MLPY/Lib/site-packages/onnx/defs/tensor/defs.cc

/*
 * SPDX-License-Identifier: Apache-2.0
 */

#include <algorithm>
#include <cmath>
#include <numeric>

#include "onnx/defs/data_propagators.h"
#include "onnx/defs/function.h"
#include "onnx/defs/tensor/utils.h"
#include "onnx/defs/tensor_proto_util.h"

namespace ONNX_NAMESPACE {

static const char* Cast_ver19_doc = R"DOC(
The operator casts the elements of a given input tensor to a data type
specified by the 'to' argument and returns an output tensor of the same size in
the converted type. The 'to' argument must be one of the data types specified
in the 'DataType' enum field in the TensorProto message.

Casting from string tensor in plain (e.g., "3.14" and "1000") and scientific numeric representations
(e.g., "1e-5" and "1E8") to float types is supported. For example, converting string "100.5" to an integer may
yield result 100. There are some string literals reserved for special floating-point values;
"+INF" (and "INF"), "-INF", and "NaN" are positive infinity, negative infinity, and not-a-number, respectively.
Any string which can exactly match "+INF" in a case-insensitive way would be mapped to positive infinite. Similarly,
this case-insensitive rule is applied to "INF" and "NaN". When casting from numeric tensors
to string tensors, plain floating-point representation (such as "314.15926") would be used.
Converting non-numerical-literal string such as "Hello World!" is an undefined behavior. Cases
of converting string representing floating-point arithmetic value, such as "2.718", to INT is an undefined behavior.

Conversion from a numerical type to any numerical type is always allowed.
User must be aware of precision loss and value change caused by range difference between two types.
For example, a 64-bit float 3.1415926459 may be round to a 32-bit float 3.141592. Similarly, converting
an integer 36 to Boolean may produce 1 because we truncate bits which can't be stored in the targeted type.

In more detail, the conversion among numerical types should follow these rules
if the destination type is not a float 8 type.

* Casting from floating point to:
  * floating point: +/- infinity if OOR (out of range).
  * fixed point: undefined if OOR.
  * bool: +/- 0.0 to False; all else to True.
* Casting from fixed point to:
  * floating point: +/- infinity if OOR. (+ infinity in the case of uint)
  * fixed point: when OOR, discard higher bits and reinterpret (with respect to two's complement representation for
    signed types). For example, 200 (int16) -> -56 (int8).
  * bool: zero to False; nonzero to True.
* Casting from bool to:
  * floating point: `{1.0, 0.0}`.
  * fixed point: `{1, 0}`.
  * bool: no change.

Float 8 type were introduced to speed up the training of
deep models. By default the conversion of a float *x* obeys
to the following rules. `[x]` means the value rounded to
the target mantissa width.

| x | E4M3FN | E4M3FNUZ | E5M2 | E5M2FNUZ |
|------|----|----|----|----|
| 0 | 0 | 0 | 0 | 0 |
|-0 | -0 | 0 | -0 | 0 |
| NaN | NaN | NaN | NaN | NaN |
| +/- Inf | +/- FLT_MAX | NaN | FLT_MAX | NaN |
| [x] > FLT_MAX | FLT_MAX | FLT_MAX | FLT_MAX | FLT_MAX |
| [x] < -FLT_MAX | -FLT_MAX | -FLT_MAX | -FLT_MAX | -FLT_MAX |
| else | RNE | RNE | RNE | RNE |

The behavior changes if the parameter 'saturate' is set to False.
The rules then become:

| x | E4M3FN | E4M3FNUZ | E5M2 | E5M2FNUZ |
|------|----|----|----|----|
| 0 | 0 | 0 | 0 | 0 |
|-0 | -0 | 0 | -0 | 0 |
| NaN | NaN | NaN | NaN | NaN |
| +/- Inf | NaN | NaN | +/- Inf | NaN |
| [x] > FLT_MAX | NaN | NaN | Inf | NaN |
| [x] < -FLT_MAX | NaN | NaN | -Inf | NaN |
| else | RNE | RNE | RNE | RNE |
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Cast,
    21,
    OpSchema()
        .SetDoc(Cast_ver19_doc)
        .Attr(
            "to",
            "The data type to which the elements of the input tensor are cast. "
            "Strictly must be one of the types from DataType enum in TensorProto",
            AttributeProto::INT)
        .Attr(
            "saturate",
            "The parameter defines how the conversion behaves if an input value is out of "
            "range of the destination type. It only applies for float 8 conversion "
            "(float8e4m3fn, float8e4m3fnuz, float8e5m2, float8e5m2fnuz). It is true by default. "
            "All cases are fully described in two tables inserted in the operator description.",
            AttributeProto::INT,
            static_cast<int64_t>(1))
        .Input(0, "input", "Input tensor to be cast.", "T1", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Output(
            0,
            "output",
            "Output tensor with the same shape as input with type "
            "specified by the 'to' argument",
            "T2",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint(
            "T1",
            {"tensor(float16)",    "tensor(float)",          "tensor(double)",       "tensor(int8)",
             "tensor(int16)",      "tensor(int32)",          "tensor(int64)",        "tensor(uint8)",
             "tensor(uint16)",     "tensor(uint32)",         "tensor(uint64)",       "tensor(bool)",
             "tensor(string)",     "tensor(bfloat16)",       "tensor(float8e4m3fn)", "tensor(float8e4m3fnuz)",
             "tensor(float8e5m2)", "tensor(float8e5m2fnuz)", "tensor(uint4)",        "tensor(int4)"},
            "Constrain input types. Casting from complex is not supported.")
        .TypeConstraint(
            "T2",
            {"tensor(float16)",    "tensor(float)",          "tensor(double)",       "tensor(int8)",
             "tensor(int16)",      "tensor(int32)",          "tensor(int64)",        "tensor(uint8)",
             "tensor(uint16)",     "tensor(uint32)",         "tensor(uint64)",       "tensor(bool)",
             "tensor(string)",     "tensor(bfloat16)",       "tensor(float8e4m3fn)", "tensor(float8e4m3fnuz)",
             "tensor(float8e5m2)", "tensor(float8e5m2fnuz)", "tensor(uint4)",        "tensor(int4)"},
            "Constrain output types. Casting to complex is not supported.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromAttributeToOutput(ctx, "to", 0);
          if (hasNInputShapes(ctx, 1)) {
            propagateShapeFromInputToOutput(ctx, 0, 0);
          }
        })
        .PartialDataPropagationFunction([](DataPropagationContext& ctx) {
          PropagateShapeDataFromInputToOutput(ctx, 0);
        }));

static const char* CastLike_ver19_doc = R"DOC(
The operator casts the elements of a given input tensor (the first input) to
the same data type as the elements of the second input tensor.
See documentation of the Cast operator for further details.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    CastLike,
    21,
    OpSchema()
        .SetDoc(CastLike_ver19_doc)
        .Attr(
            "saturate",
            "The parameter defines how the conversion behaves if an input value is out of "
            "range of the destination type. It only applies for float 8 conversion "
            "(float8e4m3fn, float8e4m3fnuz, float8e5m2, float8e5m2fnuz). It is true by default. "
            "Please refer to operator Cast description for further details.",
            AttributeProto::INT,
            static_cast<int64_t>(1))
        .Input(0, "input", "Input tensor to be cast.", "T1", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "target_type",
            "The (first) input tensor will be cast to produce a tensor of the same type as this (second input) tensor.",
            "T2",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "output",
            "Output tensor produced by casting the first input tensor to have the same type as the second input tensor.",
            "T2",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint(
            "T1",
            {"tensor(float16)",    "tensor(float)",          "tensor(double)",       "tensor(int8)",
             "tensor(int16)",      "tensor(int32)",          "tensor(int64)",        "tensor(uint8)",
             "tensor(uint16)",     "tensor(uint32)",         "tensor(uint64)",       "tensor(bool)",
             "tensor(string)",     "tensor(bfloat16)",       "tensor(float8e4m3fn)", "tensor(float8e4m3fnuz)",
             "tensor(float8e5m2)", "tensor(float8e5m2fnuz)", "tensor(uint4)",        "tensor(int4)"},
            "Constrain input types. Casting from complex is not supported.")
        .TypeConstraint(
            "T2",
            {"tensor(float16)",    "tensor(float)",          "tensor(double)",       "tensor(int8)",
             "tensor(int16)",      "tensor(int32)",          "tensor(int64)",        "tensor(uint8)",
             "tensor(uint16)",     "tensor(uint32)",         "tensor(uint64)",       "tensor(bool)",
             "tensor(string)",     "tensor(bfloat16)",       "tensor(float8e4m3fn)", "tensor(float8e4m3fnuz)",
             "tensor(float8e5m2)", "tensor(float8e5m2fnuz)", "tensor(uint4)",        "tensor(int4)"},
            "Constrain output types. Casting to complex is not supported.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 1, 0);
          if (hasNInputShapes(ctx, 1)) {
            propagateShapeFromInputToOutput(ctx, 0, 0);
          }
        })
        .SetContextDependentFunctionBodyBuilder(
            [](const FunctionBodyBuildContext& ctx, const OpSchema& schema, FunctionProto& functionProto) -> bool {
              auto target_type = ctx.getInputType(1);
              if ((target_type == nullptr) || (!target_type->has_tensor_type())) {
                // we cannot create a correct function body without knowing the target element type
                return false;
              }
              auto target_elt_type = target_type->tensor_type().elem_type();
              FunctionBuilder builder(functionProto);
              builder.Add(
                  MakeString("output = Cast <to= ", (int64_t)(target_elt_type), ", saturate: int = @saturate> (input)")
                      .c_str());
              schema.BuildFunction(functionProto);
              return true;
            }));

static const char* Reshape_ver19_doc = R"DOC(
Reshape the input tensor similar to numpy.reshape.
First input is the data tensor, second input is a shape tensor which specifies the output shape. It outputs the reshaped tensor.
At most one dimension of the new shape can be -1. In this case, the value is
inferred from the size of the tensor and the remaining dimensions. A dimension
could also be 0, in which case the actual dimension value is unchanged (i.e. taken
from the input tensor). If 'allowzero' is set, and the new shape includes 0, the
dimension will be set explicitly to zero (i.e. not taken from input tensor).
Shape (second input) could be an empty shape, which means converting to a scalar.
The input tensor's shape and the output tensor's shape are required to have the same number of elements.

