preset is optional parameter of nncf.quantize(...)

nncf/experimental/torch/quantization/quantize_model.py

@@ -87,7 +87,7 @@ def send_to_device(tensor):
 def quantize_impl(
     model: torch.nn.Module,
     calibration_dataset: Dataset,
-    preset: QuantizationPreset,
+    preset: Optional[QuantizationPreset],
     target_device: TargetDevice,
     subset_size: int,
     fast_bias_correction: bool,

nncf/onnx/quantization/quantize_model.py

@@ -31,7 +31,7 @@
 def quantize_impl(
     model: onnx.ModelProto,
     calibration_dataset: Dataset,
-    preset: QuantizationPreset,
+    preset: Optional[QuantizationPreset],
     target_device: TargetDevice,
     subset_size: int,
     fast_bias_correction: bool,

nncf/openvino/pot/quantization/quantize_model.py

@@ -192,7 +192,7 @@ def _create_quantization_group_config(
 
 
 def _create_quantization_config(
-    preset: QuantizationPreset,
+    preset: Optional[QuantizationPreset],
     target_device: TargetDevice,
     subset_size: int,
     fast_bias_correction: bool,
@@ -224,6 +224,12 @@ def _create_quantization_config(
         fine-tuning the quantization algorithm.
     :return: A POT quantization configuration as dict.
     """
+    if preset is None:
+        if model_type == ModelType.TRANSFORMER:
+            preset = QuantizationPreset.MIXED
+        else:
+            preset = QuantizationPreset.PERFORMANCE
+
     config = {
         "target_device": target_device.value,
         "preset": preset.value,
@@ -320,7 +326,7 @@ def _create_engine_config(
 def quantize_impl(
     model: ov.Model,
     calibration_dataset: Dataset,
-    preset: QuantizationPreset = QuantizationPreset.PERFORMANCE,
+    preset: Optional[QuantizationPreset] = None,
     target_device: TargetDevice = TargetDevice.ANY,
     subset_size: int = 300,
     fast_bias_correction: bool = True,
@@ -423,7 +429,7 @@ def quantize_with_accuracy_control_impl(
     validation_fn: Callable[[ov.CompiledModel, Iterable[Any]], float],
     max_drop: float = 0.01,
     drop_type: DropType = DropType.ABSOLUTE,
-    preset: QuantizationPreset = QuantizationPreset.PERFORMANCE,
+    preset: Optional[QuantizationPreset] = None,
     target_device: TargetDevice = TargetDevice.ANY,
     subset_size: int = 300,
     fast_bias_correction: bool = True,

nncf/quantization/algorithms/min_max/algorithm.py

@@ -96,7 +96,7 @@ class MinMaxQuantization(Algorithm):
 
     def __init__(
         self,
-        preset: QuantizationPreset = QuantizationPreset.PERFORMANCE,
+        preset: Optional[QuantizationPreset] = None,
         target_device: TargetDevice = TargetDevice.ANY,
         subset_size: int = 300,
         model_type: Optional[ModelType] = None,
@@ -157,6 +157,13 @@ def __init__(
             QuantizerGroup.ACTIVATIONS: activations_range_estimator_params,
         }
 
+        # preset definition
+        if preset is None:
+            if model_type == ModelType.TRANSFORMER:
+                preset = QuantizationPreset.MIXED
+            else:
+                preset = QuantizationPreset.PERFORMANCE
+
         # Calculates global quantizer constraints
         self._global_quantizer_constraints = {}
         for quantizer_group in QuantizerGroup:

nncf/quantization/algorithms/post_training/algorithm.py

@@ -38,7 +38,7 @@ class PostTrainingQuantization(Algorithm):
 
     def __init__(
         self,
-        preset: QuantizationPreset = QuantizationPreset.PERFORMANCE,
+        preset: Optional[QuantizationPreset] = None,
         target_device: TargetDevice = TargetDevice.ANY,
         subset_size: int = 300,
         fast_bias_correction: bool = True,

nncf/quantization/algorithms/post_training/pipeline.py

@@ -30,7 +30,7 @@
 
 
 def create_ptq_pipeline(
-    preset: QuantizationPreset = QuantizationPreset.PERFORMANCE,
+    preset: Optional[QuantizationPreset] = None,
     target_device: TargetDevice = TargetDevice.ANY,
     subset_size: int = 300,
     fast_bias_correction: bool = True,

