[NNCF] (#3249) Remove backend-specific methods from common layer attributes (#3287)

shumaari · web-flow · commit 7996e7358da4 · 2025-02-25T15:10:44.000+02:00
### Changes 1. Moved and renamed backend-specific class methods from common layer attributes in nncf/common/graph/layer_attributes.py to self-contained functions in nncf/common/graph/utils.py - layer_attributes.get_bias_shape -> get_bias_shape_legacy(layer_attributes: WeightedLayerAttributes) - layer_attributes.get_num_filters -> get_num_filters_legacy(layer_attributes: WeightedLayerAttributes) - layer_attributes.get_target_dim_for_compression -> get_target_dim_for_compression_legacy(layer_attributes: WeightedLayerAttributes) - layer_attributes.get_weight_shape -> get_weight_shape_legacy(layer_attributes: WeightedLayerAttributes) 2. Calls to above class methods were replaced with calls to their corresponding legacy functions in the following locations - /nncf/experimental/common/ folder - pruning/nodes_grouping.py - /nncf/experimental/tensorflow/ folder - quantization/algorithm.py - /nncf/experimental/torch/ folder - nas/bootstrapNAS/elasticity/elastic_width.py - sparsity/movement/layers.py - sparsity/movement/structured_mask_handler.py - /nncf/torch/ folder - quantization/algo.py - quantization/init_range.py - sparsity/const/algo.py - sparsity/magnitude/algo.py - sparsity/rb/algo.py ### Reason for changes Torch and Tensorflow backend-specific methods need to be removed from common layer attributes and all related calls need to be replaced by their corresponding legacy function calls (resolves #3249)
diff --git a/nncf/common/graph/layer_attributes.py b/nncf/common/graph/layer_attributes.py
@@ -10,7 +10,6 @@
 # limitations under the License.
 
 from abc import ABC
-from abc import abstractmethod
 from dataclasses import dataclass
 from enum import Enum
 from typing import Any, List, Optional, Tuple, Union
@@ -66,18 +65,6 @@ def __init__(self, weight_requires_grad: bool, dtype: Dtype = Dtype.FLOAT, with_
         self.dtype = dtype
         self.with_bias = with_bias
 
-    @abstractmethod
-    def get_weight_shape(self) -> List[int]:
-        pass
-
-    def get_num_filters(self) -> int:
-        weight_shape = self.get_weight_shape()
-        return weight_shape[self.get_target_dim_for_compression()]
-
-    @abstractmethod
-    def get_target_dim_for_compression(self) -> int:
-        pass
-
 
 class GenericWeightedLayerAttributes(WeightedLayerAttributes):
     """
@@ -103,12 +90,6 @@ def __init__(
         self.weight_shape = weight_shape
         self.filter_dimension_idx = filter_dimension_idx
 
-    def get_weight_shape(self) -> List[int]:
-        return self.weight_shape
-
-    def get_target_dim_for_compression(self) -> int:
-        return 0
-
 
 class LinearLayerAttributes(WeightedLayerAttributes):
     def __init__(
@@ -129,15 +110,6 @@ def __init__(
         self.in_features = in_features
         self.out_features = out_features
 
-    def get_weight_shape(self) -> List[int]:
-        return [self.out_features, self.in_features]
-
-    def get_bias_shape(self) -> int:
-        return self.out_features if self.with_bias is True else 0
-
-    def get_target_dim_for_compression(self) -> int:
-        return 0
-
 
 class ConvolutionLayerAttributes(WeightedLayerAttributes):
     def __init__(
@@ -179,17 +151,6 @@ def __init__(
         self.padding_values = padding_values
         self.output_padding_values = output_padding_values
 
