StringConcat Operator

docs/Changelog.md

@@ -24022,6 +24022,37 @@ This version of the operator has been available since version 20 of the default
 <dd>Constrain grid types to float tensors.</dd>
 </dl>
 
+### <a name="StringConcat-20"></a>**StringConcat-20**</a>
+
+  StringConcat concatenates string tensors elementwise (with NumPy-style broadcasting support)
+
+#### Version
+
+This version of the operator has been available since version 20 of the default ONNX operator set.
+
+#### Inputs
+
+<dl>
+<dt><tt>X</tt> (non-differentiable) : T</dt>
+<dd>Tensor to prepend in concatenation</dd>
+<dt><tt>Y</tt> (non-differentiable) : T</dt>
+<dd>Tensor to append in concatenation</dd>
+</dl>
+
+#### Outputs
+
+<dl>
+<dt><tt>Z</tt> (non-differentiable) : T</dt>
+<dd>Concatenated string tensor</dd>
+</dl>
+
+#### Type Constraints
+
+<dl>
+<dt><tt>T</tt> : tensor(string)</dt>
+<dd>Inputs and outputs must be UTF-8 strings</dd>
+</dl>
+
 # ai.onnx.preview.training
 ## Version 1 of the 'ai.onnx.preview.training' operator set
 ### <a name="ai.onnx.preview.training.Adagrad-1"></a>**ai.onnx.preview.training.Adagrad-1**</a>

docs/Operators.md

@@ -146,6 +146,7 @@ For an operator input/output's differentiability, it can be differentiable,
 |<a href="#SplitToSequence">SplitToSequence</a>|<a href="Changelog.md#SplitToSequence-11">11</a>|
 |<a href="#Sqrt">Sqrt</a>|<a href="Changelog.md#Sqrt-13">13</a>, <a href="Changelog.md#Sqrt-6">6</a>, <a href="Changelog.md#Sqrt-1">1</a>|
 |<a href="#Squeeze">Squeeze</a>|<a href="Changelog.md#Squeeze-13">13</a>, <a href="Changelog.md#Squeeze-11">11</a>, <a href="Changelog.md#Squeeze-1">1</a>|
+|<a href="#StringConcat">StringConcat</a>|<a href="Changelog.md#StringConcat-20">20</a>|
 |<a href="#StringNormalizer">StringNormalizer</a>|<a href="Changelog.md#StringNormalizer-10">10</a>|
 |<a href="#Sub">Sub</a>|<a href="Changelog.md#Sub-14">14</a>, <a href="Changelog.md#Sub-13">13</a>, <a href="Changelog.md#Sub-7">7</a>, <a href="Changelog.md#Sub-6">6</a>, <a href="Changelog.md#Sub-1">1</a>|
 |<a href="#Sum">Sum</a>|<a href="Changelog.md#Sum-13">13</a>, <a href="Changelog.md#Sum-8">8</a>, <a href="Changelog.md#Sum-6">6</a>, <a href="Changelog.md#Sum-1">1</a>|
@@ -30080,6 +30081,81 @@ expect(node, inputs=[x, axes], outputs=[y], name="test_squeeze_negative_axes")
 </details>
 
 
+### <a name="StringConcat"></a><a name="stringconcat">**StringConcat**</a>
+
+  StringConcat concatenates string tensors elementwise (with NumPy-style broadcasting support)
+
+#### Version
+
+This version of the operator has been available since version 20 of the default ONNX operator set.
+
+#### Inputs
+
+<dl>
+<dt><tt>X</tt> (non-differentiable) : T</dt>
+<dd>Tensor to prepend in concatenation</dd>
+<dt><tt>Y</tt> (non-differentiable) : T</dt>
+<dd>Tensor to append in concatenation</dd>
+</dl>
+
+#### Outputs
+
+<dl>
+<dt><tt>Z</tt> (non-differentiable) : T</dt>
+<dd>Concatenated string tensor</dd>
+</dl>
+
+#### Type Constraints
+
+<dl>
+<dt><tt>T</tt> : tensor(string)</dt>
+<dd>Inputs and outputs must be UTF-8 strings</dd>
+</dl>
+
+
+#### Examples
+
+<details>
+<summary>stringconcat</summary>
+
+```python
+node = onnx.helper.make_node(
+    "StringConcat",
+    inputs=["x", "y"],
+    outputs=["result"],
+)
+x = np.array(["abc", "def"]).astype("object")
+y = np.array([".com", ".net"]).astype("object")
+result = np.array(["abc.com", "def.net"]).astype("object")
+
+expect(node, inputs=[x, y], outputs=[result], name="test_string_concat")
+
+x = np.array(["cat", "dog", "snake"]).astype("object")
+y = np.array(["s"]).astype("object")
+result = np.array(["cats", "dogs", "snakes"]).astype("object")
+
+expect(
+    node,
+    inputs=[x, y],
+    outputs=[result],
+    name="test_string_concat_broadcasting",
+)
+
+x = np.array("cat").astype("object")
+y = np.array("s").astype("object")
+result = np.array("cats").astype("object")
+
+expect(
+    node,
+    inputs=[x, y],
+    outputs=[result],
+    name="test_string_concat_zero_dimensional",
+)
+```
+
+</details>
+
+
 ### <a name="StringNormalizer"></a><a name="stringnormalizer">**StringNormalizer**</a>
 
