[Torch] SE block HW fusing pattern is presented (#2177)

### Changes * SE block ignored pattern is presented for Torch backend ![image](https://github.com/openvinotoolkit/nncf/assets/74656388/2bbf0d3c-ac9d-4db0-9c28-ff6b1bf299af) * SE block ignored pattern for OV backend is adjusted * NON_PATTERN_NODE_WITH_TYPE is introduced in the graph matcher ### Reason for changes To align FQ layout between TORCH and OV backend for efficientnet_b0 model To fix FQ placement for SE block: Before: ![image](https://github.com/openvinotoolkit/nncf/assets/74656388/7e8b27a3-fafc-41f9-ace2-1945bce32b65) After: ![image](https://github.com/openvinotoolkit/nncf/assets/74656388/e4479c90-bd4f-47fb-b2c2-acd76a685030) ### Related tickets 121647 ### Tests * test_non_pattern_node_with_type for the graph matcher Post training quantization manual build 180 | Model | Backend | Metric name | Metric value | Metric diff | Num FQ | RAM MiB | Quant. time | Total time | Status | |-------|----------------------|-------------|--------------|-------------|---------|---------|-------------|------------|---------| | 1 | timm/efficientnet_b0 | OV | Acc@1 | 0.7688 | -0.0042 | 92 | 1033 | 0:00:25 | 1:03:58 | | | 2 | timm/efficientnet_b0 | TORCH | Acc@1 | 0.7680 | -0.0049 | 92 | 892 | 0:00:48 | 1:04:27 | * pattern manager tests
openvinotoolkit · Oct 24, 2023 · 5336405 · 5336405
1 parent f415d90
commit 5336405
Show file tree

Hide file tree

Showing 14 changed files with 875 additions and 687 deletions.
diff --git a/nncf/common/graph/graph_matching.py b/nncf/common/graph/graph_matching.py
@@ -15,10 +15,12 @@
 
 from nncf.common.graph.patterns import GraphPattern
 
+ATTRS_TO_SKIP = [GraphPattern.LABEL_ATTR, GraphPattern.PATTERN_NODE_TO_EXCLUDE]
+
 
 def _are_nodes_matched(node_1, node_2) -> bool:
     for attr in node_2:
-        if attr == GraphPattern.LABEL_ATTR:
+        if attr in ATTRS_TO_SKIP:
             continue
         if attr == GraphPattern.METATYPE_ATTR:
             # GraphPattern.ANY_PATTERN_NODE_TYPE and GraphPattern.NON_PATTERN_NODE_TYPE
@@ -103,7 +105,8 @@ def _is_subgraph_matching_strict(graph: nx.DiGraph, pattern: nx.DiGraph, subgrap
 
 def _copy_subgraph_excluding_non_pattern_node(subgraph: Dict[str, str], pattern_graph: GraphPattern) -> Dict[str, str]:
     """
-    Copies a matching subgraph excluding the nodes having GraphPattern.NON_PATTERN_NODE_TYPE.
+    Copies a matching subgraph excluding the nodes having GraphPattern.NON_PATTERN_NODE_TYPE
+       or GraphPattern.PATTERN_NODE_TO_EXCLUDE.
 
     :param subgraph: Subgraph
     :param pattern_graph: A graph consists of patterns to match.
@@ -113,8 +116,12 @@ def _copy_subgraph_excluding_non_pattern_node(subgraph: Dict[str, str], pattern_
     for node_from_graph, node_from_pattern in subgraph.items():
         pattern_node = pattern_graph.graph.nodes[node_from_pattern]
         pattern_node_types = pattern_node.get(GraphPattern.METATYPE_ATTR, [])
-        if GraphPattern.NON_PATTERN_NODE_TYPE not in pattern_node_types:
-            output[node_from_graph] = node_from_pattern
+        if GraphPattern.NON_PATTERN_NODE_TYPE in pattern_node_types:
+            continue
+        if pattern_node.get(GraphPattern.PATTERN_NODE_TO_EXCLUDE, False):
+            continue
+        output[node_from_graph] = node_from_pattern
+
     return output
 
 

diff --git a/nncf/common/graph/patterns/patterns.py b/nncf/common/graph/patterns/patterns.py
@@ -80,6 +80,7 @@ class GraphPattern:
     NODE_TYPE_ATTR = "metatype"
     ANY_PATTERN_NODE_TYPE = "ANY_PATTERN_NODE"
     NON_PATTERN_NODE_TYPE = "NON_PATTERN_NODE"
+    PATTERN_NODE_TO_EXCLUDE = "PATTERN_NODE_TO_EXCLUDE"
 
     def __init__(self):
         self._graph = nx.DiGraph()
@@ -298,7 +299,6 @@ class HWFusedPatternNames(Enum):
     NORMALIZE_L2_MULTIPLY = PatternDesc("normalize_l2_multiply")
     SCALE_SHIFT = PatternDesc("scale_shift")
     SHIFT_SCALE = PatternDesc("shift_scale")
-    SE_BLOCK = PatternDesc("se_block")
     SOFTMAX_DIV = PatternDesc("softmax_div")
 
