rewrite Pattern_3.pattern_condition_checker to filter new test case: …

…Conv + Mul(w/ one init) + Add(w/o init)

python/furiosa-quantizer/furiosa/quantizer/frontend/onnx/transformer/fuse_bn_into_conv.py

@@ -230,18 +230,10 @@ def pattern_matching(self, base_node):
     def pattern_condition_checker(self, nodes_to_check):
         _, mul_node, add_node = nodes_to_check
 
-        if self.check_condition_1(mul_node):
-            return True
-
-        if self.check_condition_1(add_node):
-            return True
-
-        return False
-
-    def check_condition_1(self, node):
-        if self.get_init_node_input(node):
-            return True
-        return False
+        # This checks if a node has a initializer, \
+        # assuming a node has exactly one initializer if it has one. \
+        # That is, there is no node with two initialzier.
+        return self.get_init_node_input(mul_node) and self.get_init_node_input(add_node)
 
     def make_new_node(self, matched_nodes):
         top_node, middle_node, bottom_node = matched_nodes

python/furiosa-quantizer/tests/frontend/onnx/transformer/test_fuse_bn_into_conv.py

@@ -60,6 +60,21 @@ def forward(self, x):
         return x
 
 
+class UnitTestModel3_1(UnitTestModel):
+    """
+    This creates Conv + Mul + Add graph for testing Pattern_3
+    """
+
+    def __init__(self, in_channel, out_channel):
+        super().__init__(in_channel, out_channel)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = torch.mul(x, torch.ones((1, x.shape[1], 1, 1)))
+        x = torch.add(x, x)
+        return x
+
+
 class MultiTestModel(UnitTestModel):
     def __init__(self, in_channel, out_channel):
         super(MultiTestModel, self).__init__(in_channel, out_channel)
@@ -163,7 +178,7 @@ def test_case5(self):
         self.check_output_value(orig_model, trans_model, input_shapes)
         self.check_value_info(trans_model)
 
-    def test_case5(self):
+    def test_case5_1(self):
         """
         This tests Pattern_3
         """
@@ -179,3 +194,17 @@ def test_case5(self):
         self.check_graph_node(trans_model, op_types)
         self.check_output_value(orig_model, trans_model, input_shapes)
         self.check_value_info(trans_model)
+
+    def test_case5_2(self):
+        input_shapes = [(1, 4, 4, 4)]
+        in_channel = 4
+        out_channel = 8
+
+        op_types = ['Conv', 'Mul', 'Add']
+
+        orig_model, trans_model = self._make_test_model(
+            UnitTestModel3_1(in_channel, out_channel), input_shapes
+        )
+        self.check_graph_node(trans_model, op_types)
+        self.check_output_value(orig_model, trans_model, input_shapes)
+        self.check_value_info(trans_model)