   StringNormalization performs string operations for basic cleaning.

docs/TestCoverage.md

@@ -6,7 +6,7 @@
 * [Overall Test Coverage](#overall-test-coverage)
 # Node Test Coverage
 ## Summary
-Node tests have covered 174/187 (93.05%, 5 generators excluded) common operators.
+Node tests have covered 175/188 (93.09%, 5 generators excluded) common operators.
 
 Node tests have covered 0/0 (N/A) experimental operators.
 
@@ -20594,6 +20594,49 @@ expect(node, inputs=[x, axes], outputs=[y], name="test_squeeze_negative_axes")
 </details>
 
 
+### StringConcat
+There are 1 test cases, listed as following:
+<details>
+<summary>stringconcat</summary>
+
+```python
+node = onnx.helper.make_node(
+    "StringConcat",
+    inputs=["x", "y"],
+    outputs=["result"],
+)
+x = np.array(["abc", "def"]).astype("object")
+y = np.array([".com", ".net"]).astype("object")
+result = np.array(["abc.com", "def.net"]).astype("object")
+
+expect(node, inputs=[x, y], outputs=[result], name="test_string_concat")
+
+x = np.array(["cat", "dog", "snake"]).astype("object")
+y = np.array(["s"]).astype("object")
+result = np.array(["cats", "dogs", "snakes"]).astype("object")
+
+expect(
+    node,
+    inputs=[x, y],
+    outputs=[result],
+    name="test_string_concat_broadcasting",
+)
+
+x = np.array("cat").astype("object")
+y = np.array("s").astype("object")
+result = np.array("cats").astype("object")
+
+expect(
+    node,
+    inputs=[x, y],
+    outputs=[result],
+    name="test_string_concat_zero_dimensional",
+)
+```
+
+</details>
+
+
 ### StringNormalizer
 There are 6 test cases, listed as following:
 <details>

onnx/backend/test/case/node/string_concat.py

@@ -0,0 +1,46 @@
+# Copyright (c) ONNX Project Contributors
+#
+# SPDX-License-Identifier: Apache-2.0
+
+import numpy as np
+
+import onnx
+from onnx.backend.test.case.base import Base
+from onnx.backend.test.case.node import expect
+
+
+class StringConcat(Base):
+    @staticmethod
+    def export() -> None:
+        node = onnx.helper.make_node(
+            "StringConcat",
+            inputs=["x", "y"],
+            outputs=["result"],
+        )
+        x = np.array(["abc", "def"]).astype("object")
+        y = np.array([".com", ".net"]).astype("object")
+        result = np.array(["abc.com", "def.net"]).astype("object")
+
+        expect(node, inputs=[x, y], outputs=[result], name="test_string_concat")
+
+        x = np.array(["cat", "dog", "snake"]).astype("object")
+        y = np.array(["s"]).astype("object")
+        result = np.array(["cats", "dogs", "snakes"]).astype("object")
+
+        expect(
+            node,
+            inputs=[x, y],
+            outputs=[result],
+            name="test_string_concat_broadcasting",
+        )
+
+        x = np.array("cat").astype("object")
+        y = np.array("s").astype("object")
+        result = np.array("cats").astype("object")
+
+        expect(
+            node,
+            inputs=[x, y],
+            outputs=[result],
+            name="test_string_concat_zero_dimensional",
+        )

onnx/backend/test/data/node/test_string_concat/test_data_set_0/input_0.pb

@@ -0,0 +1 @@
+2abc2defB
x

onnx/backend/test/data/node/test_string_concat/test_data_set_0/input_1.pb

@@ -0,0 +1 @@
+2.com2.netB
y

onnx/backend/test/data/node/test_string_concat/test_data_set_0/output_0.pb

@@ -0,0 +1 @@
+2abc.com2def.netBresult

onnx/backend/test/data/node/test_string_concat_broadcasting/test_data_set_0/input_0.pb

@@ -0,0 +1 @@
+2cat2dog2snakeB
x

onnx/backend/test/data/node/test_string_concat_broadcasting/test_data_set_0/input_1.pb

@@ -0,0 +1 @@
+
2
sB
y

onnx/backend/test/data/node/test_string_concat_broadcasting/test_data_set_0/output_0.pb

