[PTQ][OV] ChannelAlignment algorithm is ready to be enabled by default (

openvinotoolkit#2073)

### Changes

* ChannelAlignment algorithm is enabled by default
* Biases are added only for operations that are affected by CA algorithm

### Reason for changes

* To increase models mertics by using ChannelAlignment algorithm by
default

### Related tickets

114328
114583

### Tests

tests/post_training/test_templates/test_channel_alignment.py is updated

nncf/common/graph/transformations/command_creation.py

@@ -47,6 +47,17 @@ def create_command_to_update_bias(
         :return: The command to update bias value.
         """
 
+    @staticmethod
+    @abstractmethod
+    def create_command_to_insert_bias(node_without_bias: NNCFNode, bias_value: Any) -> TransformationCommand:
+        """
+        Creates command to insert bias after given node.
+
+        :param node_without_bias: The node that corresponds to the operation without bias.
+        :param bias_value: Bias value to insert.
+        :return: The command to insert bias value.
+        """
+
     @staticmethod
     @abstractmethod
     def create_command_to_update_weight(

nncf/quantization/algorithms/channel_alignment/algorithm.py

@@ -59,20 +59,17 @@ def __init__(
         self,
         subset_size: int = 100,
         inplace_statistics: bool = True,
-        backend_params: Optional[Dict[str, Any]] = None,
     ):
         """
         :param subset_size: Size of a subset for the statistics collection,
             defaults to 100.
         :param inplace_statistics: Defines wheather to calculate quantizers statistics
             by backend graph operations or by default Python implementation, defaults
             to True.
-        :param backend_params: Backend specific parameters.
         """
         super().__init__()
         self.subset_size = subset_size
         self.inplace_statistics = inplace_statistics
-        self.backend_params = backend_params
         self._backend_entity = None
         self._quantile = 1e-4
         self._algorithm_key = f"CA_{hash(self)}"
@@ -120,15 +117,14 @@ def filter_func(point: StatisticPoint) -> bool:
             conv_in_cont = ConvParamsContainer(conv_in, model, graph, self._backend_entity)
             conv_out_cont = ConvParamsContainer(conv_out, model, graph, self._backend_entity)
 
-            if conv_in_cont.has_bias() and conv_out_cont.has_bias():
-                amean = (stat.max_values + stat.min_values) * 0.5
-                conv_in_cont.bias, conv_out_cont.bias = self._align_means(
-                    conv_in_cont.bias,
-                    conv_out_cont.bias,
-                    conv_out_cont.weight,
-                    amean,
-                    conv_out_cont.dims,
-                )
+            amean = (stat.max_values + stat.min_values) * 0.5
+            conv_in_cont.bias, conv_out_cont.bias = self._align_means(
+                conv_in_cont.bias,
+                conv_out_cont.bias,
+                conv_out_cont.weight,
+                amean,
+                conv_out_cont.dims,
+            )
 
             ascale = (stat.max_values - stat.min_values).astype(np.float32)
             eps = np.finfo(ascale.dtype).eps
@@ -153,9 +149,11 @@ def filter_func(point: StatisticPoint) -> bool:
                     )
 
                 if container.stated_bias.is_modified():
-                    transformation_layout.register(
-                        command_creator.create_command_to_update_bias(container.op, container.bias, graph),
-                    )
+                    if container.bias_op_exist():
+                        command = command_creator.create_command_to_update_bias(container.op, container.bias, graph)
+                    else:
+                        command = command_creator.create_command_to_insert_bias(container.op, container.bias)
+                    transformation_layout.register(command)
 
         transformed_model = model_transformer.transform(transformation_layout)
         return transformed_model
@@ -239,10 +237,7 @@ def _align_scales(
         scale_in_shape[conv_in_descr.conv_weight_out_channels_dim] = scale_factor.shape[conv_in_descr.bias_channels_dim]
         updated_conv_in_value = conv_in_value / scale_factor.reshape(scale_in_shape)
 
