[ONNX] ONNX FBC/BC metatypes groups (#2599)

### Changes

Update metatypes list for FBC/BC
Fix bug with determing depthwise conv.

### Reason for changes

Red PTQ conformance

### Related tickets

N/A

### Tests

Extend test_metatypes by quantized models

nncf/onnx/graph/metatypes/groups.py

@@ -117,12 +117,14 @@
 
 
 # Contains the operation metatypes for which bias can be applied.
-OPERATIONS_WITH_BIAS = [
+OPERATIONS_WITH_BIAS_REDUCED = [
     onnx_metatypes.ONNXConvolutionMetatype,
     onnx_metatypes.ONNXGemmMetatype,
     # TODO: Need to add MatMul with the separate bias support (CVS-135433)
 ]
 
+OPERATIONS_WITH_BIAS = [*OPERATIONS_WITH_BIAS_REDUCED, onnx_metatypes.ONNXDepthwiseConvolutionMetatype]
+
 
 QUANTIZE_DEQUANTIZE_OPERATIONS = [
     onnx_metatypes.ONNXQuantizeLinearMetatype,

nncf/onnx/graph/metatypes/onnx_metatypes.py

@@ -749,7 +749,7 @@ def _is_depthwise_conv(model: onnx.ModelProto, node: onnx.NodeProto) -> bool:
         if attribute.name == "group":
             conv_group = onnx.helper.get_attribute_value(attribute)
     weight_tensor_value = None
-    initializer_name = node.input[1]
+    initializer_name = get_tensor_edge_name(model, node, 1, get_parents_node_mapping(model))
     for init in model.graph.initializer:
         if init.name == initializer_name:
             weight_tensor_value = onnx.numpy_helper.to_array(init)

nncf/quantization/algorithms/fast_bias_correction/onnx_backend.py

@@ -19,6 +19,7 @@
 from nncf.common.graph.transformations.commands import TargetType
 from nncf.experimental.common.tensor_statistics.collectors import TensorCollector
 from nncf.experimental.tensor import Tensor
+from nncf.onnx.graph.metatypes.groups import OPERATIONS_WITH_BIAS_REDUCED
 from nncf.onnx.graph.node_utils import get_bias_value
 from nncf.onnx.graph.node_utils import is_any_weight_quantized
 from nncf.onnx.graph.node_utils import is_node_with_bias
@@ -89,7 +90,7 @@ def is_quantized_weights(node: NNCFNode, nncf_graph: NNCFGraph) -> bool:
 
     @staticmethod
     def is_node_with_bias(node: NNCFNode, nncf_graph: NNCFGraph) -> bool:
-        return is_node_with_bias(node)
+        return is_node_with_bias(node) and node.metatype in OPERATIONS_WITH_BIAS_REDUCED
 
     @staticmethod
     def get_node_names_for_input_output_statistics(node: NNCFNode, nncf_graph: NNCFGraph) -> Tuple[str, str]:

tests/onnx/test_metatypes.py

@@ -20,14 +20,17 @@
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXConstantOfShapeMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXConvolutionMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXDepthwiseConvolutionMetatype
+from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXDequantizeLinearMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXGlobalAveragePoolMetatype
+from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXQuantizeLinearMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXReluMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXShapeMetatype
 from nncf.onnx.graph.nncf_graph_builder import GraphConverter
 from tests.onnx.models import LinearModel
 from tests.onnx.models import ModelWithIntEdges
 from tests.onnx.models import MultiInputOutputModel
 from tests.onnx.models import OneDepthwiseConvolutionalModel
+from tests.onnx.quantization.common import min_max_quantize_model
 
 TEST_MODELS = [LinearModel, MultiInputOutputModel, ModelWithIntEdges, OneDepthwiseConvolutionalModel]
 REF_METATYPES_COUNTERS = [
@@ -52,11 +55,45 @@
     [InputNoopMetatype, ONNXConstantOfShapeMetatype, ONNXShapeMetatype, OutputNoopMetatype],
     [InputNoopMetatype, ONNXDepthwiseConvolutionMetatype, OutputNoopMetatype],
 ]
+QUANTIZED_REF_METATYPES_COUNTERS = [
+    REF_METATYPES_COUNTERS[0]
+    + [
+        ONNXQuantizeLinearMetatype,
+        ONNXDequantizeLinearMetatype,
+    ]
+    * 5,
+    REF_METATYPES_COUNTERS[1]
+    + [
+        ONNXQuantizeLinearMetatype,
+        ONNXDequantizeLinearMetatype,
+    ]
+    * 2,
+    REF_METATYPES_COUNTERS[2]
+    + [
+        ONNXQuantizeLinearMetatype,
+        ONNXDequantizeLinearMetatype,
+    ]
+    * 0,
+    REF_METATYPES_COUNTERS[3]
+    + [
+        ONNXQuantizeLinearMetatype,
+        ONNXDequantizeLinearMetatype,
+    ]
+    * 2,
+]
+
 
+@pytest.mark.parametrize(
+    ("model_creator_func, ref_metatypes, q_ref_metatypes"),
+    zip(TEST_MODELS, REF_METATYPES_COUNTERS, QUANTIZED_REF_METATYPES_COUNTERS),
+)
+def test_mapping_onnx_metatypes(model_creator_func, ref_metatypes, q_ref_metatypes):
+    def _check_metatypes(model, ref_metatypes):
+        nncf_graph = GraphConverter.create_nncf_graph(model)
+        actual_metatypes = [node.metatype for node in nncf_graph.get_all_nodes()]
+        assert Counter(ref_metatypes) == Counter(actual_metatypes)
 
-@pytest.mark.parametrize(("model_creator_func, ref_metatypes"), zip(TEST_MODELS, REF_METATYPES_COUNTERS))
-def test_mapping_onnx_metatypes(model_creator_func, ref_metatypes):
-    model = model_creator_func()
-    nncf_graph = GraphConverter.create_nncf_graph(model.onnx_model)
-    actual_metatypes = [node.metatype for node in nncf_graph.get_all_nodes()]
-    assert Counter(ref_metatypes) == Counter(actual_metatypes)
+    model = model_creator_func().onnx_model
+    q_model = min_max_quantize_model(model)
+    _check_metatypes(model, ref_metatypes)
+    _check_metatypes(q_model, q_ref_metatypes)