fix(fuse_batchnorm): support convtranpose + bn fusion with group != 1 (#2879)

Fixes #2867

Author: AyoubMDL

onnxscript/rewriter/rules/common/_fuse_batchnorm.py

@@ -2,7 +2,7 @@
 # Licensed under the MIT License.
 """Fuses BatchNormalization nodes into preceding nodes. Supported fusion patterns:
 - BatchNormalization ∘ Conv         -> Conv
-- BatchNormalization ∘ ConvTranpose -> ConvTranpose
+- BatchNormalization ∘ ConvTranspose -> ConvTranspose
 - BatchNormalization ∘ Gemm         -> Gemm
 
 Approach:
@@ -14,7 +14,7 @@
         - B_fused = (B - μ) * (gamma / std) + β
 """
 
-from abc import ABC, abstractmethod
+from abc import ABC
 from typing import ClassVar, Mapping
 
 import numpy as np
@@ -33,9 +33,18 @@ def _reshape_for_broadcast(x: np.ndarray, rank: int, axis: int = 1) -> np.ndarra
 class _FuseBatchNormBase(RewriteRuleClassBase, ABC):
     """Interface for BatchNormalization nodes fusion."""
 
-    @abstractmethod
     def get_filters_axis(self, attributes: Mapping[str, ir.Attr]) -> int:
         """Return the axis along which BatchNorm scale should be broadcasted."""
+        raise NotImplementedError()
+
+    def _scale_weights(
+        self,
+        weights: np.ndarray,
+        scale_factor: np.ndarray,
+        attributes: Mapping[str, ir.Attr],
+    ) -> np.ndarray:
+        axis = self.get_filters_axis(attributes)
+        return weights * _reshape_for_broadcast(scale_factor, weights.ndim, axis=axis)
 
     def rewrite(self, op, x: ir.Value, inbound_out: ir.Value, batchnorm_out: ir.Value):
         batchnorm_node = batchnorm_out.producer()
@@ -56,10 +65,8 @@ def rewrite(self, op, x: ir.Value, inbound_out: ir.Value, batchnorm_out: ir.Valu
         inbound_node = inbound_out.producer()
         weights = inbound_node.inputs[1].const_value.numpy()
 
-        # Reshape scale factor so it is broadcastable
-        axis = self.get_filters_axis(inbound_node.attributes)
         fused_weights = ir.tensor(
-            weights * _reshape_for_broadcast(scale_factor, weights.ndim, axis=axis)
+            self._scale_weights(weights, scale_factor, inbound_node.attributes)
         )
 
         # Update bias
@@ -144,8 +151,26 @@ class FuseBatchNormIntoConvTranspose(_FuseBatchNormBase):
 
     op_type: ClassVar = "ConvTranspose"
 
-    def get_filters_axis(self, attributes: Mapping[str, ir.Attr]) -> int:
-        return 1
+    def _scale_weights(
+        self,
+        weights: np.ndarray,
+        scale_factor: np.ndarray,
+        attributes: Mapping[str, ir.Attr],
+    ) -> np.ndarray:
+        # ConvTranspose weight: (in_channels, out_channels/group, *kernel)
+        # Reshape weights: [in_channels, out_channels/group, *kernel] → [group, in_channels/group, out_channels/group, *kernel]
+        in_channels = weights.shape[0]
+        out_channels_per_group = weights.shape[1]
+        kernel_shape = weights.shape[2:]
+        group = attributes.get("group", ir.AttrInt64("group", 1)).as_int()
+        w = weights.reshape(group, in_channels // group, out_channels_per_group, *kernel_shape)
+
+        # Per group scale_factor (out_channels,) -> (group, out_channels/group) -> (group, 1, out_channels/group, 1, ..., 1)
+        s = scale_factor.reshape((group, out_channels_per_group) + (1,) * len(kernel_shape))
+        # insert in_channels/group axis -> (group, 1, out_channels/group, *ones)
+        s = s[:, None, ...]
+
+        return (w * s).reshape(weights.shape)
 
     def pattern(self, op, x):
         return op.BatchNormalization(
@@ -154,6 +179,25 @@ def pattern(self, op, x):
             _outputs=["batchnorm_out"],
         )
 
+    def check(self, context, x, inbound_out, batchnorm_out):
+        check_result = super().check(context, x, inbound_out, batchnorm_out)
+        if not check_result:
+            return check_result
+
+        inbound_node = inbound_out.producer()
+
+        in_channels = inbound_node.inputs[1].const_value.numpy().shape[0]
+        group = inbound_node.attributes.get("group", ir.AttrInt64("group", 1)).as_int()
+
+        # Check that in_channels is divisible by group as ONNX checker allows it
+        # But this is invalid case
+        if in_channels % group != 0:
+            return check_result.fail(
+                f"ConvTranspose in_channels ({in_channels}) is not divisible by group ({group})."
+            )
+
+        return check_result
+
 
 class FuseBatchNormIntoGemm(_FuseBatchNormBase):
     """Replaces ``BatchNormalization(Gemm(x))`` with ``Gemm(x)``."""

onnxscript/rewriter/rules/common/_fuse_batchnorm_test.py

@@ -3,8 +3,7 @@
 import unittest
 
 import numpy as np
-import onnx.checker
-import onnx.parser
+import onnx
 import parameterized
 
 from onnxscript import ir
@@ -31,29 +30,33 @@ def _create_batchnorm_params(self, size: int):
 