-        if bias_in_value is not None:
-            updated_bias_in_value = bias_in_value / scale_factor.reshape(bias_in_value.shape)
-        else:
-            updated_bias_in_value = None
+        updated_bias_in_value = bias_in_value / scale_factor.reshape(bias_in_value.shape)
 
         scale_out_shape = np.ones(len(conv_out_value.shape), dtype=int)
         scale_out_shape[conv_out_descr.conv_weight_in_channels_dim] = scale_factor.shape[
@@ -433,18 +428,23 @@ class ConvParamsContainer:
     Convolution container class which is incapsulating common convolutional parameters collection.
     """
 
-    def __init__(self, conv_op, model, nncf_graph, backend_entity: ChannelAlignmentAlgoBackend):
+    def __init__(
+        self, conv_op: NNCFNode, model: TModel, nncf_graph: NNCFGraph, backend_entity: ChannelAlignmentAlgoBackend
+    ):
         """
-        :param conv_op: Backend-specific conv node.
+        :param conv_op: NNCF conv node.
         :param model: Backend-specific model instance.
         :param nncf_graph: NNCFGraph of given backend-specific model.
         :param backend_entity: Current backend entity to retrieve parameters from given conv node
         """
         _, self._weights_port_id = backend_entity.get_weights_port_ids_for_node(conv_op)
         self.stated_weight = StatedTensor(backend_entity.get_weight_value(conv_op, model, self._weights_port_id))
-        bias = None
+        self._bias_op_exist = False
         if backend_entity.is_node_with_bias(conv_op, nncf_graph):
             bias = backend_entity.get_bias_value(conv_op, model, nncf_graph)
+            self._bias_op_exist = True
+        else:
+            bias = backend_entity.create_bias_tensor(conv_op, nncf_graph, 0)
         self.stated_bias = StatedTensor(bias)
         self._op = conv_op
         self._dims = backend_entity.get_dims_descriptor(conv_op)
@@ -477,5 +477,5 @@ def weight_port_id(self):
     def dims(self) -> LayoutDescriptor:
         return self._dims
 
-    def has_bias(self) -> bool:
-        return self.bias is not None
+    def bias_op_exist(self) -> bool:
+        return self._bias_op_exist

nncf/quantization/algorithms/channel_alignment/backend.py

@@ -11,7 +11,7 @@
 
 from abc import abstractmethod
 from dataclasses import dataclass
-from typing import Optional, Tuple, TypeVar
+from typing import Any, Optional, Tuple, TypeVar
 
 import numpy as np
 
@@ -142,3 +142,15 @@ def get_conv_layer_attributes(node: NNCFNode) -> Optional[ConvolutionLayerAttrib
         :param node: NNCFNode to take convolutional layer attributes from.
         :return: Convolutional layer attributes of given node if they are present and None otherwise
         """
+
+    @staticmethod
+    @abstractmethod
+    def create_bias_tensor(node: NNCFNode, nncf_graph: NNCFGraph, value: Any) -> np.ndarray:
+        """
+        Creates bias value constant array filled by given value.
+
+        :param node: NNCFNode to add bias to.
+        :param nncf_graph: Target NNCFgraph.
+        :param value: Value to fill bias constant array.
+        :return: Bias value constant array filled by given value.
+        """

nncf/quantization/algorithms/channel_alignment/openvino_backend.py

@@ -9,7 +9,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Optional, Tuple
+from typing import Any, Optional, Tuple
 
 import numpy as np
 import openvino.runtime as ov
@@ -29,6 +29,7 @@
 from nncf.openvino.graph.metatypes.openvino_metatypes import OVGroupConvolutionMetatype
 from nncf.openvino.graph.metatypes.openvino_metatypes import OVMatMulMetatype
 from nncf.openvino.graph.metatypes.openvino_metatypes import OVSubtractMetatype
+from nncf.openvino.graph.node_utils import create_bias_tensor
 from nncf.openvino.graph.node_utils import get_bias_value
 from nncf.openvino.graph.node_utils import get_node_with_bias_value
 from nncf.openvino.graph.node_utils import get_weight_value
@@ -146,3 +147,7 @@ def get_conv_layer_attributes(node: NNCFNode) -> Optional[ConvolutionLayerAttrib
         if node.layer_attributes is None:
             return None
         return node.layer_attributes.layer_attributes[1]
+
+    @staticmethod
+    def create_bias_tensor(node: NNCFNode, nncf_graph: NNCFGraph, value: Any) -> np.ndarray:
+        return create_bias_tensor(node, nncf_graph, value)

