Add Pipeline class

nncf/experimental/torch/quantization/quantize_model.py

@@ -18,7 +18,7 @@
 from nncf.parameters import ModelType
 from nncf.parameters import TargetDevice
 from nncf.quantization.advanced_parameters import AdvancedQuantizationParameters
-from nncf.quantization.algorithms.post_training.algorithm import PostTrainingQuantization
+from nncf.quantization.pipelines.post_training.pipeline import create_ptq_pipeline
 from nncf.scopes import IgnoredScope
 from nncf.torch.dynamic_graph.context import no_nncf_trace
 from nncf.torch.dynamic_graph.io_handling import replicate_same_tensors
@@ -105,7 +105,7 @@ def quantize_impl(
 
     nncf_network = create_nncf_network(model.eval(), calibration_dataset)
 
-    quantization_algorithm = PostTrainingQuantization(
+    quantization_pipeline = create_ptq_pipeline(
         preset=preset,
         target_device=target_device,
         subset_size=subset_size,
@@ -115,9 +115,7 @@ def quantize_impl(
         advanced_parameters=advanced_parameters,
     )
 
-    quantized_model = quantization_algorithm.apply(
-        nncf_network, nncf_network.nncf.get_graph(), dataset=calibration_dataset
-    )
+    quantized_model = quantization_pipeline.run(nncf_network, calibration_dataset)
 
     quantized_model.nncf.disable_dynamic_graph_building()
 

nncf/onnx/quantization/quantize_model.py

@@ -16,11 +16,10 @@
 from nncf.common.logging.logger import nncf_logger
 from nncf.common.quantization.structs import QuantizationPreset
 from nncf.data import Dataset
-from nncf.onnx.graph.nncf_graph_builder import GraphConverter
 from nncf.parameters import ModelType
 from nncf.parameters import TargetDevice
 from nncf.quantization.advanced_parameters import AdvancedQuantizationParameters
-from nncf.quantization.algorithms.post_training.algorithm import PostTrainingQuantization
+from nncf.quantization.pipelines.post_training.pipeline import create_ptq_pipeline
 from nncf.quantization.telemetry_extractors import CompressionStartedWithQuantizeApi
 from nncf.scopes import IgnoredScope
 from nncf.telemetry import tracked_function
@@ -56,7 +55,7 @@ def quantize_impl(
         advanced_parameters.weights_quantization_params.per_channel = False
         advanced_parameters.activations_quantization_params.per_channel = False
 
-    quantization_algorithm = PostTrainingQuantization(
+    quantization_pipeline = create_ptq_pipeline(
         preset=preset,
         target_device=target_device,
         subset_size=subset_size,
@@ -66,7 +65,6 @@ def quantize_impl(
         advanced_parameters=advanced_parameters,
     )
 
-    graph = GraphConverter.create_nncf_graph(model)
-    quantized_model = quantization_algorithm.apply(model, graph, dataset=calibration_dataset)
+    quantized_model = quantization_pipeline.run(model, calibration_dataset)
 
     return quantized_model

nncf/openvino/quantization/quantize_model.py

@@ -19,7 +19,6 @@
 from nncf.common.logging import nncf_logger
 from nncf.common.quantization.structs import QuantizationPreset
 from nncf.data import Dataset
-from nncf.openvino.graph.nncf_graph_builder import GraphConverter
 from nncf.openvino.quantization.backend_parameters import BackendParameters
 from nncf.openvino.quantization.backend_parameters import is_weight_compression_needed
 from nncf.openvino.quantization.weights_compression import insert_pre_compression_operations
@@ -32,7 +31,7 @@
 from nncf.quantization.algorithms.accuracy_control.algorithm import QuantizationAccuracyRestorer
 from nncf.quantization.algorithms.accuracy_control.algorithm import calculate_accuracy_drop
 from nncf.quantization.algorithms.accuracy_control.evaluator import Evaluator
-from nncf.quantization.algorithms.post_training.algorithm import PostTrainingQuantization
+from nncf.quantization.pipelines.post_training.pipeline import create_ptq_pipeline
 from nncf.quantization.quantize_model import quantize_with_tune_hyperparams
 from nncf.quantization.telemetry_extractors import CompressionStartedWithQuantizeApi
 from nncf.scopes import IgnoredScope
@@ -106,7 +105,7 @@ def native_quantize_impl(
     """
     Implementation of the `quantize()` method for the OpenVINO backend via the OpenVINO Runtime API.
     """
-    quantization_algorithm = PostTrainingQuantization(
+    quantization_algorithm = create_ptq_pipeline(
         preset=preset,
         target_device=target_device,
         subset_size=subset_size,
@@ -116,8 +115,7 @@ def native_quantize_impl(
         advanced_parameters=advanced_parameters,
     )
 