-    def get_weight_shape(self) -> List[int]:
-        if not self.transpose:
-            return [self.out_channels, self.in_channels // self.groups, *self.kernel_size]
-        return [self.in_channels, self.out_channels // self.groups, *self.kernel_size]
-
-    def get_target_dim_for_compression(self) -> int:
-        # Always quantize per each "out" channel
-        if self.transpose:
-            return 1
-        return 0
-
 
 class GroupNormLayerAttributes(WeightedLayerAttributes):
     def __init__(self, weight_requires_grad: bool, num_channels: int, num_groups: int):
@@ -204,12 +165,6 @@ def __init__(self, weight_requires_grad: bool, num_channels: int, num_groups: in
         self.num_channels = num_channels
         self.num_groups = num_groups
 
-    def get_weight_shape(self) -> List[int]:
-        return [self.num_channels]
-
-    def get_target_dim_for_compression(self) -> int:
-        return 0
-
 
 @dataclass
 class ReshapeLayerAttributes(BaseLayerAttributes):
diff --git a/nncf/common/graph/utils.py b/nncf/common/graph/utils.py
@@ -14,6 +14,11 @@
 
 from nncf.common.graph import NNCFGraph
 from nncf.common.graph import NNCFNode
+from nncf.common.graph.layer_attributes import ConvolutionLayerAttributes
+from nncf.common.graph.layer_attributes import GenericWeightedLayerAttributes
+from nncf.common.graph.layer_attributes import GroupNormLayerAttributes
+from nncf.common.graph.layer_attributes import LinearLayerAttributes
+from nncf.common.graph.layer_attributes import WeightedLayerAttributes
 from nncf.common.graph.operator_metatypes import OperatorMetatype
 from nncf.common.logging import nncf_logger
 from nncf.common.pruning.utils import traverse_function
@@ -132,3 +137,78 @@ def get_reduction_axes(
     for channel_axis in sorted(channel_axes, reverse=True):
         del reduction_axes[channel_axis]
     return tuple(reduction_axes)
+
+
+def get_weight_shape_legacy(layer_attributes: WeightedLayerAttributes) -> List[int]:
+    """
+    Returns hard-coded weights shape layout for the given layer attributes.
+    Applicable only for eager PyTorch and Tensorflow models.
+
+    :param layer_attributes: Layer attributes of a NNCFNode.
+    :return: Weights shape layout.
+    """
+    if isinstance(layer_attributes, LinearLayerAttributes):
+        return [layer_attributes.out_features, layer_attributes.in_features]
+
+    if isinstance(layer_attributes, ConvolutionLayerAttributes):
+        if not layer_attributes.transpose:
+            return [
+                layer_attributes.out_channels,
+                layer_attributes.in_channels // layer_attributes.groups,
+                *layer_attributes.kernel_size,
+            ]
+        return [
+            layer_attributes.in_channels,
+            layer_attributes.out_channels // layer_attributes.groups,
+            *layer_attributes.kernel_size,
+        ]
+
+    if isinstance(layer_attributes, GroupNormLayerAttributes):
+        return [layer_attributes.num_channels]
+
+    assert isinstance(layer_attributes, GenericWeightedLayerAttributes)
+    return layer_attributes.weight_shape
+
+
+def get_target_dim_for_compression_legacy(layer_attributes: WeightedLayerAttributes) -> int:
+    """
+    Returns hard-coded target dim for compression for the given layer attributes.
+    Applicable only for eager PyTorch and Tensorflow models.
+
+    :param layer_attributes: Layer attributes of a NNCFNode.
+    :return: Target dim for compression.
+    """
+    if isinstance(layer_attributes, ConvolutionLayerAttributes):
+        # Always quantize per each "out" channel
+        return 1 if layer_attributes.transpose else 0
+
+    else:
+        assert isinstance(
+            layer_attributes, (GenericWeightedLayerAttributes, LinearLayerAttributes, GroupNormLayerAttributes)
+        )
+        return 0
+
+
+def get_bias_shape_legacy(layer_attributes: WeightedLayerAttributes) -> int:
+    """
+    Returns hard-coded bias shape for the given linear layer attributes.
+    Applicable only for eager PyTorch and Tensorflow models.
+
+    :param layer_attributes: Linear layer attributes of a NNCFNode.
+    :return: Correspondent bias shape.
+    """
+    assert isinstance(layer_attributes, LinearLayerAttributes)
+    return layer_attributes.out_features if layer_attributes.with_bias is True else 0
+
+
+def get_num_filters_legacy(layer_attributes: WeightedLayerAttributes) -> int:
+    """
+    Returns hard-coded number of filters for the given layer attribues.
+    Applicable only for eager PyTorch and Tensorflow models.
+
+    :param layer_attributes: Layer attributes of a NNCFNode.
+    :return: Correspondent number of filters.
+    """
+    assert isinstance(layer_attributes, WeightedLayerAttributes)
+    weight_shape = get_weight_shape_legacy(layer_attributes)
+    return weight_shape[get_target_dim_for_compression_legacy(layer_attributes)]
diff --git a/nncf/experimental/common/pruning/nodes_grouping.py b/nncf/experimental/common/pruning/nodes_grouping.py
@@ -18,6 +18,7 @@
 from nncf.common.graph.graph import NNCFNode
 from nncf.common.graph.layer_attributes import ConvolutionLayerAttributes
 from nncf.common.graph.layer_attributes import LinearLayerAttributes
+from nncf.common.graph.utils import get_target_dim_for_compression_legacy
 from nncf.common.pruning.mask_propagation import MaskPropagationAlgorithm
 from nncf.common.pruning.utils import PruningOperationsMetatypeRegistry
 from nncf.experimental.common.graph.netron import save_for_netron
@@ -76,7 +77,7 @@ def get_pruning_groups(
     roots = {}
     for node in all_nodes_to_prune:
         assert isinstance(node.layer_attributes, (LinearLayerAttributes, ConvolutionLayerAttributes))
-        pruning_dim = node.layer_attributes.get_target_dim_for_compression()
+        pruning_dim = get_target_dim_for_compression_legacy(node.layer_attributes)
         output_tensors_shapes = [x.tensor_shape for x in graph.get_output_edges(node)]
         assert not len(set(output_tensors_shapes)) > 1, node.node_name
         output_tensors_shape = output_tensors_shapes[0]
diff --git a/nncf/experimental/tensorflow/quantization/algorithm.py b/nncf/experimental/tensorflow/quantization/algorithm.py
@@ -18,6 +18,7 @@
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.common.graph.transformations.commands import TransformationPriority
 from nncf.common.graph.utils import get_first_nodes_of_type
+from nncf.common.graph.utils import get_weight_shape_legacy
 from nncf.common.logging import nncf_logger
 from nncf.common.quantization.quantizer_setup import ActivationQuantizationInsertionPoint
 from nncf.common.quantization.quantizer_setup import QuantizationPointId
@@ -133,7 +134,7 @@ def _get_tensor_specs(
         assert len(metatype.weight_definitions) == 1
 