nncf/quantization/algorithms/post_training/algorithm.py

@@ -9,7 +9,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from dataclasses import dataclass
 from typing import Callable, Dict, List, Optional, TypeVar
 
 from nncf import Dataset
@@ -34,7 +33,6 @@
 from nncf.quantization.algorithms.fast_bias_correction.algorithm import FastBiasCorrection
 from nncf.quantization.algorithms.min_max.algorithm import MinMaxQuantization
 from nncf.quantization.algorithms.smooth_quant.algorithm import SmoothQuant
-from nncf.quantization.passes import insert_null_biases_pass
 from nncf.scopes import IgnoredScope
 
 TModel = TypeVar("TModel")
@@ -49,11 +47,6 @@ class PostTrainingQuantization(Algorithm):
     3) FastBiasCorrection or BiasCorrection
     """
 
-    @dataclass
-    class FirstStageAlgorithm:
-        algorithm: "Algorithm"
-        pre_passes: List[TPass]
-
     def __init__(
         self,
         preset: QuantizationPreset = QuantizationPreset.PERFORMANCE,
@@ -87,7 +80,7 @@ def __init__(
         """
         super().__init__()
         self.algorithms = []
-        self.first_stage_algorithms: List[self.FirstStageAlgorithm] = []
+        self.first_stage_algorithms: List[Algorithm] = []
 
         if target_device is TargetDevice.VPU:
             warning_deprecated("VPU device is deprecated and will no longer be supported in the future.")
@@ -101,15 +94,14 @@ def __init__(
                 inplace_statistics=advanced_parameters.inplace_statistics,
                 alpha=advanced_parameters.smooth_quant_alpha,
             )
-            self.first_stage_algorithms.append(self.FirstStageAlgorithm(smooth_quant_algorithm, []))
+            self.first_stage_algorithms.append(smooth_quant_algorithm)
 
         if not advanced_parameters.disable_channel_alignment:
             channel_alignment = ChannelAlignment(
                 subset_size=subset_size,
                 inplace_statistics=advanced_parameters.inplace_statistics,
-                backend_params=advanced_parameters.backend_params,
             )
-            self.first_stage_algorithms.append(self.FirstStageAlgorithm(channel_alignment, [insert_null_biases_pass]))
+            self.first_stage_algorithms.append(channel_alignment)
 
         min_max_quantization = MinMaxQuantization(
             preset=preset,
@@ -187,9 +179,7 @@ def apply(
         modified_model_graph = graph
         backend = get_backend(modified_model)
 
-        for first_stage_algorithm in self.first_stage_algorithms:
-            algorithm = first_stage_algorithm.algorithm
-
+        for algorithm in self.first_stage_algorithms:
             if isinstance(algorithm, SmoothQuant) and backend != BackendType.OPENVINO:
                 nncf_logger.debug(f"{backend.name} does not support SmoothQuant algorithm yet.")
                 continue
@@ -198,10 +188,6 @@ def apply(
                 nncf_logger.debug(f"{backend.name} does not support ChannelAlignment algorithm yet.")
                 continue
 
-            for pre_pass in first_stage_algorithm.pre_passes:
-                modified_model = pre_pass(modified_model, modified_model_graph)
-                modified_model_graph = NNCFGraphFactory.create(modified_model)
-
             statistics_aggregator = StatisticsAggregatorFactory.create(modified_model, dataset)
             algo_statistic_points = algorithm.get_statistic_points(modified_model, modified_model_graph)
             statistics_aggregator.register_statistic_points(algo_statistic_points)