     # ACTIVATIONS
@@ -396,5 +396,6 @@ class IgnoredPatternNames(Enum):
         model_types=[ModelType.TRANSFORMER],
         devices=[TargetDevice.ANY, TargetDevice.CPU, TargetDevice.GPU, TargetDevice.VPU],
     )
+    SE_BLOCK = PatternDesc("se_block")
     FC_BN_HSWISH_ACTIVATION = PatternDesc("fc_bn_hswish_activation")
     EQUAL_LOGICALNOT = PatternDesc("equal_logicalnot")
diff --git a/nncf/openvino/hardware/fused_patterns.py b/nncf/openvino/hardware/fused_patterns.py
@@ -147,48 +147,6 @@ def create_shift_scale() -> GraphPattern:
     return pattern
 
 
-@OPENVINO_HW_FUSED_PATTERNS.register(HWFusedPatternNames.SE_BLOCK)
-def create_se_block() -> GraphPattern:
-    pattern = GraphPattern()
-    any_node = pattern.add_node(
-        **{GraphPattern.LABEL_ATTR: "ANY", GraphPattern.METATYPE_ATTR: GraphPattern.NON_PATTERN_NODE_TYPE}
-    )
-    reduce_mean_node = pattern.add_node(
-        **{GraphPattern.LABEL_ATTR: "REDUCE_MEAN", GraphPattern.METATYPE_ATTR: om.OVReduceMeanMetatype}
-    )
-    linear_node_1 = pattern.add_node(
-        **{GraphPattern.METATYPE_ATTR: LINEAR_OPERATIONS, GraphPattern.LABEL_ATTR: "LINEAR"}
-    )
-    add_node_1 = pattern.add_node(**{GraphPattern.LABEL_ATTR: "ADD_BIAS", GraphPattern.METATYPE_ATTR: om.OVAddMetatype})
-    activation_node_1 = pattern.add_node(
-        **{
-            GraphPattern.LABEL_ATTR: "RELU, PRELU, SWISH",
-            GraphPattern.METATYPE_ATTR: [om.OVReluMetatype, om.OVPReluMetatype, om.OVSwishMetatype],
-        }
-    )
-    linear_node_2 = pattern.add_node(
-        **{GraphPattern.METATYPE_ATTR: LINEAR_OPERATIONS, GraphPattern.LABEL_ATTR: "LINEAR"}
-    )
-    add_node_2 = pattern.add_node(**{GraphPattern.LABEL_ATTR: "ADD_BIAS", GraphPattern.METATYPE_ATTR: om.OVAddMetatype})
-    activation_node_2 = pattern.add_node(
-        **{GraphPattern.LABEL_ATTR: "SIGMOID", GraphPattern.METATYPE_ATTR: om.OVSigmoidMetatype}
-    )
-    multiply_node = pattern.add_node(
-        **{GraphPattern.LABEL_ATTR: "MULTIPLY", GraphPattern.METATYPE_ATTR: om.OVMultiplyMetatype}
-    )
-
-    pattern.add_edge(any_node, reduce_mean_node)
-    pattern.add_edge(reduce_mean_node, linear_node_1)
-    pattern.add_edge(linear_node_1, add_node_1)
-    pattern.add_edge(add_node_1, activation_node_1)
-    pattern.add_edge(activation_node_1, linear_node_2)
-    pattern.add_edge(linear_node_2, add_node_2)
-    pattern.add_edge(add_node_2, activation_node_2)
-    pattern.add_edge(activation_node_2, multiply_node)
-    pattern.add_edge(any_node, multiply_node)
-    return pattern
-
-
 @OPENVINO_HW_FUSED_PATTERNS.register(HWFusedPatternNames.SOFTMAX_DIV)
 def create_softmax_div() -> GraphPattern:
     pattern = GraphPattern()

diff --git a/nncf/openvino/quantization/ignored_patterns.py b/nncf/openvino/quantization/ignored_patterns.py
@@ -12,6 +12,7 @@
 from nncf.common.graph.patterns.patterns import IgnoredPatternNames
 from nncf.common.utils.registry import Registry
 from nncf.openvino.graph.metatypes import openvino_metatypes as om
+from nncf.openvino.graph.metatypes.groups import LINEAR_OPERATIONS
 
