int8:enable change observer algorithm

Author: XiaobingSuper

intel_pytorch_extension_py/__init__.py

@@ -59,16 +59,6 @@ def get_auto_mix_precision():
     else:
         return None
 
-'''
-def quarry_int8_configure(model, inputs_shape):
-    dummy_input = torch.randn(input_shapes).to(DEVICE)
-    core.enable_mix_int8_fp32()
-    with torch.no_grad():
-        y = model(dummy_input)
-    observer_configures = core.get_int8_observer_configures()
-    return observer_configures
-'''
-
 def calibration_reset():
     if core.get_int8_calibration():
         core.calibration_reset()
@@ -104,14 +94,16 @@ def generator_context(*args, **kwargs):
 
 class int8_calibration(_DecoratorContextManager):
     def __init__(self, file_name, observer_configure=None):
-        #self.observer_configure = observer_configure
         self.configure_file = file_name
 
     def __enter__(self):
         if not core.get_mix_int8_fp32():
             raise ValueError("please first run enable_auto_mix_precision(torch.int8) before int8 calibration")
         core.enable_int8_calibration()
-        #core.set_int8_observer_configure(self.observer_configure)
+        if os.path.exists(self.configure_file) and os.stat(self.configure_file).st_size != 0:
+            f = open(self.configure_file)
+            configures = json.load(f)
+            core.load_indicators_file(configures)
 
     def __exit__(self, *args):
         core.disable_int8_calibration()

torch_ipex/csrc/auto_opt_config.h

@@ -64,8 +64,12 @@ class AutoOptConfig {
     std::vector<std::vector<float>> i_min_max_values, std::vector<std::vector<float>> o_min_max_values) {
     num_ops_id++;
     if (observers_.size() < num_ops_id) {
-      // this path is that user not set int8 op's configure, using default configures
-      Observer new_observer = {num_ops_id - 1, op_name, i_min_max_values, o_min_max_values};
+      // this path is that user not set int8 op's configure, using default configures if user not set it.
+      std::string observer_algorithm = "min_max";
+      if (!indicators_.empty()) {
+        observer_algorithm = indicators_[num_ops_id - 1].get_indicator_algorithm();
+      }
+      Observer new_observer = {num_ops_id - 1, op_name, i_min_max_values, o_min_max_values, observer_algorithm};
       observers_.push_back(new_observer);
     } else {
       // user has set configure or have run one interation
@@ -94,24 +98,9 @@ class AutoOptConfig {
     }
   }
 
-  /*
-  inline void print_observer() {
-    for (auto i = 0; i< observers_.size(); i++) {
-      for (auto j = 0; j < observers_[i].max_values.size(); j++)
-        std::cout<<observers_[i].max_values[j]<<std::endl;
-    }
-  }
-  inline void print_indicator() {
-    for (auto i = 0; i< indicators_.size(); i++) {
-      auto scales = indicators_[i].get_indicator_scales();
-      for (auto j = 0; j< scales.size(); j++)
-          std::cout<<scales[j]<<std::endl;
-    }
-  }
-  */
-
   inline void add_indicators() {
     num_ops_id = 0;
+    indicators_.clear();
     // default used is s8
     for (auto i = 0; i < observers_.size(); i++) {
       std::vector<float> inputs_scale, outputs_scale;

torch_ipex/csrc/init_python_bindings.cpp

@@ -147,8 +147,6 @@ void InitIpexModuleBindings(py::module m) {
   m.def("get_int8_calibration", []() { return AutoOptConfig::singleton().get_int8_calibration(); });
   m.def("calibration_reset", []() { AutoOptConfig::singleton().calibration_reset(); });
   m.def("add_indicators", []() { AutoOptConfig::singleton().add_indicators(); });
-  //m.def("print_observer", []() { AutoOptConfig::singleton().print_observer(); });
-  // m.def("print_indicator", []() { AutoOptConfig::singleton().print_indicator(); });
   m.def("get_int8_configures", []() {
       py::list output_list;
       auto indicators = AutoOptConfig::singleton().get_indicators();