         channel_axes = metatype.weight_definitions[0].channel_axes
-        weight_shape = node.layer_attributes.get_weight_shape()
+        weight_shape = get_weight_shape_legacy(node.layer_attributes)
         tensor_specs.append((weight_shape, channel_axes))
     else:
         data_format = node.layer_attributes.get_data_format()
diff --git a/nncf/experimental/torch/nas/bootstrapNAS/elasticity/elastic_width.py b/nncf/experimental/torch/nas/bootstrapNAS/elasticity/elastic_width.py
@@ -27,6 +27,7 @@
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.common.graph.transformations.commands import TransformationCommand
 from nncf.common.graph.transformations.commands import TransformationPriority
+from nncf.common.graph.utils import get_num_filters_legacy
 from nncf.common.logging import nncf_logger
 from nncf.common.pruning.clusterization import Cluster
 from nncf.common.pruning.clusterization import Clusterization
@@ -1204,15 +1205,15 @@ def _create_dynamic_dw_conv_input_op(conv_layer_attrs: BaseLayerAttributes, node
     def _create_dynamic_bn_input_op(generic_layer_attrs: BaseLayerAttributes, node_name: str) -> UpdateBatchNormParams:
         assert isinstance(generic_layer_attrs, GenericWeightedLayerAttributes)
         dynamic_bn_input_op = ElasticInputWidthBatchNormOp(
-            max_width=generic_layer_attrs.get_num_filters(), node_name=node_name
+            max_width=get_num_filters_legacy(generic_layer_attrs), node_name=node_name
         )
         return UpdateBatchNormParams(dynamic_bn_input_op)
 