tests/openvino/native/quantization/test_channel_alignment.py

@@ -23,6 +23,7 @@
 from nncf.openvino.graph.metatypes.openvino_metatypes import OVMatMulMetatype
 from nncf.openvino.graph.transformations.command_creation import OVCommandCreator
 from nncf.openvino.graph.transformations.commands import OVBiasCorrectionCommand
+from nncf.openvino.graph.transformations.commands import OVBiasInsertionCommand
 from nncf.openvino.graph.transformations.commands import OVTargetPoint
 from nncf.openvino.graph.transformations.commands import OVWeightUpdateCommand
 from nncf.quantization.algorithms.channel_alignment.backend import LayoutDescriptor
@@ -64,7 +65,7 @@ def get_constant_metatype(self):
         return OVConstantMetatype
 
     def get_transformation_commands(self):
-        return OVBiasCorrectionCommand, OVWeightUpdateCommand
+        return OVBiasInsertionCommand, OVBiasCorrectionCommand, OVWeightUpdateCommand
 
     def mock_command_creation_factory(self, mocker) -> None:
         mocker.patch("nncf.common.factory.CommandCreatorFactory.create", return_value=OVCommandCreator)

tests/openvino/native/test_model_utils.py

@@ -42,7 +42,7 @@ def get_nncf_graph_for_test(edge_shape, dtype):
     "edge_shape,dtype,ref_shape",
     [((2, 3, 4, 5), np.float32, (1, 3, 1, 1)), ((1, 1, 2, 3), np.float64, (1, 1, 1, 1))],
 )
-def test_create_bias_constant_value(edge_shape, dtype, ref_shape):
+def test_create_bias_tensor(edge_shape, dtype, ref_shape):
     graph = get_nncf_graph_for_test(edge_shape, dtype)
     val = create_bias_tensor(graph.get_node_by_name("/Conv_1_0"), graph, 5)
     assert val.shape == ref_shape

tests/post_training/test_templates/test_channel_alignment.py

@@ -183,18 +183,15 @@ def test_align_means(self, conv_out_value, refs, transposed):
     REF_UPDATED_CONV_OUT = np.array([[0.0, 2.0, 0.04, 600, 8], [10, 12, 0.14, 1600, 18]])
     REF_UPDATED_BIAS_IN = np.array([2, 4, 600, 0.08, 10])
 
-    @pytest.mark.parametrize("bias_in_value", [np.array([2, 4, 6, 8, 10]), None])
+    @pytest.mark.parametrize("bias_in_value", [np.array([2, 4, 6, 8, 10])])
     def test_align_scales(self, bias_in_value):
         def check_updated_values(updated_conv_in, updated_conv_out, updated_bias_in):
             assert updated_conv_in.shape == self.REF_UPDATED_CONV_IN.shape
             assert np.allclose(updated_conv_in, self.REF_UPDATED_CONV_IN)
             assert updated_conv_out.shape == self.REF_UPDATED_CONV_OUT.shape
             assert np.allclose(updated_conv_out, self.REF_UPDATED_CONV_OUT)
-            if bias_in_value is None:
-                assert updated_bias_in is None
-            else:
-                assert updated_bias_in.shape == self.REF_UPDATED_BIAS_IN.shape
-                assert np.allclose(updated_bias_in, self.REF_UPDATED_BIAS_IN)
+            assert updated_bias_in.shape == self.REF_UPDATED_BIAS_IN.shape
+            assert np.allclose(updated_bias_in, self.REF_UPDATED_BIAS_IN)
 
         conv_in_value = np.arange(5).reshape(5, 1)
         conv_out_value = np.arange(10).reshape(2, 5) * 2
@@ -225,7 +222,6 @@ def check_updated_values(updated_conv_in, updated_conv_out, updated_bias_in):
         updated_conv_in = updated_conv_in.reshape(updated_conv_in.shape[1:])
         check_updated_values(updated_conv_in, updated_conv_out, updated_bias_in)
 
-    GET_NODES_TEST_CASES = []
     GET_NODES_TEST_CASES = [(VALID_CONV_LAYER_ATTR, VALID_CONV_LAYER_ATTR, True)]
     GET_NODES_TEST_CASES.extend([(attr, VALID_CONV_LAYER_ATTR, True) for attr in INVALID_CONSUMER_CONV_LAYER_ATTRS])
     GET_NODES_TEST_CASES.extend([(VALID_CONV_LAYER_ATTR, attr, False) for attr in INVALID_CONSUMER_CONV_LAYER_ATTRS])
@@ -370,27 +366,24 @@ class MockBackend(backend_cls):
             assert len(arg.transformations) == 0
             return
 