     @parameterized.parameterized.expand(
         [
-            ("bias_false", False),
-            ("bias_true", True),
+            ("bias_false_group1", False, 1),
+            ("bias_true_group1", True, 1),
+            ("bias_false_group4", False, 4),
+            ("bias_true_group4", True, 4),
         ]
     )
-    def test_fuse_batchnorm_convtranspose(self, _: str, convtranspose_bias: bool):
+    def test_fuse_batchnorm_convtranspose(self, _: str, convtranspose_bias: bool, group: int):
+        # ConvTranspose weight: [in_channels, out_channels/group, kH, kW]
+        out_channels = 64 * group
         convtranspose_inputs = "X, W"
         parameters = (
-            "float[32, 64, 3, 3] W, "
-            "float[64] gamma, "
-            "float[64] beta, "
-            "float[64] input_mean, "
-            "float[64] input_var"
+            f"float[32, 64, 3, 3] W, "
+            f"float[{out_channels}] gamma, "
+            f"float[{out_channels}] beta, "
+            f"float[{out_channels}] input_mean, "
+            f"float[{out_channels}] input_var"
         )
         if convtranspose_bias:
-            parameters += ", float[64] B"
+            parameters += f", float[{out_channels}] B"
             convtranspose_inputs += ", B"
 
         model_proto = onnx.parser.parse_model(f"""
             < ir_version: 7, opset_import: ["" : 17] >
             test_model (float[N, 32, 14, 16] X) => (float [N, ?, ?, ?] Y)
             <{parameters}>
             {{
-                X1 = ConvTranspose({convtranspose_inputs})
+                X1 = ConvTranspose<group={group}>({convtranspose_inputs})
                 Y = BatchNormalization(X1, gamma, beta, input_mean, input_var)
             }}
         """)
@@ -62,11 +65,13 @@ def test_fuse_batchnorm_convtranspose(self, _: str, convtranspose_bias: bool):
             onnx.numpy_helper.from_array(
                 np.random.randn(32, 64, 3, 3).astype(np.float32), name="W"
             ),
-            *self._create_batchnorm_params(size=64),
+            *self._create_batchnorm_params(size=out_channels),
         ]
         if convtranspose_bias:
             initializers.append(
-                onnx.numpy_helper.from_array(np.random.randn(64).astype(np.float32), name="B")
+                onnx.numpy_helper.from_array(
+                    np.random.randn(out_channels).astype(np.float32), name="B"
+                )
             )
         model_proto.graph.initializer.extend(initializers)
 
@@ -90,14 +95,18 @@ def test_fuse_batchnorm_convtranspose(self, _: str, convtranspose_bias: bool):
 
     @parameterized.parameterized.expand(
         [
-            ("bias_false", False),
-            ("bias_true", True),
+            ("bias_false_group1", False, 1),
+            ("bias_true_group1", True, 1),
+            ("bias_false_group2", False, 2),
+            ("bias_true_group2", True, 2),
         ]
     )
-    def test_fuse_batchnorm_conv(self, _: str, conv_bias: bool):
+    def test_fuse_batchnorm_conv(self, _: str, conv_bias: bool, group: int):
+        # Conv weight: [out_channels, in_channels/group, kH, kW]
+        in_channels_per_group = 32 // group
         conv_inputs = "X, W"
         parameters = (
-            "float[64, 32, 3, 3] W, "
+            f"float[64, {in_channels_per_group}, 3, 3] W, "
             "float[64] gamma, "
             "float[64] beta, "
             "float[64] input_mean, "
@@ -112,14 +121,14 @@ def test_fuse_batchnorm_conv(self, _: str, conv_bias: bool):
             test_model (float[N, 32, 14, 16] X) => (float [N, ?, ?, ?] Y)
             <{parameters}>
             {{
-                X1 = Conv({conv_inputs})
+                X1 = Conv<group={group}>({conv_inputs})
                 Y = BatchNormalization(X1, gamma, beta, input_mean, input_var)
             }}
         """)
         # Add initializers
         initializers = [
             onnx.numpy_helper.from_array(
-                np.random.randn(64, 32, 3, 3).astype(np.float32), name="W"
+                np.random.randn(64, in_channels_per_group, 3, 3).astype(np.float32), name="W"
             ),
             *self._create_batchnorm_params(size=64),
         ]
@@ -211,6 +220,32 @@ def test_fuse_batchnorm_gemm(self, _: str, gemm_bias: bool, transB: int):
         output_model_proto = ir.serde.serialize_model(model)
         onnx.checker.check_model(output_model_proto, True)
 
+    def test_fuse_batchnorm_convtranspose_grouped_invalid_skipped(self):
+        """Fusion is skipped when in_channels is not divisible by group (semantically invalid model)."""
+        # in_channels=32 is not divisible by group=3, the ONNX checker won't catch this.
+        model_proto = onnx.parser.parse_model("""
+            < ir_version: 7, opset_import: ["" : 17] >
+            test_model (float[N, 32, 14, 14] X) => (float[N, ?, ?, ?] Y)
+            <float[32, 64, 3, 3] W,
+             float[192] gamma, float[192] beta, float[192] input_mean, float[192] input_var>
+            {
+                X1 = ConvTranspose<group=3>(X, W)
+                Y = BatchNormalization(X1, gamma, beta, input_mean, input_var)
+            }
+        """)
+        initializers = [
+            onnx.numpy_helper.from_array(
+                np.random.randn(32, 64, 3, 3).astype(np.float32), name="W"
+            ),
+            *self._create_batchnorm_params(size=192),
+        ]
+        model_proto.graph.initializer.extend(initializers)
+        model = ir.serde.deserialize_model(model_proto)
+        count = _fuse_batchnorm.rules.apply_to_model(model)
+
+        # Fusion must be skipped, applying it would crash on the invalid dimensions.
+        self.assertEqual(count, 0)
+
     def test_fuse_batchnorm_non_initializers(self):
         model_proto = onnx.parser.parse_model("""
             < ir_version: 7, opset_import: ["" : 17] >