     @staticmethod
     def _create_dynamic_ln_input_op(generic_layer_attrs: BaseLayerAttributes, node_name: str) -> UpdateLayerNormParams:
         assert isinstance(generic_layer_attrs, GenericWeightedLayerAttributes)
         dynamic_ln_input_op = ElasticInputWidthLayerNormOp(
-            max_width=generic_layer_attrs.get_num_filters(), node_name=node_name
+            max_width=get_num_filters_legacy(generic_layer_attrs), node_name=node_name
         )
         return UpdateLayerNormParams(dynamic_ln_input_op)
 
diff --git a/nncf/experimental/torch/sparsity/movement/layers.py b/nncf/experimental/torch/sparsity/movement/layers.py
@@ -18,6 +18,8 @@
 
 import nncf
 from nncf.common.graph import NNCFNode
+from nncf.common.graph.utils import get_bias_shape_legacy
+from nncf.common.graph.utils import get_weight_shape_legacy
 from nncf.experimental.torch.sparsity.movement.functions import binary_mask_by_threshold
 from nncf.torch.layer_utils import COMPRESSION_MODULES
 from nncf.torch.layer_utils import CompressionParameter
@@ -167,7 +169,7 @@ def __init__(
         self._importance_threshold = -math.inf
         self._importance_regularization_factor = 0.0
 
-        weight_shape: List[int] = target_module_node.layer_attributes.get_weight_shape()
+        weight_shape: List[int] = get_weight_shape_legacy(target_module_node.layer_attributes)
         assert len(weight_shape) == 2, "Unsupported module with weight shape not in 2D."
         self.weight_ctx = BinaryMask(weight_shape)
         self.sparse_factors = self._get_sparse_factors(weight_shape, sparse_cfg)
@@ -185,7 +187,7 @@ def __init__(
         self.weight_ctx.binary_mask = self._calc_training_binary_mask()
 
         if self.prune_bias:
-            bias_shape = target_module_node.layer_attributes.get_bias_shape()
+            bias_shape = get_bias_shape_legacy(target_module_node.layer_attributes)
             self.bias_ctx = BinaryMask(bias_shape)
             bias_importance_shape = weight_importance_shape[0]
             self.bias_importance = CompressionParameter(
diff --git a/nncf/experimental/torch/sparsity/movement/structured_mask_handler.py b/nncf/experimental/torch/sparsity/movement/structured_mask_handler.py
@@ -18,6 +18,8 @@
 
 from nncf.common.graph.graph import NNCFNodeName
 from nncf.common.graph.layer_attributes import LinearLayerAttributes
+from nncf.common.graph.utils import get_bias_shape_legacy
+from nncf.common.graph.utils import get_weight_shape_legacy
 from nncf.common.logging import nncf_logger
 from nncf.experimental.common.pruning.nodes_grouping import get_pruning_groups
 from nncf.experimental.common.pruning.nodes_grouping import select_largest_groups
@@ -208,9 +210,9 @@ def gather_statistics_from_operand(self) -> StructuredMaskContextStatistics:
         """
         node = self.sparsifier_operand.target_module_node
         assert isinstance(node.layer_attributes, tuple(EXPECTED_NODE_LAYER_ATTRS))
-        weight_shape: Tuple[int, int] = tuple(node.layer_attributes.get_weight_shape())
+        weight_shape: Tuple[int, int] = tuple(get_weight_shape_legacy(node.layer_attributes))
         bias_shape: Tuple[int] = (
-            (node.layer_attributes.get_bias_shape(),) if self.sparsifier_operand.prune_bias else (0,)
+            (get_bias_shape_legacy(node.layer_attributes),) if self.sparsifier_operand.prune_bias else (0,)
         )
 