If the attribute 'allowzero' is set, it is invalid for the specified shape to
contain both a zero value and -1, as the value of the dimension corresponding
to -1 cannot be determined uniquely.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Reshape,
    21,
    OpSchema()
        .SetDoc(Reshape_ver19_doc)
        .Attr(
            "allowzero",
            "(Optional) By default, when any value in the 'shape' input is equal to zero "
            "the corresponding dimension value is copied from the input tensor dynamically. "
            "allowzero=1 indicates that if any value in the 'shape' input is set to zero, "
            "the zero value is honored, similar to NumPy.",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Input(0, "data", "An input tensor.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "shape",
            "Specified shape for output.",
            "tensor(int64)",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(0, "reshaped", "Reshaped data.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir10(), "Constrain input and output types to all tensor types.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          // Type inference
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          bool found;
          TensorShapeProto targetShapeProto = getShapeInput(ctx, 1, found);
          if (!found) {
            return;
          }

          int allowzero = static_cast<int>(getAttribute(ctx, "allowzero", 0));

          // Iterate through targetShape, adding dimensions in the outputShape
          // TensorProto. If the targetShape dimension is -1, we do not set the
          // dimension value in this iteration, but we record the Dimension. If
          // targetShape dimension is 0, we attempt to propagate the dimension
          // value/param. If the value cannot be inferred, we set the flag in
          // the unresolveZeros vector. If targetShape dimension is positive, we
          // set the dimension value in the outputShape. We track the product of
          // the dimensions we are setting outputShape in the outputProduct
          // variable. The outputProduct will potentially be used for inferring
          // a dimension marked -1.
          auto* outputShape = ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();
          TensorShapeProto::Dimension* negativeOneDim = nullptr;
          const auto& dataInputTensorType = ctx.getInputType(0)->tensor_type();
          std::vector<bool> unresolvedZeros(targetShapeProto.dim_size(), false);
          int64_t outputProduct = 1;
          bool outputProductValid = true;
          for (int i = 0; i < static_cast<int>(targetShapeProto.dim_size()); ++i) {
            // Add a new dimension to outputShape
            auto* new_dim = outputShape->add_dim();
            if (targetShapeProto.dim(i).has_dim_param()) {
              // There is a tricky edge case here. It is possible that the value of
              // symbolic dim can be -1 or 0 at runtime. In that case simply propgating this
              // symbol can be erroneous. This should be a very rare scenario and in such a
              // case an option is to turn off data propagation during shape inference.
              new_dim->set_dim_param(targetShapeProto.dim(i).dim_param());
              outputProductValid = false;
            } else {
              if (!targetShapeProto.dim(i).has_dim_value()) {
                outputProductValid = false;
                // treat this dim as unknown dim
                continue;
              }

              const auto dim_value = targetShapeProto.dim(i).dim_value();

              if (dim_value == -1) {
                // Check if multiple -1's. If not, set negativeOneDim, marking
                // this dimension to potentially be filled in later.
                if (negativeOneDim) {
                  fail_shape_inference("Target shape may not have multiple -1 dimensions.");
                }
                negativeOneDim = new_dim;
              } else if (dim_value == 0) {
                // Check if data input has a shape and if the index i is within
                // its bounds. If these conditions are satisfied, any dimension
                // value/param should be propagated. If dimension value cannot be
                // inferred, set the corresponding  unresolvedZeros flag to true.
                // If allowzero is set however, do not propagate values, since output
                // dimension is explicitly zero.
                if (allowzero == 0) {
                  unresolvedZeros[i] = true;
                  if (dataInputTensorType.has_shape()) {
                    if (i >= dataInputTensorType.shape().dim_size()) {
                      fail_shape_inference("Invalid position of 0.");
                    }
                    if (dataInputTensorType.shape().dim(i).has_dim_value()) {
                      const auto& input_dim_value = dataInputTensorType.shape().dim(i).dim_value();
                      new_dim->set_dim_value(input_dim_value);
                      outputProduct *= input_dim_value;
                      unresolvedZeros[i] = false;
                    } else if (dataInputTensorType.shape().dim(i).has_dim_param()) {
                      new_dim->set_dim_param(dataInputTensorType.shape().dim(i).dim_param());
                    }
                  }
                } else {
                  new_dim->set_dim_value(dim_value);
                  outputProduct *= dim_value;
                }
              } else if (dim_value > 0) {
                // Set the dimension value to dim_value
                new_dim->set_dim_value(dim_value);
                outputProduct *= dim_value;
              } else {
                // Check if value is less than -1; fail if so
                fail_shape_inference("Invalid dimension value: ", dim_value);
              }
            }
          }
          // If negativeOneDim has been set, we attempt to infer its value. This
          // can be done if all dimension values for the data input tensor shape
          // are known other than the ones corresponding to unresolvedZeros
          // flags.
          if (negativeOneDim && outputProductValid) {
            // First, attempt to compute product of data input shape dimensions
            // that are not marked by unresolvedZeros. If not possible, set the
            // inputProductValid flag to false.
            if (!outputProduct) {
              fail_shape_inference("Invalid Target shape product of 0. Product cannot be 0 in combination with -1");
            }
            int64_t inputProduct = 1;
            bool inputProductValid = true;
            if (!dataInputTensorType.has_shape()) {
              inputProductValid = false;
            } else {
              for (int i = 0; i < dataInputTensorType.shape().dim_size(); ++i) {
                if (dataInputTensorType.shape().dim(i).has_dim_value()) {
                  inputProduct *= dataInputTensorType.shape().dim(i).dim_value();
                } else if (i >= static_cast<int>(unresolvedZeros.size()) || !unresolvedZeros[i]) {
                  inputProductValid = false;
                  break;
                }
              }
            }
            if (inputProductValid) {
              if (inputProduct % outputProduct != 0) {
                fail_shape_inference("Dimension could not be inferred: incompatible shapes");
              }
              negativeOneDim->set_dim_value(inputProduct / outputProduct);
            }
          }
        }));

static const char* Shape_ver19_doc = R"DOC(
Takes a tensor as input and outputs an 1D int64 tensor containing the shape of the input tensor.
Optional attributes start and end can be used to compute a slice of the input tensor's shape.
If start axis is omitted, the slice starts from axis 0.
The end axis, if specified, is exclusive (and the returned value will not include the size of that axis).
If the end axis is omitted, the axes upto the last one will be included.
Negative axes indicate counting back from the last axis.
Note that axes will be clamped to the range [0, r-1], where r is the
rank of the input tensor if they are out-of-range (after adding r in the case of
negative axis). Thus, specifying any end value > r is equivalent to specifying an end
value of r, and specifying any start value < -r is equivalent to specifying a start
value of 0.