-    graph = GraphConverter.create_nncf_graph(model)
-    quantized_model = quantization_algorithm.apply(model, graph, dataset=calibration_dataset)
+    quantized_model = quantization_algorithm.run(model, calibration_dataset)
 
     if is_weight_compression_needed(advanced_parameters):
         compress_quantize_weights_transformation(quantized_model)

nncf/quantization/algorithms/channel_alignment/algorithm.py

@@ -22,6 +22,7 @@
 from nncf.common.graph.transformations.commands import TargetPoint
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.common.graph.transformations.layout import TransformationLayout
+from nncf.common.logging import nncf_logger
 from nncf.common.logging.track_progress import track
 from nncf.common.tensor_statistics.statistic_point import StatisticPoint
 from nncf.common.tensor_statistics.statistic_point import StatisticPointsContainer
@@ -97,6 +98,12 @@ def apply(
         statistic_points: Optional[StatisticPointsContainer] = None,
         dataset: Optional[Dataset] = None,
     ) -> TModel:
+        if model is not None:
+            backend = get_backend(model)
+            if backend != BackendType.OPENVINO:
+                nncf_logger.debug(f"{backend.name} does not support ChannelAlignment algorithm yet.")
+                return model
+
         self._set_backend_entity(model)
         model_transformer = ModelTransformerFactory.create(model)
         transformation_layout = TransformationLayout()
@@ -368,9 +375,15 @@ def _get_target_point_and_node_in(self, conv_in, add_in) -> Tuple[TargetPoint, N
         )
 
     def get_statistic_points(self, model: TModel, graph: NNCFGraph) -> StatisticPointsContainer:
-        self._set_backend_entity(model)
-
         statistic_container = StatisticPointsContainer()
+
+        if model is not None:
+            backend = get_backend(model)
+            if backend != BackendType.OPENVINO:
+                nncf_logger.debug(f"{backend.name} does not support ChannelAlignment algorithm yet.")
+                return statistic_container
+
+        self._set_backend_entity(model)
         for conv_in, add_in, _ in self._get_node_pairs(graph):
             target_point, node_in = self._get_target_point_and_node_in(conv_in, add_in)
             channel_axis = conv_in.metatype.output_channel_axis

nncf/quantization/algorithms/smooth_quant/algorithm.py

@@ -101,6 +101,11 @@ def apply(
         statistic_points: Optional[StatisticPointsContainer] = None,
         dataset: Optional[Dataset] = None,
     ) -> TModel:
+        backend = get_backend(model)
+        if backend != BackendType.OPENVINO:
+            nncf_logger.debug(f"{backend.name} does not support SmoothQuant algorithm yet.")
+            return model
+
         self._set_backend_entity(model)
 
         nodes_to_smooth_data = self._get_nodes_to_smooth_data(graph)
@@ -221,6 +226,11 @@ def filter_func(point: StatisticPoint) -> bool:
     def get_statistic_points(self, model: TModel, graph: NNCFGraph) -> StatisticPointsContainer:
         statistic_container = StatisticPointsContainer()
 
+        backend = get_backend(model)
+        if backend != BackendType.OPENVINO:
+            nncf_logger.debug(f"{backend.name} does not support SmoothQuant algorithm yet.")
+            return statistic_container
+
         self._set_backend_entity(model)
 
         nodes_to_smooth_data = self._get_nodes_to_smooth_data(graph)