         pruned_weight_shape = list(weight_shape)
diff --git a/nncf/torch/quantization/algo.py b/nncf/torch/quantization/algo.py
@@ -35,6 +35,8 @@
 from nncf.common.graph.patterns.manager import TargetDevice
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.common.graph.utils import get_first_nodes_of_type
+from nncf.common.graph.utils import get_target_dim_for_compression_legacy
+from nncf.common.graph.utils import get_weight_shape_legacy
 from nncf.common.hardware.config import HWConfig
 from nncf.common.hardware.config import HWConfigType
 from nncf.common.hardware.config import get_hw_config_type
@@ -603,8 +605,8 @@ def _get_minmax_values_for_quantizer_locations(
                 if qp.is_weight_quantization_point():
                     layer_attrs = target_node.layer_attributes
                     assert isinstance(layer_attrs, WeightedLayerAttributes)
-                    input_shape = layer_attrs.get_weight_shape()
-                    channel_idx = layer_attrs.get_target_dim_for_compression()
+                    input_shape = get_weight_shape_legacy(layer_attrs)
+                    channel_idx = get_target_dim_for_compression_legacy(layer_attrs)
                 else:
                     input_shape = target_model_graph.get_input_shape_for_insertion_point(qp.insertion_point)
                     channel_idx = 1  # channel dim for activations
@@ -773,10 +775,10 @@ def _get_quantizer_setup(self, target_model: NNCFNetwork) -> PTQuantizerSetup:
                 layer_attributes = target_node.layer_attributes
                 assert isinstance(layer_attributes, WeightedLayerAttributes)
                 scale_shape = get_scale_shape(
-                    layer_attributes.get_weight_shape(),
+                    get_weight_shape_legacy(layer_attributes),
                     is_weights=True,
                     per_channel=qconfig.per_channel,
-                    channel_idx=layer_attributes.get_target_dim_for_compression(),
+                    channel_idx=get_target_dim_for_compression_legacy(layer_attributes),
                 )
             else:
                 input_shape = target_model_graph.get_input_shape_for_insertion_point(insertion_point)
@@ -1181,7 +1183,7 @@ def is_weights(ip: PTTargetPoint) -> bool:
             )
             module_node = target_model_graph.get_node_by_name(primary_ip.target_node_name)
             layer_attributes = module_node.layer_attributes
-            input_shape = layer_attributes.get_weight_shape()
+            input_shape = get_weight_shape_legacy(layer_attributes)
             self._quantizers_input_shapes[primary_qid] = tuple(input_shape)
         else:
             primary_qid = NonWeightQuantizerId(primary_ip.target_node_name, primary_ip.input_port_id)
diff --git a/nncf/torch/quantization/init_range.py b/nncf/torch/quantization/init_range.py
@@ -18,6 +18,8 @@
 
 import nncf
 from nncf.common.graph.layer_attributes import WeightedLayerAttributes
+from nncf.common.graph.utils import get_target_dim_for_compression_legacy
+from nncf.common.graph.utils import get_weight_shape_legacy
 from nncf.common.quantization.initialization.range import RangeInitCollectorParams
 from nncf.common.quantization.initialization.range import RangeInitConfig
 from nncf.common.quantization.initialization.range import RangeInitParams
@@ -226,8 +228,8 @@ def get_all_scale_shapes_with_params(
             module_node = target_nncf_graph.get_node_by_name(qp.insertion_point.target_node_name)
             layer_attributes = module_node.layer_attributes
             assert isinstance(layer_attributes, WeightedLayerAttributes)
-            input_shape = layer_attributes.get_weight_shape()
-            channel_idx = layer_attributes.get_target_dim_for_compression()
+            input_shape = get_weight_shape_legacy(layer_attributes)
+            channel_idx = get_target_dim_for_compression_legacy(layer_attributes)
         else:
             input_shape = target_nncf_graph.get_input_shape_for_insertion_point(qp.insertion_point)
             channel_idx = 1  # channel dim for activations
diff --git a/nncf/torch/sparsity/const/algo.py b/nncf/torch/sparsity/const/algo.py
@@ -11,6 +11,7 @@
 from typing import Tuple
 