nncf/quantization/quantize_model.py

@@ -38,7 +38,7 @@
 def quantize(
     model: TModel,
     calibration_dataset: Dataset,
-    preset: QuantizationPreset = QuantizationPreset.PERFORMANCE,
+    preset: Optional[QuantizationPreset] = None,
     target_device: TargetDevice = TargetDevice.ANY,
     subset_size: int = 300,
     fast_bias_correction: bool = True,
@@ -152,7 +152,7 @@ def quantize_with_accuracy_control(
     validation_fn: Callable[[Any, Iterable[Any]], float],
     max_drop: float = 0.01,
     drop_type: DropType = DropType.ABSOLUTE,
-    preset: QuantizationPreset = QuantizationPreset.PERFORMANCE,
+    preset: Optional[QuantizationPreset] = None,
     target_device: TargetDevice = TargetDevice.ANY,
     subset_size: int = 300,
     fast_bias_correction: bool = True,
@@ -288,7 +288,7 @@ def quantize_with_tune_hyperparams(
     initial_metric_results: MetricResults,
     quantized_metric_results: MetricResults,
     tuner_subset_size: int = 300,
-    preset: QuantizationPreset = QuantizationPreset.PERFORMANCE,
+    preset: Optional[QuantizationPreset] = None,
     target_device: TargetDevice = TargetDevice.ANY,
     subset_size: int = 300,
     fast_bias_correction: bool = True,

nncf/tensorflow/quantization/quantize_model.py

@@ -133,7 +133,7 @@ def _create_nncf_config(
 def quantize_impl(
     model: tf.Module,
     calibration_dataset: Dataset,
-    preset: QuantizationPreset,
+    preset: Optional[QuantizationPreset],
     target_device: TargetDevice,
     subset_size: int,
     fast_bias_correction: bool,
@@ -157,6 +157,9 @@ def quantize_impl(
     if target_device == TargetDevice.CPU_SPR:
         raise RuntimeError("target_device == CPU_SPR is not supported.")
 
+    if preset is None:
+        preset = QuantizationPreset.PERFORMANCE
+
     nncf_config = _create_nncf_config(preset, target_device, subset_size, ignored_scope, advanced_parameters)
 
     calibration_data_loader = CalibrationDataLoader(calibration_dataset)

nncf/torch/quantization/quantize_model.py

@@ -81,7 +81,7 @@ def _get_length(iterable) -> int:
         return length
 
 
-def _get_transformer_quantization_config(subset_size: int) -> Dict[str, Any]:
+def _get_transformer_quantization_config(preset: QuantizationPreset, subset_size: int) -> Dict[str, Any]:
     """
     Returns the quantization config for transformer-based models.
 
@@ -91,7 +91,7 @@ def _get_transformer_quantization_config(subset_size: int) -> Dict[str, Any]:
     """
     return {
         "algorithm": "quantization",
-        "preset": "mixed",
+        "preset": preset.value,
         "initializer": {
             "range": {"num_init_samples": subset_size, "type": DEFAULT_RANGE_TYPE},
             "batchnorm_adaptation": {"num_bn_adaptation_samples": 0},
@@ -133,7 +133,7 @@ def _get_default_quantization_config(preset: QuantizationPreset, subset_size: in
 
 
 def _create_nncf_config(
-    preset: QuantizationPreset,
+    preset: Optional[QuantizationPreset],
     target_device: TargetDevice,
     subset_size: int,
     model_type: Optional[ModelType],
@@ -161,10 +161,16 @@ def _create_nncf_config(
         fine-tuning the quantization algorithm.
     :return: NNCFConfig for the quantization algorithm.
     """
-    if model_type is None:
+    if preset is None:
+        if model_type == ModelType.TRANSFORMER:
+            preset = QuantizationPreset.MIXED
+        else:
+            preset = QuantizationPreset.PERFORMANCE
+
+    if model_type == ModelType.TRANSFORMER:
+        compression_config = _get_transformer_quantization_config(preset, subset_size)
+    else:
         compression_config = _get_default_quantization_config(preset, subset_size)
-    elif model_type == ModelType.TRANSFORMER:
-        compression_config = _get_transformer_quantization_config(subset_size)
 
     if ignored_scope is not None:
         _ignored_scope = convert_ignored_scope_to_list(ignored_scope)
@@ -186,7 +192,7 @@ def _create_nncf_config(
 def quantize_impl(
     model: torch.nn.Module,
     calibration_dataset: Dataset,
-    preset: QuantizationPreset,
+    preset: Optional[QuantizationPreset],
     target_device: TargetDevice,
     subset_size: int,
     fast_bias_correction: bool,