Examples:

```
Input tensor with shape: [2, 3, 4]
No attributes specified.
Output: [2, 3, 4]
```

```
Input tensor with shape: [2, 3, 4]
start: -1
Output: [4]
```

```
Input tensor with shape: [2, 3, 4]
end: -1
Output: [2, 3]
```

```
Input tensor with shape: [2, 3, 4]
start: 1
end: 2
Output: [3]
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Shape,
    21,
    OpSchema()
        .SetDoc(Shape_ver19_doc)
        .Input(0, "data", "An input tensor.", "T", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
        .Output(0, "shape", "Shape of the input tensor", "T1", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
        .Attr(
            "start",
            "(Optional) Starting axis for slicing the shape. Default value is 0."
            "Negative value means counting dimensions from the back.",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Attr(
            "end",
            "(Optional) Ending axis for slicing the shape. "
            "Negative value means counting dimensions from the back. "
            "If omitted, sizes of all axes upto (including) the last one will be included.",
            AttributeProto::INT,
            OPTIONAL_VALUE)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir10(), "Input tensor can be of arbitrary type.")
        .TypeConstraint("T1", {"tensor(int64)"}, "Constrain output to int64 tensor.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          ctx.getOutputType(0)->mutable_tensor_type()->set_elem_type(TensorProto::INT64);
          auto* output_shape = ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();
          auto* output_length = output_shape->add_dim();

          if (!hasNInputShapes(ctx, 1)) {
            return;
          }

          int64_t rank = static_cast<int64_t>(ctx.getInputType(0)->tensor_type().shape().dim_size());
          int64_t start = getAttribute(ctx, "start", 0);
          if (start < 0)
            start += rank;
          start = (start < 0) ? 0 : (start > rank) ? rank : start;
          int64_t end = getAttribute(ctx, "end", rank);
          if (end < 0)
            end += rank;
          end = (end < 0) ? 0 : (end > rank) ? rank : end;
          output_length->set_dim_value((end - start) < 0 ? 0 : (end - start));
        })
        .PartialDataPropagationFunction([](DataPropagationContext& ctx) {
          if (hasInputShape(ctx, 0)) {
            auto& input_shape = ctx.getInputType(0)->tensor_type().shape();
            int64_t rank = static_cast<int64_t>(input_shape.dim_size());
            int64_t start = getAttribute(ctx, "start", 0);
            if (start < 0)
              start += rank;
            start = (start < 0) ? 0 : (start > rank) ? rank : start;
            int64_t end = getAttribute(ctx, "end", rank);
            if (end < 0)
              end += rank;
            end = (end < 0) ? 0 : (end > rank) ? rank : end;
            TensorShapeProto output_shape;
            for (int64_t d = start; d < end; ++d) {
              *output_shape.add_dim() = input_shape.dim(static_cast<int>(d));
            }
            ctx.addOutputData(0, std::move(output_shape));
          }
        }));

static const char* Size_ver19_doc = R"DOC(
Takes a tensor as input and outputs a int64 scalar that equals to the total number of elements of the input tensor.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Size,
    21,
    OpSchema()
        .SetDoc(Size_ver19_doc)
        .Input(0, "data", "An input tensor.", "T", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
        .Output(
            0,
            "size",
            "Total number of elements of the input tensor",
            "T1",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir10(), "Input tensor can be of arbitrary type.")
        .TypeConstraint("T1", {"tensor(int64)"}, "Constrain output to int64 tensor, which should be a scalar though.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          ctx.getOutputType(0)->mutable_tensor_type()->set_elem_type(TensorProto::INT64);
          ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();
        })
        .PartialDataPropagationFunction([](DataPropagationContext& ctx) {
          const auto input_data = ctx.getInputData(0);
          if (input_data != nullptr) {
            TensorShapeProto tsp;
            tsp.mutable_dim()->Add()->set_dim_value(input_data->dim_size());
            ctx.addOutputData(0, std::move(tsp));
          }
        }));

ONNX_OPERATOR_SET_SCHEMA(
    Concat,
    13,
    OpSchema()
        .Attr(
            "axis",
            "Which axis to concat on. A negative value means counting dimensions from the back. "
            "Accepted range is [-r, r-1] where r = rank(inputs)..",
            AttributeProto::INT)
        .SetDoc(
            "Concatenate a list of tensors into a single tensor. "
            "All input tensors must have the same shape, except for the dimension size of the axis to concatenate on.")
        .Input(
            0,
            "inputs",
            "List of tensors for concatenation",
            "T",
            OpSchema::Variadic,
            true,
            1,
            OpSchema::Differentiable)
        .Output(0, "concat_result", "Concatenated tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain output types to any tensor type.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          auto numInputs = ctx.getNumInputs();
          if (numInputs < 1 || !hasNInputShapes(ctx, static_cast<int>(numInputs))) {
            return;
          }

          auto rank = ctx.getInputType(0)->tensor_type().shape().dim_size();

          auto axisAttr = ctx.getAttribute("axis");
          if (!axisAttr) {
            fail_shape_inference("Required attribute axis is missing");
          }
          int axis = static_cast<int>(axisAttr->i());
          if (axis < -rank || axis >= rank) {
            fail_shape_inference("axis must be in [-rank, rank-1].");
          }
          if (axis < 0) {
            axis += rank;
          }

          if (numInputs == 1) {
            propagateShapeFromInputToOutput(ctx, 0, 0);
            return;
          }

          bool all_lengths_known = true;
          int total_length = 0;

          auto* output_shape = ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();

          for (int64_t i = 0; i < rank; ++i) {
            output_shape->add_dim();
          }

          for (size_t i = 0; i < numInputs; i++) {
            const auto& shape = ctx.getInputType(i)->tensor_type().shape();
            if (shape.dim_size() != rank) {
              fail_shape_inference(
                  "All inputs to Concat must have same rank. Input ", i, " has rank ", shape.dim_size(), " != ", rank);
            }
            for (int j = 0; j < rank; j++) {
              if (j == axis) {
                if (shape.dim(j).has_dim_value()) {
                  total_length += static_cast<int>(shape.dim(j).dim_value());
                } else {
                  all_lengths_known = false;
                }
              } else {
                auto& output_dim = *output_shape->mutable_dim(j);
                const auto& input_dim = shape.dim(j);
                mergeInDimensionInfo(input_dim, output_dim, j);
              }
            }
          }

          if (all_lengths_known) {
            output_shape->mutable_dim(axis)->set_dim_value(total_length);
          }
        })
        .PartialDataPropagationFunction([](DataPropagationContext& ctx) {
          if (!axisIsZero(ctx)) {
            return;
          }
          TensorShapeProto tsp;
          for (size_t i = 0; i < ctx.getNumInputs(); ++i) {
            const auto input_data = ctx.getInputData(i);
            if (input_data == nullptr) {
              return;
            }
            for (int j = 0; j < input_data->dim_size(); ++j) {
              *tsp.add_dim() = input_data->dim(j);
            }
          }
          if (tsp.dim_size() > 0) {
            ctx.addOutputData(0, std::move(tsp));
          }
        }));

static const char* Split_ver18_doc =
    R"DOC(Split a tensor into a list of tensors, along the specified 'axis'.
Either input 'split' or the attribute 'num_outputs' should be specified, but not both.
If the attribute 'num_outputs' is specified, then the tensor is split into equal sized parts.
If the tensor is not evenly splittable into `num_outputs`, the last chunk will be smaller.
If the input 'split' is specified, it indicates the sizes of each output in the split.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Split,
    18,
    OpSchema()
        .Input(0, "input", "The tensor to split", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "split",
            "Optional length of each output. Values should be >= 0."
            "Sum of the values must be equal to the dim value at 'axis' specified.",
            "tensor(int64)",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "outputs",
            "One or more outputs forming list of tensors after splitting",
            "T",
            OpSchema::Variadic,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to all tensor types.")
        .Attr(
            "axis",
            "Which axis to split on. "
            "A negative value means counting dimensions from the back. Accepted range is [-rank, rank-1] "
            "where r = rank(input).",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Attr(
            "num_outputs",
            "Number of outputs to split parts of the tensor into. "
            "If the tensor is not evenly splittable the last chunk will be smaller.",
            AttributeProto::INT,
            false)
        .SetDoc(Split_ver18_doc)
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          for (int i = 0; i < static_cast<int>(ctx.getNumOutputs()); ++i) {
            propagateElemTypeFromInputToOutput(ctx, 0, i);
          }
          if (!hasNInputShapes(ctx, 1)) {
            return;
          }
          const auto& shape = ctx.getInputType(0)->tensor_type().shape();
          int rank = shape.dim_size();
          int axis = static_cast<int>(getAttribute(ctx, "axis", 0));
          if (axis < -rank || axis >= rank) {
            fail_type_inference("Invalid value of attribute 'axis'. Rank=", rank, " Value=", axis);
          }
          if (axis < 0) {
            axis += rank;
          }
          const auto& split_dim = shape.dim(axis);
          if (!split_dim.has_dim_value()) {
            for (size_t i = 0; i < ctx.getNumOutputs(); i++) {
              *ctx.getOutputType(i)->mutable_tensor_type()->mutable_shape() = shape;
              ctx.getOutputType(i)->mutable_tensor_type()->mutable_shape()->mutable_dim(axis)->Clear();
            }
            return;
          }
          int split_dim_value = static_cast<int>(split_dim.dim_value());

          std::vector<int64_t> split;
          const auto num_outputs_attr = ctx.getAttribute("num_outputs");
          if (ctx.hasInput(1) && num_outputs_attr) {
            fail_shape_inference("Both 'split' input and 'num_outputs' attribute were given");
          }
          if (ctx.hasInput(1)) { //'split' is input
            auto split_proto = ctx.getInputData(1);
            if (split_proto == nullptr) {
              // skip if split is not an initializer
              return;
            }
            split = ParseData<int64_t>(split_proto);
            if (split.size() != ctx.getNumOutputs()) {
              fail_shape_inference(
                  "Mismatch between number of splits (", split.size(), ") and outputs (", ctx.getNumOutputs(), ")");
            }
            int64_t total_dim = 0;
            for (int64_t d : split) {
              total_dim += d;
            }
            if (total_dim != split_dim_value) {
              fail_shape_inference(
                  "Mismatch between the sum of 'split' (",
                  total_dim,
                  ") and the split dimension of the input (",
                  split_dim_value,
                  ")");
            }
          } else { // no value available for 'split'
            if (num_outputs_attr) {
              const auto num_outputs = num_outputs_attr->i();
              if (num_outputs < 1) {
                fail_shape_inference("Attribute `num_outputs` value cannot be lower than 1");
              }
              if (split_dim_value % num_outputs == 0) { // tensor is evenly splittable
                int chunk_size = split_dim_value / num_outputs;
                split.resize(num_outputs, chunk_size);
              } else { // tensor needs to be split unevenly
                int chunk_size = (split_dim_value / num_outputs) + 1;
                int last_chunk_size = split_dim_value - (chunk_size * (num_outputs - 1));
                split.resize(num_outputs - 1, chunk_size);
                split.push_back(last_chunk_size);
              }
            } else {
              fail_shape_inference("Neither 'split' input nor 'num_outputs' attribute has been given");
            }
          }
          for (size_t i = 0; i < ctx.getNumOutputs(); i++) {
            *ctx.getOutputType(i)->mutable_tensor_type()->mutable_shape() = shape;
            ctx.getOutputType(i)->mutable_tensor_type()->mutable_shape()->mutable_dim(axis)->set_dim_value(split[i]);
          }
        }));

static const char* Slice_ver13_doc = R"DOC(
Produces a slice of the input tensor along multiple axes. Similar to numpy:
https://numpy.org/doc/stable/user/basics.indexing.html?highlight=slice#slicing-and-striding

Slice uses the `starts`, `ends`, `axes` and `steps` inputs to select a sub-tensor
of its input `data` tensor.

An effective `starts[i]`, `ends[i]`, and `steps[i]` must be computed for each `i`
in `[0, ... r-1]` where `r = rank(input)` as follows:

If `axes` are omitted, they are set to `[0, ..., r-1]`.
If `steps` are omitted, they are set to `[1, ..., 1]` of length `len(starts)`

The effective values are initialized as `start[i] = 0`, `ends[i] = dims[i]` where
`dims` are the dimensions of `input` and `steps[i] = 1`.

All negative elements of `axes` are made non-negative by adding `r` to them, where
`r =rank(input)`.

All negative values in `starts[i]` and `ends[i]` have `dims[axes[i]]` added to them,
where `dims` are the dimensions of `input`. Then `start[axes[i]]` is the adjusted
`starts[i]` is clamped into the range `[0, dims[axes[i]]]` for positive stepping
and `[0, dims[axes[i]]-1]` for negative stepping.

The clamping for the adjusted `ends[i]` depends on the sign of `steps[i]` and must
accommodate copying 0 through `dims[axes[i]]` elements, so for positive stepping
`ends[axes[i]]` is clamped to `[0, dims[axes[i]]]`, while for negative stepping it
is clamped to `[-1, dims[axes[i]]-1]`.

Finally, `steps[axes[i]] = steps[i]`.

For slicing to the end of a dimension with unknown size, it is recommended to pass
in `INT_MAX` when slicing forward and 'INT_MIN' when slicing backward.

Example 1:

```
data = [
    [1, 2, 3, 4],
    [5, 6, 7, 8],
]
axes = [0, 1]
starts = [1, 0]
ends = [2, 3]
steps = [1, 2]
result = [
    [5, 7],
]
```

Example 2:

```
data = [
    [1, 2, 3, 4],
    [5, 6, 7, 8],
]
starts = [0, 1]
ends = [-1, 1000]
result = [
    [2, 3, 4],
]
```
)DOC";

inline void processSliceInputs(const int64_t input_rank, int64_t& start, int64_t& end, int64_t& step) {
  auto clamp = [](int64_t val, int64_t min, int64_t max) -> int64_t {
    return (val < min) ? min : (val > max) ? max : val;
  };
  // process step
  if (step == 0) {
    fail_shape_inference("'step' cannot be 0 for Slice");
  }
  // process start
  if (start < 0)
    start += input_rank;
  if (step < 0)
    start = clamp(start, 0, input_rank - 1);
  else
    start = clamp(start, 0, input_rank);
  // process end
  if (end < 0)
    end += input_rank;
  if (step < 0)
    end = clamp(end, -1, input_rank - 1);
  else
    end = clamp(end, 0, input_rank);
}

ONNX_OPERATOR_SET_SCHEMA(
    Slice,
    13,
    OpSchema()
        .SetDoc(Slice_ver13_doc)
        .Input(
            0,
            "data",
            "Tensor of data to extract slices from.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Input(
            1,
            "starts",
            "1-D tensor of starting indices of corresponding axis in `axes`",
            "Tind",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            2,
            "ends",
            "1-D tensor of ending indices (exclusive) of corresponding axis in `axes`",
            "Tind",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            3,
            "axes",
            "1-D tensor of axes that `starts` and `ends` apply to. Negative value means counting dimensions "
            "from the back. Accepted range is [-r, r-1] where r = rank(data). Behavior is undefined if an "
            "axis is repeated.",
            "Tind",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            4,
            "steps",
            "1-D tensor of slice step of corresponding axis in `axes`. "
            "Negative value means slicing backward. 'steps' cannot be 0. "
            "Defaults to 1s.",
            "Tind",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(0, "output", "Sliced data tensor.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to all tensor types.")
        .TypeConstraint("Tind", {"tensor(int32)", "tensor(int64)"}, "Constrain indices to integer types")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          size_t num_inputs = ctx.getNumInputs();
          if (num_inputs != 3 && num_inputs != 4 && num_inputs != 5) {
            fail_type_inference("Slice op must have either three, four or five inputs.");
          }
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (!hasNInputShapes(ctx, 1)) {
            return;
          }
          // Shape Inference if
          //     1. 2nd and 3rd input data (starts, ends) are available.
          // and 2. 4th and 5th optional input (axes, steps) are either not set,
          // or set and is initializer.
          const TensorProto* startsInitializer = ctx.getInputData(1);
          const TensorProto* endsInitializer = ctx.getInputData(2);
          const TensorProto* axesInitializer = hasInputShape(ctx, 3) ? ctx.getInputData(3) : nullptr;
          const TensorProto* stepsInitializer = hasInputShape(ctx, 4) ? ctx.getInputData(4) : nullptr;

          if (!startsInitializer || !endsInitializer || (hasInputShape(ctx, 3) && !ctx.getInputData(3)) ||
              (hasInputShape(ctx, 4) && !ctx.getInputData(4))) {
            const auto input_rank = ctx.getInputType(0)->tensor_type().shape().dim_size();
            // we can infer the output rank - it never changes
            for (size_t i = 0; (int64_t)i < input_rank; ++i) {
              ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->add_dim();
            }
            return;
          }

          // don't know data_type- can't proceed
          if (!startsInitializer->has_data_type())
            return;

          auto get_initializer_data = [](const TensorProto* initializer) -> std::vector<int64_t> {
            std::vector<int64_t> vec;
            if (initializer->data_type() == TensorProto::INT64) {
              const auto& data = ParseData<int64_t>(initializer);
              vec.insert(vec.end(), data.begin(), data.end());
            } else if (initializer->data_type() == TensorProto::INT32) {
              const auto& data = ParseData<int32_t>(initializer);
              vec.insert(vec.end(), data.begin(), data.end());
            } else {
              // unaccepted data type
              fail_shape_inference("Only supports `int32_t` or `int64_t` inputs for starts/ends/axes/steps");
            }
            return vec;
          };

          std::vector<int64_t> starts = get_initializer_data(startsInitializer);
          std::vector<int64_t> ends = get_initializer_data(endsInitializer);

          if (starts.size() != ends.size()) {
            fail_shape_inference("Incorrect or missing input value for starts and ends");
          }

          const auto& input_shape = ctx.getInputType(0)->tensor_type().shape();
          const auto input_rank = input_shape.dim_size();
          std::vector<int64_t> axes(starts.size());
          if (!axesInitializer) {
            std::iota(axes.begin(), axes.end(), 0);
          } else {
            axes = get_initializer_data(axesInitializer);
            if (axes.size() != starts.size()) {
              fail_shape_inference("Input axes has incorrect length");
            }
          }
          checkAxesRange(axes, input_rank);
          adjustNegativeAxes(axes, input_rank);
          checkDuplicateAxes(axes, input_rank);
          std::vector<int64_t> steps;
          if (!stepsInitializer) {
            steps = std::vector<int64_t>(starts.size(), 1);
          } else {
            steps = get_initializer_data(stepsInitializer);
            if (steps.size() != axes.size()) {
              fail_shape_inference("Input steps has incorrect length");
            }
          }

          for (size_t i = 0; (int64_t)i < input_rank; ++i) {
            // first update rank of output dim
            auto* output_dim = ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->add_dim();
            const auto& input_dim = input_shape.dim((int)i);
            if (input_dim.has_dim_value()) {
              output_dim->set_dim_value(input_dim.dim_value());
            } else if (input_dim.has_dim_param()) {
              output_dim->set_dim_param(input_dim.dim_param());
            }
          }

          size_t axes_size = axes.size();
          for (size_t axis_index = 0; axis_index < axes_size; ++axis_index) {
            auto axis = axes[axis_index] < 0 ? axes[axis_index] + static_cast<int64_t>(input_rank) : axes[axis_index];

            auto input_dim = ctx.getInputType(0)->tensor_type().shape().dim((int)axis);

            // input dim value is missing - cannot perform shape inference for
            // this axis
            if (!input_dim.has_dim_value()) {
              // Clear any previously propagated dim_param and leave this dimension "empty",
              // before moving on to the next dimension
              ctx.getOutputType(0)
                  ->mutable_tensor_type()
                  ->mutable_shape()
                  ->mutable_dim(static_cast<int>(axis))
                  ->clear_dim_param();
              continue;
            }
            auto start = starts[axis_index];
            auto end = ends[axis_index];
            auto step = steps[axis_index];
            processSliceInputs(input_dim.dim_value(), start, end, step);

            // find output dim value for this axis
            auto temp = static_cast<int64_t>(ceil(1.0 * (end - start) / step));
            if (temp < 0)
              temp = 0;

            // assign output value
            ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->mutable_dim((int)axis)->set_dim_value(temp);
          }
        })
        .PartialDataPropagationFunction([](DataPropagationContext& ctx) {
          const auto input_data = ctx.getInputData(0);
          const auto starts = ctx.getInputData(1);
          const auto ends = ctx.getInputData(2);
          bool axes_specified = ctx.getNumInputs() >= 4;
          bool steps_specified = ctx.getNumInputs() >= 5;

          const TensorShapeProto* axes = nullptr;
          const TensorShapeProto* steps = nullptr;
          if (axes_specified) {
            axes = ctx.getInputData(3);
            if (axes == nullptr) {
              return;
            }
          }
          if (steps_specified) {
            steps = ctx.getInputData(4);
            if (steps == nullptr) {
              return;
            }
          }

          if (input_data == nullptr || starts == nullptr || ends == nullptr) {
            return;
          }
          if (starts->dim_size() != ends->dim_size()) {
            fail_shape_inference(
                "Input rank for starts and ends should be the same: (",
                starts->dim_size(),
                ") vs (",
                ends->dim_size(),
                ").");
          }
          // Only supports axis = 0 since the data comes from Shape
          if ((!axes_specified || (axes->dim_size() == 1 && axes->dim(0).dim_value() == 0)) &&
              starts->dim_size() == 1 && ends->dim_size() == 1) {
            auto start = starts->dim(0).dim_value();
            auto end = ends->dim(0).dim_value();
            int64_t step = 1; // Default step is 1
            if (steps_specified) {
              if (steps->dim_size() != 1) {
                return;
              }
              if (!steps->dim(0).has_dim_value()) {
                return;
              }
              step = steps->dim(0).dim_value();
            }
            processSliceInputs(input_data->dim_size(), start, end, step);

            TensorShapeProto tsp;
            if (step > 0) {
              for (int i = start; i < end; i += step) {
                *tsp.add_dim() = input_data->dim(i);
              }
            } else {
              for (int i = start; i > end; i += step) {
                *tsp.add_dim() = input_data->dim(i);
              }
            }
            if (tsp.dim_size() > 0) {
              ctx.addOutputData(0, std::move(tsp));
            }
          }
        }));

static const char* Transpose_ver13_doc = R"DOC(
Transpose the input tensor similar to numpy.transpose. For example, when
perm=(1, 0, 2), given an input tensor of shape (1, 2, 3), the output shape
will be (2, 1, 3).
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Transpose,
    21,
    OpSchema()
        .SetDoc(Transpose_ver13_doc)
        .Attr(
            "perm",
            "A list of integers. By default, reverse the dimensions, "
            "otherwise permute the axes according to the values given. "
            "Its length must be equal to the rank of the input.",
            AttributeProto::INTS,
            OPTIONAL_VALUE)
        .Input(0, "data", "An input tensor.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Output(0, "transposed", "Transposed output.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir10(), "Constrain input and output types to all tensor types.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (!hasNInputShapes(ctx, 1)) {
            return;
          }
          auto input_type = ctx.getInputType(0);
          const TensorShapeProto& shape = input_type->tensor_type().shape();
          std::vector<int64_t> perm;
          bool has_perm_attr = getRepeatedAttribute(ctx, "perm", perm);
          if (!has_perm_attr) {
            perm.reserve(shape.dim_size());
            for (int i = shape.dim_size() - 1; i >= 0; --i)
              perm.push_back(i);
          } else if (!perm.empty()) {
            // check if every index is valid
            std::vector<bool> seen(shape.dim_size(), false);
            for (int64_t fromDimIndex : perm) {
              if (!(0 <= fromDimIndex && fromDimIndex < shape.dim_size())) {
                std::ostringstream oss;
                oss << "Invalid attribute perm {" << perm[0];
                for (size_t i = 1; i != perm.size(); ++i) {
                  oss << ", " << perm[i];
                }
                oss << "}, input shape = {";
                if (shape.dim_size() > 0) {
                  oss << shape.dim(0).dim_value();
                  for (int i = 1; i != shape.dim_size(); ++i) {
                    oss << ", " << shape.dim(i).dim_value();
                  }
                  oss << "}";
                }
                fail_type_inference(oss.str());
              } else {
                // check if any perm is repeated
                if (seen[fromDimIndex]) {
                  fail_type_inference("Attribute perm for Transpose has repeated value: ", fromDimIndex);
                }
                seen[fromDimIndex] = true;
              }
            }
          }

          getOutputShape(ctx, 0);

          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          for (size_t i = 0; i < perm.size(); ++i) {
            appendSingleDimCopiedFromInputTypeToOutputType(ctx, 0, 0, static_cast<size_t>(perm[i]));
          }
        }));

static const char* Scatter_ver11_doc = R"DOC(
This operator is deprecated. Please use ScatterElements, which provides the same functionality.

Scatter takes three inputs `data`, `updates`, and `indices` of the same
rank r >= 1 and an optional attribute axis that identifies an axis of `data`
(by default, the outer-most axis, that is axis 0). The output of the operation
is produced by creating a copy of the input `data`, and then updating its value
to values specified by `updates` at specific index positions specified by
`indices`. Its output shape is the same as the shape of `data`.

For each entry in `updates`, the target index in `data` is obtained by combining
the corresponding entry in `indices` with the index of the entry itself: the
index-value for dimension = axis is obtained from the value of the corresponding
entry in `indices` and the index-value for dimension != axis is obtained from the
index of the entry itself.

For instance, in a 2-D tensor case, the update corresponding to the [i][j] entry
is performed as below:
```
  output[indices[i][j]][j] = updates[i][j] if axis = 0,
  output[i][indices[i][j]] = updates[i][j] if axis = 1,
```

This operator is the inverse of GatherElements. It is similar to Torch's Scatter operation.

Example 1:
```
  data = [
      [0.0, 0.0, 0.0],
      [0.0, 0.0, 0.0],
      [0.0, 0.0, 0.0],
  ]
  indices = [
      [1, 0, 2],
      [0, 2, 1],
  ]
  updates = [
      [1.0, 1.1, 1.2],
      [2.0, 2.1, 2.2],
  ]
  output = [
      [2.0, 1.1, 0.0]
      [1.0, 0.0, 2.2]
      [0.0, 2.1, 1.2]
  ]
```
Example 2:
```
  data = [[1.0, 2.0, 3.0, 4.0, 5.0]]
  indices = [[1, 3]]
  updates = [[1.1, 2.1]]
  axis = 1
  output = [[1.0, 1.1, 3.0, 2.1, 5.0]]
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Scatter,
    11,
    OpSchema()
        .Deprecate()
        .SetDoc(Scatter_ver11_doc)
        .Attr(
            "axis",
            "Which axis to scatter on. Negative value means "
            "counting dimensions from the back. Accepted range is [-r, r-1] where r = rank(data).",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Input(0, "data", "Tensor of rank r >= 1.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "indices",
            "Tensor of int32/int64 indices, of r >= 1 (same rank as input). All index values are expected to be "
            "within bounds [-s, s-1] along axis of size s. It is an error if any of the index values are out of bounds.",
            "Tind",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            2,
            "updates",
            "Tensor of rank r >=1 (same rank and shape as indices)",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Output(
            0,
            "output",
            "Tensor of rank r >= 1 (same rank as input).",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Input and output types can be of any tensor type.")
        .TypeConstraint("Tind", {"tensor(int32)", "tensor(int64)"}, "Constrain indices to integer types")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (hasNInputShapes(ctx, 1)) {
            propagateShapeFromInputToOutput(ctx, 0, 0);
          }
        }));