 OPENVINO_IGNORED_PATTERNS = Registry("IGNORED_PATTERNS")
 
@@ -108,3 +109,53 @@ def create_equal_logicalnot() -> GraphPattern:
 
     pattern.add_edge(equal_node, logical_not_node)
     return pattern
+
+
+@OPENVINO_IGNORED_PATTERNS.register(IgnoredPatternNames.SE_BLOCK)
+def create_se_block() -> GraphPattern:
+    pattern = GraphPattern()
+    any_node = pattern.add_node(
+        **{GraphPattern.LABEL_ATTR: "ANY", GraphPattern.METATYPE_ATTR: GraphPattern.NON_PATTERN_NODE_TYPE}
+    )
+    reduce_mean_node = pattern.add_node(
+        **{
+            GraphPattern.LABEL_ATTR: "REDUCE_MEAN",
+            GraphPattern.METATYPE_ATTR: om.OVReduceMeanMetatype,
+            GraphPattern.PATTERN_NODE_TO_EXCLUDE: True,
+        }
+    )
+    linear_node_1 = pattern.add_node(
+        **{GraphPattern.METATYPE_ATTR: LINEAR_OPERATIONS, GraphPattern.LABEL_ATTR: "LINEAR"}
+    )
+    add_node_1 = pattern.add_node(**{GraphPattern.LABEL_ATTR: "ADD_BIAS", GraphPattern.METATYPE_ATTR: om.OVAddMetatype})
+    activation_node_1 = pattern.add_node(
+        **{
+            GraphPattern.LABEL_ATTR: "RELU, PRELU, SWISH",
+            GraphPattern.METATYPE_ATTR: [om.OVReluMetatype, om.OVPReluMetatype, om.OVSwishMetatype],
+        }
+    )
+    linear_node_2 = pattern.add_node(
+        **{GraphPattern.METATYPE_ATTR: LINEAR_OPERATIONS, GraphPattern.LABEL_ATTR: "LINEAR"}
+    )
+    add_node_2 = pattern.add_node(**{GraphPattern.LABEL_ATTR: "ADD_BIAS", GraphPattern.METATYPE_ATTR: om.OVAddMetatype})
+    activation_node_2 = pattern.add_node(
+        **{GraphPattern.LABEL_ATTR: "SIGMOID", GraphPattern.METATYPE_ATTR: om.OVSigmoidMetatype}
+    )
+    multiply_node = pattern.add_node(
+        **{
+            GraphPattern.LABEL_ATTR: "MULTIPLY",
+            GraphPattern.METATYPE_ATTR: om.OVMultiplyMetatype,
+            GraphPattern.PATTERN_NODE_TO_EXCLUDE: True,
+        }
+    )
+
+    pattern.add_edge(any_node, reduce_mean_node)
+    pattern.add_edge(reduce_mean_node, linear_node_1)
+    pattern.add_edge(linear_node_1, add_node_1)
+    pattern.add_edge(add_node_1, activation_node_1)
+    pattern.add_edge(activation_node_1, linear_node_2)
+    pattern.add_edge(linear_node_2, add_node_2)
+    pattern.add_edge(add_node_2, activation_node_2)
+    pattern.add_edge(activation_node_2, multiply_node)
+    pattern.add_edge(any_node, multiply_node)
+    return pattern
diff --git a/nncf/torch/quantization/ignored_patterns.py b/nncf/torch/quantization/ignored_patterns.py
@@ -11,6 +11,8 @@
 from nncf.common.graph.patterns.patterns import GraphPattern
 from nncf.common.graph.patterns.patterns import IgnoredPatternNames
 from nncf.common.utils.registry import Registry
+from nncf.torch.graph.pattern_operations import ATOMIC_ACTIVATIONS_OPERATIONS
+from nncf.torch.graph.pattern_operations import LINEAR_OPERATIONS
 
 PT_IGNORED_PATTERNS = Registry("IGNORED_PATTERNS")
 