 from nncf.common.graph import NNCFNode
+from nncf.common.graph.utils import get_weight_shape_legacy
 from nncf.common.sparsity.statistics import ConstSparsityStatistics
 from nncf.common.statistics import NNCFStatistics
 from nncf.common.utils.api_marker import api
@@ -26,7 +27,7 @@
 @PT_COMPRESSION_ALGORITHMS.register("const_sparsity")
 class ConstSparsityBuilder(BaseSparsityAlgoBuilder):
     def create_weight_sparsifying_operation(self, target_module_node: NNCFNode, compression_lr_multiplier: float):
-        return BinaryMask(target_module_node.layer_attributes.get_weight_shape())
+        return BinaryMask(get_weight_shape_legacy(target_module_node.layer_attributes))
 
     def _build_controller(self, model: NNCFNetwork) -> PTCompressionAlgorithmController:
         return ConstSparsityController(model, self._sparsified_module_info)
diff --git a/nncf/torch/sparsity/magnitude/algo.py b/nncf/torch/sparsity/magnitude/algo.py
@@ -17,6 +17,7 @@
 from nncf.api.compression import CompressionStage
 from nncf.common.accuracy_aware_training.training_loop import ADAPTIVE_COMPRESSION_CONTROLLERS
 from nncf.common.graph import NNCFNode
+from nncf.common.graph.utils import get_weight_shape_legacy
 from nncf.common.initialization.batchnorm_adaptation import BatchnormAdaptationAlgorithm
 from nncf.common.schedulers import StubCompressionScheduler
 from nncf.common.sparsity.schedulers import SPARSITY_SCHEDULERS
@@ -44,7 +45,7 @@
 @PT_COMPRESSION_ALGORITHMS.register("magnitude_sparsity")
 class MagnitudeSparsityBuilder(BaseSparsityAlgoBuilder):
     def create_weight_sparsifying_operation(self, target_module_node: NNCFNode, compression_lr_multiplier: float):
-        return BinaryMask(target_module_node.layer_attributes.get_weight_shape())
+        return BinaryMask(get_weight_shape_legacy(target_module_node.layer_attributes))
 
     def _build_controller(self, model: NNCFNetwork) -> PTCompressionAlgorithmController:
         return MagnitudeSparsityController(model, self._sparsified_module_info, self.config)
diff --git a/nncf/torch/sparsity/rb/algo.py b/nncf/torch/sparsity/rb/algo.py
@@ -18,6 +18,7 @@
 from nncf.api.compression import CompressionStage
 from nncf.common.accuracy_aware_training.training_loop import ADAPTIVE_COMPRESSION_CONTROLLERS
 from nncf.common.graph import NNCFNode
+from nncf.common.graph.utils import get_weight_shape_legacy
 from nncf.common.schedulers import StubCompressionScheduler
 from nncf.common.sparsity.schedulers import SPARSITY_SCHEDULERS
 from nncf.common.sparsity.statistics import RBSparsityStatistics
@@ -44,7 +45,7 @@
 class RBSparsityBuilder(BaseSparsityAlgoBuilder):
     def create_weight_sparsifying_operation(self, target_module_node: NNCFNode, compression_lr_multiplier: float):
         return RBSparsifyingWeight(
-            target_module_node.layer_attributes.get_weight_shape(),
+            get_weight_shape_legacy(target_module_node.layer_attributes),
             frozen=False,
             compression_lr_multiplier=compression_lr_multiplier,
         )