-        align_means_called = 1 if num_biases == 2 else 0
-        assert algorithm._align_means.call_count == align_means_called
-        if align_means_called:
-            algorithm._align_means.assert_called_once_with(
-                ref_bias_val + "1",
-                ref_bias_val + "2",
-                ref_weights_val + "2",
-                np.array(0.5, dtype=np.float32),
-                ref_dims_descr + "2",
-            )
+        assert algorithm._align_means.call_count == 1
+        args = [
+            np.zeros((1, 1, 1, 1)),
+            np.zeros((1, 1, 1, 1)),
+            ref_weights_val + "2",
+            np.array(0.5, dtype=np.float32),
+            ref_dims_descr + "2",
+        ]
+        for i in range(num_biases):
+            args[i] = f"ref_bias_val{i + 1}"
+
+        algorithm._align_means.assert_called_once_with(*args)
 
         assert algorithm._align_scales.call_count == 1
         args = algorithm._align_scales.call_args.args
         assert args[0] == ref_weights_val + "1"
         assert args[1] == ref_weights_val + "2"
-        if num_biases == 2:
-            assert args[2] == ref_bias_in_after_align
-        elif num_biases == 1:
-            assert args[2] == ref_bias_val + "1"
-        else:
-            assert args[2] is None
+        assert args[2] == ref_bias_in_after_align
         assert ((args[3] - 3) < EPS).all()
         assert args[4] == ref_dims_descr + "1"
         assert args[5] == ref_dims_descr + "2"
@@ -408,14 +401,20 @@ class MockBackend(backend_cls):
             },
             "/Conv_2_0": {"weight_value": ref_weights_out_after_scale_align, "bias_value": ref_bias_out_after_align},
         }
-        bias_update_cls, weights_update_cls = self.get_transformation_commands()
+
+        bias_insert_cls, bias_update_cls, weights_update_cls = self.get_transformation_commands()
         for transformation in transformations:
-            assert transformation.type == TransformationType.CHANGE
             tp = transformation.target_point
             if isinstance(transformation, bias_update_cls):
+                assert transformation.type == TransformationType.CHANGE
                 _class = bias_update_cls
                 _attr = "bias_value"
+            elif isinstance(transformation, bias_insert_cls):
+                assert transformation.type == TransformationType.INSERT
+                _class = bias_insert_cls
+                _attr = "bias_value"
             elif isinstance(transformation, weights_update_cls):
+                assert transformation.type == TransformationType.CHANGE
                 _class = weights_update_cls
                 _attr = "weight_value"
             else:
@@ -425,18 +424,16 @@ class MockBackend(backend_cls):
             assert ref_values[tp.target_node_name][_attr] == getattr(transformation, _attr)
 
         if num_biases == 2:
-            ref_len = {"/Conv_1_0": 2, "/Conv_2_0": 2}
+            ref_classes = {"/Conv_1_0": bias_update_cls, "/Conv_2_0": bias_update_cls}
         elif num_biases == 1:
-            ref_len = {"/Conv_1_0": 2, "/Conv_2_0": 1}
+            ref_classes = {"/Conv_1_0": bias_update_cls, "/Conv_2_0": bias_insert_cls}
         else:
-            ref_len = {"/Conv_1_0": 1, "/Conv_2_0": 1}
+            ref_classes = {"/Conv_1_0": bias_insert_cls, "/Conv_2_0": bias_insert_cls}
 
         for node_name, _transformations in target_names.items():
-            _ref_len = ref_len[node_name]
-            assert len(_transformations) == _ref_len
+            assert len(_transformations) == 2
             assert weights_update_cls in _transformations
-            if _ref_len == 2:
-                assert bias_update_cls in _transformations
+            assert ref_classes[node_name] in _transformations
 
     @pytest.mark.parametrize("num_biases", [0, 1, 2])
     def test_get_statistic_points(self, num_biases, mocker):