Fix ava-person-vehicle-detection-stage2-2_0_0 tuning failed (#1135)

Author: lvliang-intel

neural_compressor/adaptor/tf_utils/graph_rewriter/generic/fuse_decomposed_bn.py

@@ -169,6 +169,13 @@ def do_transformation(self):
 
             # Mul (input, Mul)
             input_data_op = node_from_map(input_node_map, data_scale_mul_op.input[0])
+            # Workaround for model ava-person-vehicle-detection-stage2-2_0_0
+            # FusedBatchNorm requires a 4D Tensor for input data, but MatMul only support 2D output.
+            # Don't fuse the small ops to FusedBatchNorm when the upstream has MatMul.
+            ancestor_input_data_op = node_from_map(input_node_map, input_data_op.input[0])
+            if input_data_op.op == "MatMul" or ancestor_input_data_op.op == "MatMul":
+                continue
+                    
             scale_op = node_from_map(input_node_map, data_scale_mul_op.input[1])
 
             if scale_op.op == "Rsqrt":
@@ -370,4 +377,4 @@ def get_const_dim_count(node_def):
         Number of dimensions for the Const node.
     """
     const_value = values_from_const(node_def)
-    return const_value.ndim
+    return const_value.ndim