517,6 @@ void AnalysisConfig::EnableMkldnnInt8(
   Update();
 }
 
-void AnalysisConfig::SetCalibrationFilePath(
-    const std::string &calibration_file_path) {
-  calibration_file_path_ = calibration_file_path;
-  VLOG(1) << "Set calibration file path of quantize model: " +
-                 calibration_file_path_;
-  Update();
-}
-
 MkldnnQuantizerConfig *AnalysisConfig::mkldnn_quantizer_config() const {
   PADDLE_ENFORCE_NOT_NULL(mkldnn_quantizer_config_,
                           platform::errors::PreconditionNotMet(
@@ -850,8 +841,6 @@ std::string AnalysisConfig::SerializeInfoCache() {
   ss << prog_file_;
   ss << params_file_;
 
-  ss << calibration_file_path_;
-
   ss << use_gpu_;
   ss << use_external_stream_;
   ss << exec_stream_;
@@ -1039,9 +1028,6 @@ std::string AnalysisConfig::Summary() {
     os.InsertRow({"model_file", prog_file_});
     os.InsertRow({"params_file", params_file_});
   }
-  if (!(calibration_file_path_.empty())) {
-    os.InsertRow({"calibration_file_path", calibration_file_path_});
-  }
 
   if (model_from_memory_) {
     os.InsertRow({"model_from_memory", params_file_});

paddle/fluid/inference/api/analysis_predictor.cc

@@ -1189,7 +1189,6 @@ void AnalysisPredictor::PrepareArgument() {
     argument_.SetQuantizeEnabledOpTypes(config_.quantize_enabled_op_types_);
     argument_.SetQuantizeExcludedOpIds(config_.quantize_excluded_op_ids_);
     argument_.SetQuantVarScales({});
-    argument_.SetCalibrationFilePath(config_.calibration_file_path_);
   }
 #endif
 

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -770,18 +770,6 @@ struct PD_INFER_DECL AnalysisConfig {
   ///
   void EnableMkldnnQuantizer();
 
-  ///
-  /// \brief Set the calibration ranges file path of quantize model.
-  ///
-  ///
-  void SetCalibrationFilePath(const std::string& calibration_file_path = "");
-
-  ///
-  /// \brief Return the calibration ranges file path of quantize model.
-  ///
-  ///
-  std::string CalibrationFilePath() { return calibration_file_path_; }
-
   ///
   /// \brief Turn on MKLDNN int8.
   ///
@@ -960,7 +948,6 @@ struct PD_INFER_DECL AnalysisConfig {
   std::string model_dir_;
   mutable std::string prog_file_;
   mutable std::string params_file_;
-  mutable std::string calibration_file_path_;
 
   // Mixed precision.
   std::unordered_set<std::string> mixed_black_list_;

paddle/fluid/pybind/inference_api.cc

@@ -782,9 +782,6 @@ void BindAnalysisConfig(py::module *m) {
       .def("to_native_config", &AnalysisConfig::ToNativeConfig)
       .def("enable_quantizer", &AnalysisConfig::EnableMkldnnQuantizer)
       .def("enable_mkldnn_bfloat16", &AnalysisConfig::EnableMkldnnBfloat16)
-      .def("set_calibration_file_path",
-           &AnalysisConfig::SetCalibrationFilePath,
-           py::arg("calibration_file_path") = std::string(""))
 #ifdef PADDLE_WITH_MKLDNN
       .def("quantizer_config",
            &AnalysisConfig::mkldnn_quantizer_config,

python/paddle/fluid/tests/unittests/mkldnn/test_onnx_format_quantization_mobilenetv1.py

@@ -210,9 +210,6 @@ class TestPostTrainingQuantization(unittest.TestCase):
         config.set_cpu_math_library_num_threads(1)
         config.disable_glog_info()
         if is_quantized_model:
-            calibration_file_path = os.path.join(model_path,
-                                                 'calibration_table.txt')
-            config.set_calibration_file_path(calibration_file_path)
             config.enable_mkldnn_int8()
         predictor = paddle.inference.create_predictor(config)