@@ -88,3 +90,126 @@ def create_multihead_attention_output() -> GraphPattern:
         transpose_aliases=transpose_aliases,
     )
     return pattern
+
+
+# pylint:disable=too-many-statements
+@PT_IGNORED_PATTERNS.register(IgnoredPatternNames.SE_BLOCK)
+def create_se_block() -> GraphPattern:
+    MEAN_OPERATIONS = {
+        GraphPattern.LABEL_ATTR: "REDUCE_MEAN",
+        GraphPattern.METATYPE_ATTR: ["avg_pool2d", "adaptive_avg_pool2d", "avg_pool3d", "adaptive_avg_pool3d", "mean"],
+        GraphPattern.PATTERN_NODE_TO_EXCLUDE: True,
+    }
+    SYGMOID_OPERATIONS = {
+        GraphPattern.LABEL_ATTR: "SIGMOID",
+        GraphPattern.METATYPE_ATTR: ["sigmoid", "hardsigmoid"],
+    }
+    MUL_OPERATION = {
+        GraphPattern.LABEL_ATTR: "MUL",
+        GraphPattern.METATYPE_ATTR: "__mul__",
+        GraphPattern.PATTERN_NODE_TO_EXCLUDE: True,
+    }
+
+    def get_se_block_pattern() -> GraphPattern:
+        pattern = GraphPattern()
+        any_node = pattern.add_node(label="NON_PATTERN_NODE", type=GraphPattern.NON_PATTERN_NODE_TYPE)
+        reduce_mean_node = pattern.add_node(**MEAN_OPERATIONS)
+        linear_node_1 = pattern.add_node(**LINEAR_OPERATIONS)
+        activation_node_1 = pattern.add_node(**ATOMIC_ACTIVATIONS_OPERATIONS)
+        linear_node_2 = pattern.add_node(**LINEAR_OPERATIONS)
+        activation_node_2 = pattern.add_node(**SYGMOID_OPERATIONS)
+        multiply_node = pattern.add_node(**MUL_OPERATION)
+
+        pattern.add_edge(any_node, reduce_mean_node)
+        pattern.add_edge(reduce_mean_node, linear_node_1)
+        pattern.add_edge(linear_node_1, activation_node_1)
+        pattern.add_edge(activation_node_1, linear_node_2)
+        pattern.add_edge(linear_node_2, activation_node_2)
+        pattern.add_edge(activation_node_2, multiply_node)
+        pattern.add_edge(any_node, multiply_node)
+        return pattern
+
+    def get_se_block_with_bias_pattern() -> GraphPattern:
+        pattern = GraphPattern()
+        any_node = pattern.add_node(label="NON_PATTERN_NODE", type=GraphPattern.NON_PATTERN_NODE_TYPE)
+        reduce_mean_node = pattern.add_node(**MEAN_OPERATIONS)
+        linear_node_1 = pattern.add_node(**LINEAR_OPERATIONS)
+        add_node_1 = pattern.add_node(label="ADD_BIAS", type=["__add__", "__sub__"])
+        activation_node_1 = pattern.add_node(**ATOMIC_ACTIVATIONS_OPERATIONS)
+        linear_node_2 = pattern.add_node(**LINEAR_OPERATIONS)
+        add_node_2 = pattern.add_node(label="ADD_BIAS", type=["__add__", "__sub__"])
+        activation_node_2 = pattern.add_node(**SYGMOID_OPERATIONS)
+        multiply_node = pattern.add_node(**MUL_OPERATION)
+
+        pattern.add_edge(any_node, reduce_mean_node)
+        pattern.add_edge(reduce_mean_node, linear_node_1)
+        pattern.add_edge(linear_node_1, add_node_1)
+        pattern.add_edge(add_node_1, activation_node_1)
+        pattern.add_edge(activation_node_1, linear_node_2)
+        pattern.add_edge(linear_node_2, add_node_2)
+        pattern.add_edge(add_node_2, activation_node_2)
+        pattern.add_edge(activation_node_2, multiply_node)
+        pattern.add_edge(any_node, multiply_node)
+        return pattern
+
+    RESHAPE_NODES = {
+        GraphPattern.LABEL_ATTR: "RESHAPE",
+        GraphPattern.METATYPE_ATTR: ["reshape", "view", "flatten", "unsqueeze"],
+    }
+
+    def get_se_block_with_reshape() -> GraphPattern:
+        pattern = GraphPattern()
+        any_node = pattern.add_node(label="NON_PATTERN_NODE", type=GraphPattern.NON_PATTERN_NODE_TYPE)
+        reduce_mean_node = pattern.add_node(**MEAN_OPERATIONS)
+        reshape_node_1 = pattern.add_node(**RESHAPE_NODES)
+        linear_node_1 = pattern.add_node(**LINEAR_OPERATIONS)
+        activation_node_1 = pattern.add_node(**ATOMIC_ACTIVATIONS_OPERATIONS)
+        linear_node_2 = pattern.add_node(**LINEAR_OPERATIONS)
+        activation_node_2 = pattern.add_node(**SYGMOID_OPERATIONS)
