Fix Pad's quantization (microsoft#17807)

Fix microsoft#17760. Upstream exporter creates empty string as Pad's 3rd input
and the quantization tool 1) considers that as a valid tensor name and
2) adds corresponding invalid quantization nodes. This PR adds a
condition check to make quantization tool working.

onnxruntime/python/tools/quantization/onnx_quantizer.py

@@ -645,6 +645,7 @@ def _get_quantize_input_nodes(self, node, input_index, qType, given_scale_name=N
         :return: List of newly created nodes in NodeProto format.
         """
         input_name = node.input[input_index]
+        assert input_name != "", "Cannot access undefined variable in graph."
         output_name = input_name + TENSOR_NAME_QUANT_SUFFIX
         ql_node_name = input_name + "_QuantizeLinear"
 

onnxruntime/python/tools/quantization/operators/pad.py

@@ -31,7 +31,7 @@ def quantize(self):
             kwargs.update(kv)
 
         if "mode" not in kwargs or kwargs["mode"] == b"constant":
-            if len(node.input) > 2:  # There is 3rd input 'constant_value'
+            if len(node.input) > 2 and node.input[2] != "":  # There is 3rd input 'constant_value'
                 zp_tensor = self.quantizer.model.get_initializer(quantized_input_value.zp_name)
                 scale_tensor = self.quantizer.model.get_initializer(quantized_input_value.scale_name)
                 if zp_tensor is None or scale_tensor is None:
@@ -72,7 +72,17 @@ def quantize(self):
                     self.quantizer.new_nodes.extend(pad_value_qnodes)
                     node.input[2] = pad_value_qnodes[0].output[0]
             else:
-                node.input.extend([quantized_input_value.zp_name])  # pad zero_point for original zero
+                # In quantized format, the `zero` before quantization is mapped
+                # to quantized_input_value.zp_name. Thus, padding 0 to
+                # original tensor should become padding zero point to quantized
+                # tensor.
+                if len(node.input) == 2:
+                    # Feed quantization's zero point to padding node.
+                    node.input.append(quantized_input_value.zp_name)
+                else:
+                    # Assign quantization's zero point to padding node.
+                    assert node.input[2] == ""
+                    node.input[2] = quantized_input_value.zp_name
 
         # Create an entry for output quantized value
         quantized_output_value = QuantizedValue(

onnxruntime/test/python/quantization/test_op_pad.py

@@ -5,6 +5,7 @@
 # license information.
 # --------------------------------------------------------------------------
 
+import itertools
 import unittest
 
 import numpy as np
@@ -404,6 +405,123 @@ def test_static_mode_constant_value_edge_case(self):
             "constant", constant_value=0.1, quantize_mode="static", extra_options={"dual_feed": True}
         )
 
+    @classmethod
+    def construct_model_add_pad_add(
+        cls,
+        # Name of model input, i.e., "input" in the illustration graph below.
+        name,
+        # Name of model output.
+        final_name,
+        # model input shape.
+        shape,
+    ):
+        # Graph implemented below is
+        #  `name`, `name` -> Add -> "first_add_output"
+        #  "first_add_output", "pads" -> Pad -> "pad_output"
+        #  "pad_output", "pad_output" -> Add -> `final_name`
+        # where `name` is the 2nd argument of this function,
+        # `final_name` is the 3rd argument of this function,
+        # and the rest lowercase strings are tensor names in the graph.
+
+        input_name = name
+        first_add_output_name = "first_add_output"
+        pads_name = "pads"
+        pad_output_name = "pad_output"
+        second_add_output_name = final_name
+
+        input_shape = shape
+        input_rank = len(input_shape)
+
+        input_tensor = helper.make_tensor_value_info(input_name, TensorProto.FLOAT, input_shape)
+
+        first_add_node = helper.make_node(
+            "Add",
+            [input_name, input_name],
+            [first_add_output_name],
+            name="FirstAdd",
+        )
+
+        pads = [1, 2] * input_rank
+        pads_initializer = helper.make_tensor(
+            pads_name,
+            TensorProto.INT64,
+            # 1-D tensor of shape [2 * input_rank].
+            [len(pads)],
+            pads,
+        )
+        pad_node = helper.make_node(
+            "Pad",
+            [first_add_output_name, pads_name, ""],
+            [pad_output_name],
+            name="PadNode",
+            mode="constant",
+        )
+        pad_output_shape = tuple(input_shape[i] + pads[i] + pads[i + input_rank] for i in range(input_rank))
+
+        second_add_node = helper.make_node(
+            "Add",
+            [pad_output_name, pad_output_name],
+            [second_add_output_name],
+            name="SecondAdd",
+        )
+
+        output_tensor = helper.make_tensor_value_info(second_add_output_name, TensorProto.FLOAT, pad_output_shape)
+
+        graph = helper.make_graph(
+            [first_add_node, pad_node, second_add_node],
+            "TestPadWithEmptyStringInput",
+            [input_tensor],
+            [output_tensor],
+            initializer=[pads_initializer],
+        )
+        model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
+        model.ir_version = 7  # use stable onnx ir version
+
+        return model
+
+    def test_pad_with_empty_string_input_name(self):
+        np.random.seed(108)
+        model_fp32_path = "pad_with_empty_string_input_name_fp32.onnx"
+        model_i8_path = "pad_with_empty_string_input_name_i8.onnx"
+
+        shape = [
+            3,
+        ]
+        name = "input"
+        data_reader = self.input_feeds(
+            1,
+            {
+                name: shape,
+            },
+        )
+
+        model_fp32 = TestOpQuatizerPad.construct_model_add_pad_add(name=name, shape=shape, final_name="output")
+
+        onnx.save(model_fp32, model_fp32_path)
+
+        self.quantize_model(
+            model_fp32_path,
+            model_i8_path,
+            data_reader=data_reader,
+        )
+
+        model_i8 = onnx.load(model_i8_path)
+
+        # Assert quantization really happens.
+        self.assertEqual(model_i8.graph.node[0].op_type, "QuantizeLinear")
+        self.assertEqual(model_i8.graph.node[1].op_type, "QLinearAdd")
+        self.assertEqual(model_i8.graph.node[2].op_type, "Pad")
+        self.assertEqual(model_i8.graph.node[3].op_type, "QLinearAdd")
+        self.assertEqual(model_i8.graph.node[4].op_type, "DequantizeLinear")
+
+        for node in model_i8.graph.node:
+            # Examine no empty string flows to quantization process.
+            # Previously, optional input specified by `""` in NodeProto.input
+            # may cause phantom node to generate `"_quantized"` in quantization process.
+            for name in itertools.chain(node.input, node.output):
+                self.assertNotEqual(name, "")
+                self.assertNotEqual(name, "_quantized")
+
 
 if __name__ == "__main__":
     unittest.main()