static const char* ScatterND_ver18_doc = R"DOC(
ScatterND takes three inputs `data` tensor of rank r >= 1, `indices` tensor of rank q >= 1,
and `updates` tensor of rank q + r - indices.shape[-1] - 1. The output of the operation
is produced by creating a copy of the input `data`, and then updating its value to values
specified by `updates` at specific index positions specified by `indices`. Its output shape
is the same as the shape of `data`.

`indices` is an integer tensor. Let k denote indices.shape[-1], the last dimension in the shape of `indices`.
`indices` is treated as a (q-1)-dimensional tensor of k-tuples, where each k-tuple is a partial-index into `data`.
Hence, k can be a value at most the rank of `data`. When k equals rank(data), each update entry specifies an
update to a single element of the tensor. When k is less than rank(data) each update entry specifies an
update to a slice of the tensor. Index values are allowed to be negative, as per the usual
convention for counting backwards from the end, but are expected in the valid range.

`updates` is treated as a (q-1)-dimensional tensor of replacement-slice-values. Thus, the
first (q-1) dimensions of updates.shape must match the first (q-1) dimensions of indices.shape.
The remaining dimensions of `updates` correspond to the dimensions of the
replacement-slice-values. Each replacement-slice-value is a (r-k) dimensional tensor,
corresponding to the trailing (r-k) dimensions of `data`.  Thus, the shape of `updates`
must equal indices.shape[0:q-1] ++ data.shape[k:r-1], where ++ denotes the concatenation
of shapes.

The `output` is calculated via the following equation:

```
output = np.copy(data)
update_indices = indices.shape[:-1]
for idx in np.ndindex(update_indices):
    output[indices[idx]] = updates[idx]
```

The order of iteration in the above loop is not specified.
In particular, indices should not have duplicate entries: that is, if idx1 != idx2, then indices[idx1] != indices[idx2].
This ensures that the output value does not depend on the iteration order.

`reduction` allows specification of an optional reduction operation, which is applied to all values in `updates`
tensor into `output` at the specified `indices`.
In cases where `reduction` is set to "none", indices should not have duplicate entries: that is, if idx1 != idx2,
then indices[idx1] != indices[idx2]. This ensures that the output value does not depend on the iteration order.
When `reduction` is set to some reduction function `f`, `output` is calculated as follows:

```
output = np.copy(data)
update_indices = indices.shape[:-1]
for idx in np.ndindex(update_indices):
    output[indices[idx]] = f(output[indices[idx]], updates[idx])
```

where the `f` is `+`, `*`, `max` or `min` as specified.

This operator is the inverse of GatherND.

(Opset 18 change): Adds max/min to the set of allowed reduction ops.

Example 1:
```
data    = [1, 2, 3, 4, 5, 6, 7, 8]
indices = [[4], [3], [1], [7]]
updates = [9, 10, 11, 12]
output  = [1, 11, 3, 10, 9, 6, 7, 12]
```

Example 2:
```
data    = [[[1, 2, 3, 4], [5, 6, 7, 8], [8, 7, 6, 5], [4, 3, 2, 1]],
            [[1, 2, 3, 4], [5, 6, 7, 8], [8, 7, 6, 5], [4, 3, 2, 1]],
            [[8, 7, 6, 5], [4, 3, 2, 1], [1, 2, 3, 4], [5, 6, 7, 8]],
            [[8, 7, 6, 5], [4, 3, 2, 1], [1, 2, 3, 4], [5, 6, 7, 8]]]
indices = [[0], [2]]
updates = [[[5, 5, 5, 5], [6, 6, 6, 6], [7, 7, 7, 7], [8, 8, 8, 8]],
            [[1, 1, 1, 1], [2, 2, 2, 2], [3, 3, 3, 3], [4, 4, 4, 4]]]
output  = [[[5, 5, 5, 5], [6, 6, 6, 6], [7, 7, 7, 7], [8, 8, 8, 8]],
            [[1, 2, 3, 4], [5, 6, 7, 8], [8, 7, 6, 5], [4, 3, 2, 1]],
            [[1, 1, 1, 1], [2, 2, 2, 2], [3, 3, 3, 3], [4, 4, 4, 4]],
            [[8, 7, 6, 5], [4, 3, 2, 1], [1, 2, 3, 4], [5, 6, 7, 8]]]
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    ScatterND,
    18,
    OpSchema()
        .SetDoc(ScatterND_ver18_doc)
        .Attr(
            "reduction",
            "Type of reduction to apply: none (default), add, mul, max, min. "
            "'none': no reduction applied. "
            "'add':  reduction using the addition operation. "
            "'mul':  reduction using the addition operation. "
            "'max': reduction using the maximum operation."
            "'min': reduction using the minimum operation.",
            AttributeProto::STRING,
            std::string("none"))
        .Input(0, "data", "Tensor of rank r >= 1.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "indices",
            "Tensor of rank q >= 1.",
            "tensor(int64)",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            2,
            "updates",
            "Tensor of rank q + r - indices_shape[-1] - 1.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Output(0, "output", "Tensor of rank r >= 1.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to any tensor type.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (hasNInputShapes(ctx, 1)) {
            propagateShapeFromInputToOutput(ctx, 0, 0);
          }
        }));

static const char* ScatterElements_ver18_doc = R"DOC(
ScatterElements takes three inputs `data`, `updates`, and `indices` of the same
rank r >= 1 and an optional attribute axis that identifies an axis of `data`
(by default, the outer-most axis, that is axis 0). The output of the operation
is produced by creating a copy of the input `data`, and then updating its value
to values specified by `updates` at specific index positions specified by
`indices`. Its output shape is the same as the shape of `data`.

For each entry in `updates`, the target index in `data` is obtained by combining
the corresponding entry in `indices` with the index of the entry itself: the
index-value for dimension = axis is obtained from the value of the corresponding
entry in `indices` and the index-value for dimension != axis is obtained from the
index of the entry itself.

`reduction` allows specification of an optional reduction operation, which is applied to all values in `updates`
tensor into `output` at the specified `indices`.
In cases where `reduction` is set to "none", indices should not have duplicate entries: that is, if idx1 != idx2,
then indices[idx1] != indices[idx2]. For instance, in a 2-D tensor case, the update
corresponding to the [i][j] entry is performed as below:
```
output[indices[i][j]][j] = updates[i][j] if axis = 0,
output[i][indices[i][j]] = updates[i][j] if axis = 1,
```
When `reduction` is set to some reduction function `f`, the update corresponding to the [i][j] entry is performed as below:
```
output[indices[i][j]][j] = f(output[indices[i][j]][j], updates[i][j]) if axis = 0,
output[i][indices[i][j]] = f(output[i][indices[i][j]], updates[i][j]) if axis = 1,
```
where the `f` is `+`, `*`, `max` or `min` as specified.

This operator is the inverse of GatherElements. It is similar to Torch's Scatter operation.

(Opset 18 change): Adds max/min to the set of allowed reduction ops.

Example 1:
```
data = [
    [0.0, 0.0, 0.0],
    [0.0, 0.0, 0.0],
    [0.0, 0.0, 0.0],
]
indices = [
    [1, 0, 2],
    [0, 2, 1],
]
updates = [
    [1.0, 1.1, 1.2],
    [2.0, 2.1, 2.2],
]
output = [
    [2.0, 1.1, 0.0]
    [1.0, 0.0, 2.2]
    [0.0, 2.1, 1.2]
]
```
Example 2:
```
data = [[1.0, 2.0, 3.0, 4.0, 5.0]]
indices = [[1, 3]]
updates = [[1.1, 2.1]]
axis = 1
output = [[1.0, 1.1, 3.0, 2.1, 5.0]]
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    ScatterElements,
    18,
    OpSchema()
        .SetDoc(ScatterElements_ver18_doc)
        .Attr(
            "axis",
            "Which axis to scatter on. Negative value means "
            "counting dimensions from the back. Accepted range is [-r, r-1] where r = rank(data).",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Attr(
            "reduction",
            "Type of reduction to apply: none (default), add, mul, max, min. "
            "'none': no reduction applied. "
            "'add':  reduction using the addition operation. "
            "'mul': reduction using the multiplication operation."
            "'max': reduction using the maximum operation."
            "'min': reduction using the minimum operation.",
            AttributeProto::STRING,
            std::string("none"))
        .Input(0, "data", "Tensor of rank r >= 1.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "indices",
            "Tensor of int32/int64 indices, of r >= 1 (same rank as input). All index values are expected to be "
            "within bounds [-s, s-1] along axis of size s. It is an error if any of the index values are out of bounds.",
            "Tind",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            2,
            "updates",
            "Tensor of rank r >=1 (same rank and shape as indices)",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Output(
            0,
            "output",
            "Tensor of rank r >= 1 (same rank as input).",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Input and output types can be of any tensor type.")
        .TypeConstraint("Tind", {"tensor(int32)", "tensor(int64)"}, "Constrain indices to integer types")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (hasNInputShapes(ctx, 1)) {
            propagateShapeFromInputToOutput(ctx, 0, 0);
          }
        }));

static const char* Gather_ver13_doc = R"DOC(
Given `data` tensor of rank r >= 1, and `indices` tensor of rank q, gather
entries of the axis dimension of `data` (by default outer-most one as axis=0) indexed by `indices`, and concatenates
them in an output tensor of rank q + (r - 1).

