Add parsedatetime op

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

@@ -0,0 +1,68 @@
+# Copyright (c) ONNX Project Contributors
+#
+# SPDX-License-Identifier: Apache-2.0
+from datetime import datetime
+
+import numpy as np
+
+import onnx
+from onnx import numpy_helper
+from onnx.backend.test.case.base import Base
+from onnx.backend.test.case.node import expect
+
+
+class ParseDateTime(Base):
+    @staticmethod
+    def export_float_nan_default() -> None:
+        fmt = "%d/%m/%y %H:%M"
+        default = float("NaN")
+        node = onnx.helper.make_node(
+            "ParseDateTime",
+            inputs=["x"],
+            outputs=["y"],
+            format=fmt,
+            unit="s",
+            default=onnx.helper.make_tensor(
+                name="default",
+                data_type=onnx.TensorProto.DOUBLE,
+                dims=[],
+                vals=np.array(default),
+            ),
+        )
+        x = np.array(["21/11/06 16:30", "foobar"], dtype=object)
+        y = []
+        for s in x:
+            try:
+                # datetime.timestamp() returns a float
+                y.append(datetime.strptime(s, fmt).timestamp())
+            except ValueError:
+                y.append(default)
+        expect(node, inputs=[x], outputs=[np.array(y)], name="test_parsedatetime")
+
+    @staticmethod
+    def export_int_default() -> None:
+        fmt = "%d/%m/%y %H:%M"
+        default = np.iinfo(np.int64).min
+        node = onnx.helper.make_node(
+            "ParseDateTime",
+            inputs=["x"],
+            outputs=["y"],
+            format=fmt,
+            unit="s",
+            default=onnx.helper.make_tensor(
+                name="default",
+                data_type=onnx.TensorProto.INT64,
+                dims=[],
+                vals=np.array(default),
+            ),
+        )
+        x = np.array(["21/11/06 16:30", "foobar"], dtype=object)
+        y = []
+        for s in x:
+            try:
+                y.append(datetime.strptime(s, fmt).timestamp())
+            except ValueError:
+                y.append(default)
+        expect(
+            node, inputs=[x], outputs=[np.array(y, np.int64)], name="test_parsedatetime"
+        )

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

@@ -0,0 +1 @@
+221/11/06 16:302foobarB
x

onnx/defs/operator_sets.h

@@ -1102,18 +1102,20 @@ class OpSet_Onnx_ver19 {
 };
 
 // Forward declarations for ai.onnx version 20
-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, Gelu);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, GridSample);
+class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, ParseDateTime);
 class ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, StringConcat);
 
 // Iterate over schema from ai.onnx version 20
 class OpSet_Onnx_ver20 {
  public:
   static void ForEachSchema(std::function<void(OpSchema&&)> fn) {
-    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, Gelu)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, GridSample)>());
+    fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, ParseDateTime)>());
     fn(GetOpSchema<ONNX_OPERATOR_SET_SCHEMA_CLASS_NAME(Onnx, 20, StringConcat)>());
   }
 };

onnx/defs/text/defs.cc

@@ -5,6 +5,44 @@
 #include "onnx/defs/schema.h"
 
 namespace ONNX_NAMESPACE {
+static const char* ParseDateTime_doc = R"DOC(Parse a datetime string into a (floating point) Unix time stamp.)DOC";
+ONNX_OPERATOR_SET_SCHEMA(
+    ParseDateTime,
+    20,
+    OpSchema()
+        .Input(0, "X", "Tensor with datetime strings", "T1", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
+        .Output(0, "y", "Unix time stamps", "T2", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
+        .Attr("format", "Format description in the syntax of C's `strptime`.", AttributeProto::STRING)
+        .Attr(
+            "unit",
+            "Unit of the returned time stamp. Allowed values are: 's' (second), 'ms' (millisecond), 'us' (microsecond) or 'ns' (nanosecond).",
+            AttributeProto::STRING)
+        .Attr(
+            "default",
+            "Default value to be used if the parsing fails. The tensor must be of rank 0 and either of type `tensor(int64)` or `tensor(double)`. The tensor type is the output type. If 'default' is specified, the output type is `tensor(int64)` and the behavior for failing to parse an input element is implementation defined.",
+            AttributeProto::TENSOR,
+            OPTIONAL_VALUE)
+
+        .TypeConstraint("T1", {"tensor(string)"}, "UTF-8 datetime strings")
+        .TypeConstraint("T2", {"tensor(double)", "tensor(int64)"}, "Output type depends on 'default' attribute.")
+        .SetDoc(ParseDateTime_doc)
+        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
+          auto* default_value = ctx.getAttribute("default");
+
+          if (hasInputShape(ctx, 0)) {
+            propagateShapeFromInputToOutput(ctx, 0, 0);
+          }
+
+          if (nullptr == default_value) {
+            updateOutputElemType(ctx, 0, TensorProto::INT64);
+            return;
+          } else {
+            const TensorProto& tensor_proto = default_value->t();
+            updateOutputElemType(ctx, 0, tensor_proto.data_type());
+            return;
+          }
+        }));
+
 static const char* StringConcat_doc =
     R"DOC(StringConcat concatenates string tensors elementwise (with NumPy-style broadcasting support))DOC";
 ONNX_OPERATOR_SET_SCHEMA(

onnx/reference/ops/_op_list.py

@@ -148,6 +148,7 @@
 from onnx.reference.ops.op_optional_has_element import OptionalHasElement
 from onnx.reference.ops.op_or import Or
 from onnx.reference.ops.op_pad import Pad_1, Pad_2, Pad_11, Pad_18
+from onnx.reference.ops.op_parsedatetime import ParseDateTime
 from onnx.reference.ops.op_pow import Pow
 from onnx.reference.ops.op_prelu import PRelu
 from onnx.reference.ops.op_qlinear_conv import QLinearConv

onnx/reference/ops/op_parsedatetime.py

@@ -0,0 +1,22 @@
+# Copyright (c) ONNX Project Contributors
+
+# SPDX-License-Identifier: Apache-2.0
+# pylint: disable=W0221
+from datetime import datetime
+
+import numpy as np
+
+from onnx.reference.op_run import OpRun
+
+
+class ParseDateTime(OpRun):
+    def _run(self, x, format, unit, default=None):  # type: ignore
+        def parse(el):
+            try:
+                return datetime.strptime(el, format).timestamp()
+            except ValueError:
+                return np.nan
+        out = np.array([parse(el) for el in x])
+        out[np.isnan(out)] = default
+        out = out.astype(default.dtype) if default is not None else out.astype(np.int64)
+        return (out,)