@@ -0,0 +1 @@
+2cats2dogs2snakesBresult

onnx/backend/test/data/node/test_string_concat_zero_dimensional/test_data_set_0/input_0.pb

@@ -0,0 +1 @@
+2catB
x

onnx/backend/test/data/node/test_string_concat_zero_dimensional/test_data_set_0/input_1.pb

@@ -0,0 +1 @@
+2
sB
y

onnx/backend/test/data/node/test_string_concat_zero_dimensional/test_data_set_0/output_0.pb

@@ -0,0 +1 @@
+2catsBresult

onnx/defs/nn/defs.cc

@@ -2229,75 +2229,6 @@ ONNX_OPERATOR_SET_SCHEMA(
         })
         .SetDoc(TfIdfVectorizer_ver9_doc));
 
-static const char* StringNormalizer_ver10_doc = R"DOC(
-StringNormalization performs string operations for basic cleaning.
-This operator has only one input (denoted by X) and only one output
-(denoted by Y). This operator first examines the elements in the X,
-and removes elements specified in "stopwords" attribute.
-After removing stop words, the intermediate result can be further lowercased,
-uppercased, or just returned depending the "case_change_action" attribute.
-This operator only accepts [C]- and [1, C]-tensor.
-If all elements in X are dropped, the output will be the empty value of string tensor with shape [1]
-if input shape is [C] and shape [1, 1] if input shape is [1, C].
-)DOC";
-
-ONNX_OPERATOR_SET_SCHEMA(
-    StringNormalizer,
-    10,
-    OpSchema()
-        .Input(0, "X", "UTF-8 strings to normalize", "tensor(string)")
-        .Output(0, "Y", "UTF-8 Normalized strings", "tensor(string)")
-        .Attr(
-            std::string("case_change_action"),
-            std::string("string enum that cases output to be lowercased/uppercases/unchanged."
-                        " Valid values are \"LOWER\", \"UPPER\", \"NONE\". Default is \"NONE\""),
-            AttributeProto::STRING,
-            std::string("NONE"))
-        .Attr(
-            std::string("is_case_sensitive"),
-            std::string("Boolean. Whether the identification of stop words in X is case-sensitive. Default is false"),
-            AttributeProto::INT,
-            static_cast<int64_t>(0))
-        .Attr(
-            "stopwords",
-            "List of stop words. If not set, no word would be removed from X.",
-            AttributeProto::STRINGS,
-            OPTIONAL_VALUE)
-        .Attr(
-            "locale",
-            "Environment dependent string that denotes the locale according to which output strings needs to be upper/lowercased."
-            "Default en_US or platform specific equivalent as decided by the implementation.",
-            AttributeProto::STRING,
-            OPTIONAL_VALUE)
-        .SetDoc(StringNormalizer_ver10_doc)
-        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
-          auto output_elem_type = ctx.getOutputType(0)->mutable_tensor_type();
-          output_elem_type->set_elem_type(TensorProto::STRING);
-          if (!hasInputShape(ctx, 0)) {
-            return;
-          }
-          TensorShapeProto output_shape;
-          auto& input_shape = ctx.getInputType(0)->tensor_type().shape();
-          auto dim_size = input_shape.dim_size();
-          // Last axis dimension is unknown if we have stop-words since we do
-          // not know how many stop-words are dropped
-          if (dim_size == 1) {
-            // Unknown output dimension
-            output_shape.add_dim();
-          } else if (dim_size == 2) {
-            // Copy B-dim
-            auto& b_dim = input_shape.dim(0);
-            if (!b_dim.has_dim_value() || b_dim.dim_value() != 1) {
-              fail_shape_inference("Input shape must have either [C] or [1,C] dimensions where C > 0");
-            }
-            *output_shape.add_dim() = b_dim;
-            output_shape.add_dim();
-          } else {
-            fail_shape_inference("Input shape must have either [C] or [1,C] dimensions where C > 0");
-          }
-          updateOutputShape(ctx, 0, output_shape);
-        }));
-
 static const char* mvn_ver13_doc = R"DOC(
       A MeanVarianceNormalization Function: Perform mean variance normalization
       on the input tensor X using formula: `(X-EX)/sqrt(E(X-EX)^2)`

onnx/defs/operator_sets.h

@@ -1105,6 +1105,7 @@ class OpSet_Onnx_ver19 {
 class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, GridSample);
 class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, Gelu);
 class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, ConstantOfShape);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, StringConcat);
 
 // Iterate over schema from ai.onnx version 20
 class OpSet_Onnx_ver20 {
@@ -1113,6 +1114,7 @@ class OpSet_Onnx_ver20 {
     fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, GridSample)>());
     fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, Gelu)>());
     fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, ConstantOfShape)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, StringConcat)>());
   }
 };