If `axis = 0`, let `k = indices[i_{0}, ..., i_{q-1}]`
then `output[i_{0}, ..., i_{q-1}, j_{0}, ..., j_{r-2}] = input[k , j_{0}, ..., j_{r-2}]`:

```
data = [
    [1.0, 1.2],
    [2.3, 3.4],
    [4.5, 5.7],
]
indices = [
    [0, 1],
    [1, 2],
]
output = [
    [
        [1.0, 1.2],
        [2.3, 3.4],
    ],
    [
        [2.3, 3.4],
        [4.5, 5.7],
    ],
]
```

If `axis = 1`, let `k = indices[i_{0}, ..., i_{q-1}]`
then `output[j_{0}, i_{0}, ..., i_{q-1}, j_{1}, ..., j_{r-2}] = input[j_{0}, k, j_{1}, ..., j_{r-2}]`:

```
data = [
    [1.0, 1.2, 1.9],
    [2.3, 3.4, 3.9],
    [4.5, 5.7, 5.9],
]
indices = [
    [0, 2],
]
axis = 1,
output = [
        [[1.0, 1.9]],
        [[2.3, 3.9]],
        [[4.5, 5.9]],
]
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Gather,
    13,
    OpSchema()
        .SetDoc(Gather_ver13_doc)
        .Attr(
            "axis",
            "Which axis to gather on. Negative value means "
            "counting dimensions from the back. Accepted range is [-r, r-1] where r = rank(data).",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Input(0, "data", "Tensor of rank r >= 1.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "indices",
            "Tensor of int32/int64 indices, of any rank q. All index values are expected to be within bounds [-s, s-1] "
            "along axis of size s. It is an error if any of the index values are out of bounds.",
            "Tind",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(0, "output", "Tensor of rank q + (r - 1).", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to any tensor type.")
        .TypeConstraint("Tind", {"tensor(int32)", "tensor(int64)"}, "Constrain indices to integer types")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (!hasNInputShapes(ctx, 2)) {
            return;
          }
          const TensorShapeProto& data_shape = ctx.getInputType(0)->tensor_type().shape();
          const TensorShapeProto& indices_shape = ctx.getInputType(1)->tensor_type().shape();
          int r = data_shape.dim_size();
          if (r < 1) {
            fail_shape_inference("data tensor must have rank >= 1");
          }
          int q = indices_shape.dim_size();
          int axis = static_cast<int>(getAttribute(ctx, "axis", 0));
          if (axis < -r || axis >= r) {
            fail_shape_inference("axis must be in [-r, r-1]");
          }
          if (axis < 0) {
            axis += r;
          }
          int out_rank = q + r - 1;
          if (out_rank == 0) {
            ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();
          }
          for (int i = 0; i < out_rank; ++i) {
            *ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->add_dim() = (i < axis) ? data_shape.dim(i)
                                                                                                  : // i < axis < r
                (i >= axis && i < axis + q) ? indices_shape.dim(i - axis)
                                            : // i - axis < q
                data_shape.dim(i - q + 1); // i < out_rank < q + r - 1
          }
        })
        .PartialDataPropagationFunction([](DataPropagationContext& ctx) { GatherOp13DataPropagator(ctx); }));

static const char* GatherElements_ver13_doc = R"DOC(

GatherElements takes two inputs `data` and `indices` of the same rank r >= 1
and an optional attribute `axis` that identifies an axis of `data`
(by default, the outer-most axis, that is axis 0). It is an indexing operation
that produces its output by indexing into the input data tensor at index
positions determined by elements of the `indices` tensor.
Its output shape is the same as the shape of `indices` and consists of one value
(gathered from the `data`) for each element in `indices`.

For instance, in the 3-D case (r = 3), the output produced is determined
by the following equations:
```
out[i][j][k] = input[index[i][j][k]][j][k] if axis = 0,
out[i][j][k] = input[i][index[i][j][k]][k] if axis = 1,
out[i][j][k] = input[i][j][index[i][j][k]] if axis = 2,
```

This operator is also the inverse of ScatterElements. It is similar to Torch's gather operation.

Example 1:
```
data = [
    [1, 2],
    [3, 4],
]
indices = [
    [0, 0],
    [1, 0],
]
axis = 1
output = [
    [1, 1],
    [4, 3],
]
```
Example 2:
```
data = [
    [1, 2, 3],
    [4, 5, 6],
    [7, 8, 9],
]
indices = [
    [1, 2, 0],
    [2, 0, 0],
]
axis = 0
output = [
    [4, 8, 3],
    [7, 2, 3],
]
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    GatherElements,
    13,
    OpSchema()
        .SetDoc(GatherElements_ver13_doc)
        .Attr(
            "axis",
            "Which axis to gather on. Negative value means "
            "counting dimensions from the back. Accepted range is [-r, r-1] where r = rank(data).",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Input(0, "data", "Tensor of rank r >= 1.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "indices",
            "Tensor of int32/int64 indices, with the same rank r as the input. All index values are expected to be "
            "within bounds [-s, s-1] along axis of size s. It is an error if any of the index values are out of bounds.",
            "Tind",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "output",
            "Tensor of the same shape as indices.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to any tensor type.")
        .TypeConstraint("Tind", {"tensor(int32)", "tensor(int64)"}, "Constrain indices to integer types")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          // propagate indices' shape to output if it exists
          if (hasInputShape(ctx, 1)) {
            propagateShapeFromInputToOutput(ctx, 1, 0);
          }
        }));

static const char* Squeeze_ver13_doc = R"DOC(
Remove single-dimensional entries from the shape of a tensor.
Takes an input `axes` with a list of axes to squeeze.
If `axes` is not provided, all the single dimensions will be removed from
the shape. If an axis is selected with shape entry not equal to one, an error is raised.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Squeeze,
    21,
    OpSchema()
        .SetDoc(Squeeze_ver13_doc)
        .Input(
            0,
            "data",
            "Tensors with at least max(dims) dimensions.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Input(
            1,
            "axes",
            "List of integers indicating the dimensions to squeeze. Negative value means counting dimensions "
            "from the back. Accepted range is [-r, r-1] where r = rank(data).",
            "tensor(int64)",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "squeezed",
            "Reshaped tensor with same data as input.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint(
            "T",
            OpSchema::all_tensor_types_ir10(),
            "Constrain input and output types to all tensor types up to IRv10.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (!hasNInputShapes(ctx, 1)) {
            return;
          }

          std::vector<int64_t> axes;
          size_t num_inputs = ctx.getNumInputs();
          bool axes_not_specified = false;

          if ((num_inputs == 2) && ctx.getInputType(1)) { //'axes' is input
            auto axes_proto = ctx.getInputData(1);
            if (axes_proto == nullptr) {
              // skip if axes is not an initializer
              return;
            }
            axes = ParseData<int64_t>(axes_proto);
          } else {
            // axes not specified
            axes_not_specified = true;
          }

          const auto& input_shape = ctx.getInputType(0)->tensor_type().shape();
          const auto input_ndim = input_shape.dim_size();
          checkAxesRange(axes, input_ndim);
          adjustNegativeAxes(axes, input_ndim);

          for (int i = 0; i < input_ndim; ++i) {
            if (!input_shape.dim(i).has_dim_value() && axes_not_specified) {
              // if dim has a symbolic value and the axes spec want to act on all dims,
              // return early because we can't infer the shape
              return;
            }
          }

          ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();

          for (int i = 0; i < input_ndim; ++i) {
            if (axes_not_specified && input_shape.dim(i).dim_value() == 1) {
              // if axes not specified, do not keep shape if the dimension is equal to one
              continue;
            } else if (!axes_not_specified && std::find(axes.begin(), axes.end(), i) != axes.end()) {
              // if axes wants to explicitly act on this dim, fail explicitly only if the
              // dim is numerical and != 1. If the dim is 1 or symbolic, remove it. If
              // the dim is symbolic, runtime engines should check that the dimension is
              // actually 1 when the op is evaluated
              if (input_shape.dim(i).has_dim_value() && input_shape.dim(i).dim_value() != 1) {
                fail_shape_inference(
                    "Dimension of input ", i, " must be 1 instead of ", input_shape.dim(i).dim_value());
              }
            } else {
              *ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->add_dim() = input_shape.dim(i);
            }
          }
        })
        .PartialDataPropagationFunction([](DataPropagationContext& ctx) {
          PropagateShapeDataFromInputToOutput(ctx, 0);
        }));

static const char* Unsqueeze_ver13_doc = R"DOC(
Insert single-dimensional entries to the shape of an input tensor (`data`).
Takes one required input `axes` - which contains a list of dimension indices and this operator will insert a dimension of value `1` into the corresponding index of the output tensor (`expanded`).

For example, given an input tensor (`data`) of shape [3, 4, 5], then
Unsqueeze(data, axes=[0, 4]) outputs a tensor (`expanded`) containing same data as `data` but with shape [1, 3, 4, 5, 1].