+        reshape_node_2 = pattern.add_node(**RESHAPE_NODES)
+        multiply_node = pattern.add_node(**MUL_OPERATION)
+
+        pattern.add_edge(any_node, reduce_mean_node)
+        pattern.add_edge(reduce_mean_node, reshape_node_1)
+        pattern.add_edge(reshape_node_1, linear_node_1)
+        pattern.add_edge(linear_node_1, activation_node_1)
+        pattern.add_edge(activation_node_1, linear_node_2)
+        pattern.add_edge(linear_node_2, activation_node_2)
+        pattern.add_edge(activation_node_2, reshape_node_2)
+        pattern.add_edge(reshape_node_2, multiply_node)
+        pattern.add_edge(any_node, multiply_node)
+        return pattern
+
+    def get_se_block_with_bias_and_reshape() -> GraphPattern:
+        pattern = GraphPattern()
+        any_node = pattern.add_node(label="NON_PATTERN_NODE", type=GraphPattern.NON_PATTERN_NODE_TYPE)
+        reduce_mean_node = pattern.add_node(**MEAN_OPERATIONS)
+        reshape_node_1 = pattern.add_node(**RESHAPE_NODES)
+        linear_node_1 = pattern.add_node(**LINEAR_OPERATIONS)
+        add_node_1 = pattern.add_node(label="ADD_BIAS", type=["__add__", "__sub__"])
+        activation_node_1 = pattern.add_node(**ATOMIC_ACTIVATIONS_OPERATIONS)
+        linear_node_2 = pattern.add_node(**LINEAR_OPERATIONS)
+        add_node_2 = pattern.add_node(label="ADD_BIAS", type=["__add__", "__sub__"])
+        activation_node_2 = pattern.add_node(**SYGMOID_OPERATIONS)
+        reshape_node_2 = pattern.add_node(**RESHAPE_NODES)
+        multiply_node = pattern.add_node(**MUL_OPERATION)
+
+        pattern.add_edge(any_node, reduce_mean_node)
+        pattern.add_edge(reduce_mean_node, reshape_node_1)
+        pattern.add_edge(reshape_node_1, linear_node_1)
+        pattern.add_edge(linear_node_1, add_node_1)
+        pattern.add_edge(add_node_1, activation_node_1)
+        pattern.add_edge(activation_node_1, linear_node_2)
+        pattern.add_edge(linear_node_2, add_node_2)
+        pattern.add_edge(add_node_2, activation_node_2)
+        pattern.add_edge(activation_node_2, reshape_node_2)
+        pattern.add_edge(reshape_node_2, multiply_node)
+        pattern.add_edge(any_node, multiply_node)
+        return pattern
+
+    main_pattern = GraphPattern()
+    main_pattern.add_pattern_alternative(get_se_block_pattern())
+    main_pattern.add_pattern_alternative(get_se_block_with_bias_pattern())
+    main_pattern.add_pattern_alternative(get_se_block_with_reshape())
+    main_pattern.add_pattern_alternative(get_se_block_with_bias_and_reshape())
+    return main_pattern
diff --git a/tests/common/graph/test_graph_matching.py b/tests/common/graph/test_graph_matching.py
@@ -227,3 +227,28 @@ def test_not_match_edges_inside_pattern():
     pattern.add_edge(node_1, node_3)
     matches = find_subgraphs_matching_pattern(ref_graph, pattern)
     assert matches == [["1", "2", "3"]]
+
+
+def test_non_pattern_graph_with_type():
+    for match in [False, True]:
+        ref_graph = nx.DiGraph()
+        ref_graph.add_node("0", **{GraphPattern.METATYPE_ATTR: "0"})
+        ref_graph.add_node("1", **{GraphPattern.METATYPE_ATTR: "a" if match else "0"})
+        ref_graph.add_node("2", **{GraphPattern.METATYPE_ATTR: "b"})
+        ref_graph.add_node("3", **{GraphPattern.METATYPE_ATTR: "c"})
+        ref_graph.add_edge("0", "1")
+        ref_graph.add_edge("1", "2")
+        ref_graph.add_edge("2", "3")
+
+        pattern = GraphPattern()
+        node_1 = pattern.add_node(**{GraphPattern.METATYPE_ATTR: "a", GraphPattern.PATTERN_NODE_TO_EXCLUDE: True})
+        node_2 = pattern.add_node(**{GraphPattern.METATYPE_ATTR: "b"})
+        node_3 = pattern.add_node(**{GraphPattern.METATYPE_ATTR: "c"})
+        pattern.add_edge(node_1, node_2)
+        pattern.add_edge(node_2, node_3)
+
+        matches = find_subgraphs_matching_pattern(ref_graph, pattern)
+        if not match:
+            assert not matches
+        else:
+            assert matches == [["2", "3"]]