The input `axes` should not contain any duplicate entries. It is an error if it contains duplicates.
The rank of the output tensor (`output_rank`) is the rank of the input tensor (`data`) plus the number of values in `axes`.
Each value in `axes` should be within the (inclusive) range [-output_rank , output_rank - 1].
The order of values in `axes` does not matter and can come in any order.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Unsqueeze,
    21,
    OpSchema()
        .SetDoc(Unsqueeze_ver13_doc)
        .Input(0, "data", "Original tensor", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "axes",
            "List of integers indicating the dimensions to be inserted. Negative value means counting dimensions "
            "from the back. Accepted range is [-r, r-1] where r = rank(expanded).",
            "tensor(int64)",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "expanded",
            "Reshaped tensor with same data as input.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint(
            "T",
            OpSchema::all_tensor_types_ir10(),
            "Constrain input and output types to all tensor types up to IRv10.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (!hasNInputShapes(ctx, 1)) {
            return;
          }
          std::vector<int64_t> axes;
          auto axes_proto = ctx.getInputData(1);
          if (axes_proto == nullptr) {
            // skip if axes is not an initializer
            return;
          }
          axes = ParseData<int64_t>(axes_proto);
          ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();
          const auto& input_shape = ctx.getInputType(0)->tensor_type().shape();
          const auto input_ndim = input_shape.dim_size();
          const auto output_ndim = input_ndim + static_cast<int>(axes.size());
          checkAxesRange(axes, output_ndim);
          adjustNegativeAxes(axes, output_ndim);
          checkDuplicateAxes(axes, output_ndim);
          // sort after correcting negative axes values (if any)
          std::sort(axes.begin(), axes.end());

          int j = 0;
          for (int i = 0; i < input_ndim; ++i) {
            while (static_cast<size_t>(j) < axes.size() &&
                   axes[j] == ctx.getOutputType(0)->tensor_type().shape().dim_size()) {
              ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->add_dim()->set_dim_value(1);
              ++j;
            }
            *ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->add_dim() =
                ctx.getInputType(0)->tensor_type().shape().dim(i);
          }
          while (static_cast<size_t>(j) < axes.size() &&
                 axes[j] == ctx.getOutputType(0)->tensor_type().shape().dim_size()) {
            ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->add_dim()->set_dim_value(1);
            ++j;
          }
        })
        .PartialDataPropagationFunction([](DataPropagationContext& ctx) {
          PropagateShapeDataFromInputToOutput(ctx, 0);
        }));

static const char* SpaceToDepth_ver13_doc =
    R"DOC(SpaceToDepth rearranges blocks of spatial data into depth. More specifically,
this op outputs a copy of the input tensor where values from the height and width dimensions
are moved to the depth dimension.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    SpaceToDepth,
    13,
    OpSchema()
        .Attr("blocksize", "Blocks of [blocksize, blocksize] are moved.", AttributeProto::INT)
        .SetDoc(SpaceToDepth_ver13_doc)
        .Input(
            0,
            "input",
            "Input tensor of [N,C,H,W], where N is the batch axis, C is the channel or depth"
            ", H is the height and W is the width.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Output(
            0,
            "output",
            "Output tensor of [N, C * blocksize * blocksize, H/blocksize, W/blocksize].",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to all tensor types.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          auto blocksize = getAttribute(ctx, "blocksize", 0);
          if (blocksize <= 0) {
            fail_shape_inference("Blocksize must be positive");
          }
          if (hasInputShape(ctx, 0)) {
            auto& input_shape = getInputShape(ctx, 0);
            if (input_shape.dim_size() == 4) {
              // TODO: Clarify what behavior should be if H or W is not a
              // multiple of blocksize.
              updateOutputShape(
                  ctx,
                  0,
                  {input_shape.dim(0),
                   input_shape.dim(1) * (blocksize * blocksize),
                   input_shape.dim(2) / blocksize,
                   input_shape.dim(3) / blocksize});
            } else {
              fail_shape_inference("Input tensor must be 4-dimensional");
            }
          }
        }));

static const char* DepthToSpace_ver13_doc =
    R"DOC(DepthToSpace rearranges (permutes) data from depth into blocks of spatial data.
This is the reverse transformation of SpaceToDepth. More specifically, this op outputs a copy of
the input tensor where values from the depth dimension are moved in spatial blocks to the height
and width dimensions. By default, `mode` = `DCR`.
In the DCR mode, elements along the depth dimension from the input tensor are rearranged in the
following order: depth, column, and then row. The output y is computed from the input x as below:

```
b, c, h, w = x.shape
tmp = np.reshape(x, [b, blocksize, blocksize, c // (blocksize**2), h, w])
tmp = np.transpose(tmp, [0, 3, 4, 1, 5, 2])
y = np.reshape(tmp, [b, c // (blocksize**2), h * blocksize, w * blocksize])
```

In the CRD mode, elements along the depth dimension from the input tensor are rearranged in the
following order: column, row, and the depth. The output y is computed from the input x as below:

```
b, c, h, w = x.shape
tmp = np.reshape(x, [b, c // (blocksize ** 2), blocksize, blocksize, h, w])
tmp = np.transpose(tmp, [0, 1, 4, 2, 5, 3])
y = np.reshape(tmp, [b, c // (blocksize ** 2), h * blocksize, w * blocksize])
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    DepthToSpace,
    13,
    OpSchema()
        .Attr("blocksize", "Blocks of [blocksize, blocksize] are moved.", AttributeProto::INT)
        .Attr(
            "mode",
            "DCR (default) for depth-column-row order re-arrangement. Use CRD for column-row-depth order.",
            AttributeProto::STRING,
            std::string("DCR"))
        .SetDoc(DepthToSpace_ver13_doc)
        .Input(
            0,
            "input",
            "Input tensor of [N,C,H,W], where N is the batch axis, C is the channel or depth"
            ", H is the height and W is the width.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Output(
            0,
            "output",
            "Output tensor of [N, C/(blocksize * blocksize), H * blocksize, W * blocksize].",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to all tensor types.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          auto blocksize = getAttribute(ctx, "blocksize", 0);
          if (blocksize <= 0) {
            fail_shape_inference("Blocksize must be positive");
          }
          if (hasInputShape(ctx, 0)) {
            auto& input_shape = getInputShape(ctx, 0);
            if (input_shape.dim_size() == 4) {
              // TODO: Clarify what behavior should be if C is not a multiple of
              // blocksize*blocksize.
              updateOutputShape(
                  ctx,
                  0,
                  {input_shape.dim(0),
                   input_shape.dim(1) / (blocksize * blocksize),
                   input_shape.dim(2) * blocksize,
                   input_shape.dim(3) * blocksize});
            } else {
              fail_shape_inference("Input tensor must be 4-dimensional");
            }
          }
        }));

static const char* Tile_ver13_doc =
    R"DOC(Constructs a tensor by tiling a given tensor.
This is the same as function `tile` in Numpy, but no broadcast.
For example A = [[1, 2], [3, 4]], B = [1, 2], tile(A, B) = [[1, 2, 1, 2], [3, 4, 3, 4]]
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Tile,
    13,
    OpSchema()
        .SetDoc(Tile_ver13_doc)
        .Input(0, "input", "Input tensor of any shape.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "repeats",
            "1D int64 tensor of the same length as input's dimension number, "
            "includes numbers of repeated copies along input's dimensions.",
            "T1",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "output",
            "Output tensor of the same dimensions and type as tensor input. "
            "output_dim[i] = input_dim[i] * repeats[i]",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to all tensor types.")
        .TypeConstraint("T1", {"tensor(int64)"}, "Constrain repeat's type to int64 tensors.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          // Type inference
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          // Shape inference

          // Needs at least the first input to proceed
          if (!hasNInputShapes(ctx, 1)) {
            return;
          }

          const auto& input_shape = ctx.getInputType(0)->tensor_type().shape();
          const auto input_rank = input_shape.dim_size();

          const auto* repeats_inputs = ctx.getInputData(1);

          auto* output_shape = ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();

          if (nullptr != repeats_inputs && hasNInputShapes(ctx, 2)) {
            // shape inference is possible only when 'repeats' is an initializer
            const auto& repeats_shape = ctx.getInputType(1)->tensor_type().shape();
            if (repeats_shape.dim_size() != 1 || repeats_inputs->data_type() != TensorProto::INT64) {
              fail_shape_inference("'Repeats' input must be 1D tensor of type int64");
            }

            const auto& repeats_data = ParseData<int64_t>(repeats_inputs);

            if (repeats_data.size() != static_cast<size_t>(input_rank)) {
              fail_shape_inference(
                  "'Repeats' input has incorrect number of values. "
                  "The number of values in 'repeats' must be equal "
                  "to the number of input dimensions.");
            }

            for (size_t i = 0; (int64_t)i < input_rank; ++i) {
              const auto& input_dim = input_shape.dim((int)i);
              auto* output_dim = output_shape->add_dim();
              if (input_dim.has_dim_value()) {
                output_dim->set_dim_value(input_dim.dim_value() * repeats_data[i]);
              }
            }
          } else {
            // Infer output shape's rank in any case (if repeats data is not
            // available)
            auto* output_shape_0 = getOutputShape(ctx, 0);
            for (size_t i = 0; (int64_t)i < input_rank; ++i) {
              output_shape_0->add_dim();
            }
          }
          return;
        }));

static const char* Upsample_ver10_doc = R"DOC(
Upsample the input tensor.
Each dimension value of the output tensor is:
  output_dimension = floor(input_dimension * scale).
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Upsample,
    10,
    OpSchema()
        .Deprecate()
        .Attr(
            "mode",
            "Two interpolation modes: nearest (default), and linear (including bilinear, trilinear, etc)",
            AttributeProto::STRING,
            std::string("nearest"))
        .Input(0, "X", "N-D tensor", "T", OpSchema::Single)
        .Input(
            1,
            "scales",
            "The scale array along each dimension. It takes value greater than or equal to 1."
            " The number of elements of 'scales' should be the same as the rank of input 'X'.",
            "tensor(float)",
            OpSchema::Single)
        .Output(0, "Y", "N-D tensor after resizing", "T", OpSchema::Single)
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Constrain input 'X' and output 'Y' to all tensor types.")
        .SetDoc(Upsample_ver10_doc)
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) { resizeShapeInference_opset7_to_10(ctx); }));

static const char* Resize_ver19_doc = R"DOC(
Resize the input tensor. In general, it calculates every value in the output tensor as a weighted average of neighborhood (a.k.a. sampling locations) in the input tensor.
Each dimension value of the output tensor is:
```
output_dimension = floor(input_dimension * (roi_end - roi_start) * scale)
```
if input \"sizes\" is not specified.
)DOC";

static const char* Resize_ver19_attr_coordinate_transformation_mode_doc = R"DOC(
This attribute describes how to transform the coordinate in the resized tensor to the coordinate in the original tensor.

The coordinate of each dimension is transformed individually. Let's describe a case using axis x as an example.
Denote `x_resized` as the coordinate of axis x in the resized tensor,
 `x_original` as the coordinate of axis x in the original tensor,
 `length_original` as the length of the original tensor in axis x,
 `length_resized` as the length of the resized tensor in axis x,
 `scale = length_resized / length_original`,
 `output_width` the target length on the axis x which can be a fractional number when it is calculated out of a scale factor,
 and `output_width_int` the effective output width as an integer.

if coordinate_transformation_mode is `"half_pixel"`,
```
x_original = (x_resized + 0.5) / scale - 0.5
```

if coordinate_transformation_mode is `"half_pixel_symmetric"`,
```
adjustment = output_width_int / output_width
center = input_width / 2
offset = center * (1 - adjustment)
x_ori = offset + (x + 0.5) / scale - 0.5
```

if coordinate_transformation_mode is `"pytorch_half_pixel"`,
```
x_original = length_resized > 1 ? (x_resized + 0.5) / scale - 0.5 : 0
```

if coordinate_transformation_mode is `"align_corners"`,
```
x_original = x_resized * (length_original - 1) / (length_resized - 1)
```

if coordinate_transformation_mode is `"asymmetric"`,
```
x_original = x_resized / scale
```

if coordinate_transformation_mode is `"tf_crop_and_resize"`,
```
x_original = length_resized > 1 ? start_x * (length_original - 1) + x_resized * (end_x - start_x) * (length_original - 1) / (length_resized - 1) : 0.5 * (start_x + end_x) * (length_original - 1)
```
.)DOC";

static const char* Resize_ver19_attr_keep_aspect_ratio_policy_doc = R"DOC(
This attribute describes how to interpret the `sizes` input with regard to keeping the original aspect ratio of the input, and it is not applicable when
the `scales` input is used.

Given a set of `sizes`, associated with a subset of `axes` (explicitly provided or default), and assuming `d = axes[i]`, with `i` being the index of the provided `sizes`.

If `keep_aspect_ratio_policy` is `"stretch"`, the original aspect ratio is disregarded, and the input is resized to the specified size:
`out_size[d] = sizes[i]`

If `keep_aspect_ratio_policy` is `"not_larger"`, the sizes are adjusted so that no extent of the output is larger than the specified size, while keeping the original aspect ratio:
```
scale = Min(sizes[i] / in_size[d])
out_size[d] = round_int(scale * in_size[i])
```

If `keep_aspect_ratio_policy` is `"not_smaller"`, the sizes are adjusted so that no extent of the output is smaller than the specified size, while keeping the original aspect ratio:
```
scale = Max(sizes[i] / in_size[d])
out_size[d] = round_int(scale * in_size[i])
```

For non-resizable axes (those not specified in `axes`), the output size will be equal to the input size.

Note: `round_int` stands for computing the nearest integer value, rounding halfway cases up.)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Resize,
    19,
    OpSchema()
        .Attr(
            "mode",
            "Three interpolation modes: \"nearest\" (default), \"linear\" and \"cubic\". "
            "The \"linear\" mode includes linear interpolation for 1D tensor and N-linear interpolation for N-D tensor (for example, bilinear interpolation for 2D tensor). "
            "The \"cubic\" mode includes cubic interpolation for 1D tensor and N-cubic interpolation for N-D tensor (for example, bicubic interpolation for 2D tensor).",
            AttributeProto::STRING,
            std::string("nearest"))
        .Attr(
            "cubic_coeff_a",
            "The coefficient 'a' used in cubic interpolation. Two common choice are -0.5 (in some cases of TensorFlow) and -0.75"
            " (in PyTorch). Check out Equation (4) in https://ieeexplore.ieee.org/document/1163711 for the details. "
            "This attribute is valid only if mode is \"cubic\".",
            AttributeProto::FLOAT,
            static_cast<float>(-0.75))
        .Attr(
            "exclude_outside",
            "If set to 1, the weight of sampling locations outside the tensor will be set to 0"
            " and the weight will be renormalized so that their sum is 1.0. The default value is 0.",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Attr(
            "coordinate_transformation_mode",
            Resize_ver19_attr_coordinate_transformation_mode_doc,
            AttributeProto::STRING,
            std::string("half_pixel"))
        .Attr(
            "nearest_mode",
            "Four modes: \"round_prefer_floor\" (default, as known as round half down), \"round_prefer_ceil\" (as known as round half up), \"floor\", \"ceil\". Only used by nearest interpolation. It indicates how to get \"nearest\" pixel in input tensor from x_original, so this attribute is valid only if \"mode\" is \"nearest\".",
            AttributeProto::STRING,
            std::string("round_prefer_floor"))
        .Attr(
            "extrapolation_value",
            "When coordinate_transformation_mode is \"tf_crop_and_resize\" and x_original is outside the range [0, length_original - 1], this value is used as the corresponding output value. Default is 0.0f.",
            AttributeProto::FLOAT,
            static_cast<float>(0))
        .Attr(
            "antialias",
            "If set to 1, \"linear\" and \"cubic\" interpolation modes will use an antialiasing filter when downscaling. "
            "Antialiasing is achieved by stretching the resampling filter by a factor max(1, 1 / scale), which means that when downsampling, more input pixels contribute to an output pixel.",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Attr(
            "axes",
            "If provided, it specifies a subset of axes that 'roi', 'scales' and 'sizes' refer to. "
            "If not provided, all axes are assumed [0, 1, ..., r-1], where r = rank(data). "
            "Non-specified dimensions are interpreted as non-resizable. "
            "Negative value means counting dimensions from the back. Accepted range is [-r, r-1], where r = rank(data). "
            "Behavior is undefined if an axis is repeated.",
            AttributeProto::INTS,
            false)
        .Attr(
            "keep_aspect_ratio_policy",
            Resize_ver19_attr_keep_aspect_ratio_policy_doc,
            AttributeProto::STRING,
            std::string("stretch"))
        .Input(0, "X", "N-D tensor", "T1", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "roi",
            "1-D tensor given as [start1, ..., startN, end1, ..., endN], where N is the rank of X or the length of axes, if provided. "
            "The RoIs' coordinates are normalized in the coordinate system of the input image. It only takes effect when coordinate_transformation_mode is \"tf_crop_and_resize\"",
            "T2",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            2,
            "scales",
            "The scale array along each dimension. It takes value greater than 0. If it's less than 1,"
            " it's sampling down, otherwise, it's upsampling. The number of elements of 'scales' should"
            " be the same as the rank of input 'X' or the length of 'axes', if provided. "
            "One of 'scales' and 'sizes' MUST be specified and it is an error if both are specified. If 'sizes' is needed, the user can use an empty string as the name of 'scales' in this operator's input list.",
            "tensor(float)",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            3,
            "sizes",
            "Target size of the output tensor. Its interpretation depends on the 'keep_aspect_ratio_policy' value."
            "The number of elements of 'sizes' should be the same as the"
            " rank of input 'X', or the length of 'axes', if provided. Only one of 'scales' and 'sizes' can be specified. ",
            "tensor(int64)",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(0, "Y", "N-D tensor after resizing", "T1", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .TypeConstraint(
            "T1",
            OpSchema::all_tensor_types_ir4(),
            "Constrain input 'X' and output 'Y' to all tensor types.")
        .TypeConstraint(
            "T2",
            {"tensor(float16)", "tensor(float)", "tensor(double)"},
            "Constrain roi type to float or double.")
        .SetDoc(Resize_ver19_doc)
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) { resizeShapeInference_opset18_to_19(ctx); }));

static const char* GridSample_ver20_doc = R"DOC(
Given an input `X` and a flow-field `grid`, computes the output `Y` using `X` values and pixel locations from the `grid`.
For spatial input `X` with shape (N, C, H, W), the `grid` will have shape (N, H_out, W_out, 2),
the output `Y` will have shape (N, C, H_out, W_out). For volumetric input `X` with shape (N, C, D, H, W),
the `grid` will have shape (N, D_out, H_out, W_out, 3), the output `Y` will have shape (N, C, D_out, H_out, W_out).
More generally, for an input `X` of rank r+2 with shape (N, C, d1, d2, ..., dr),
the `grid` will have shape (N, D1_out, D2_out, ..., Dr_out, r), the output `Y` will have shape (N, C, D1_out, D2_out, ..., Dr_out).

The tensor `X` contains values at centers of square pixels (voxels, etc) locations such as (n, c, d1_in, d2_in, ..., dr_in).
The (n, d1_out, d2_out, ..., dr_out, :) values from the tensor `grid` are the normalized positions for interpolating the values
at the (n, c, d1_out, d2_out, ..., dr_out) locations from the output tensor `Y` using a specified interpolation method (the mode)
and a padding mode (for `grid` positions falling outside the 2-dimensional image).

For example, the values in `grid[n, h_out, w_out, :]` are size-2 vectors specifying normalized positions in the 2-dimensional space of `X`.
They are used to interpolate output values of `Y[n, c, h_out, w_out]`.

The GridSample operator is often used in doing grid generator and sampler in the
[Spatial Transformer Networks](https://arxiv.org/abs/1506.02025).
See also in [torch.nn.functional.grid_sample](https://pytorch.org/docs/stable/generated/torch.nn.functional.grid_sample.html).
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    GridSample,
    20,
    OpSchema()
        .Attr(
            "mode",
            "Three interpolation modes: linear (default), nearest and cubic. "
            "The \"linear\" mode includes linear and N-linear interpolation modes depending on the number of spatial dimensions "
            "of the input tensor (i.e. linear for 1 spatial dimension, bilinear for 2 spatial dimensions, etc.). "
            "The \"cubic\" mode also includes N-cubic interpolation modes following the same rules. The \"nearest\" mode rounds "
            "to the nearest even index when the sampling point falls halfway between two indices.",
            AttributeProto::STRING,
            std::string("linear"))
        .Attr(
            "padding_mode",
            "Support padding modes for outside grid values: `zeros`(default), `border`, `reflection`. "
            "zeros: use 0 for out-of-bound grid locations, "
            "border: use border values for out-of-bound grid locations, "
            "reflection: use values at locations reflected by the border for out-of-bound grid locations. "
            "If index 0 represents the margin pixel, the reflected value at index -1 will be the same as the value at index 1. "
            "For location far away from the border, it will keep being reflected until becoming in bound. "
            "If pixel location x = -3.5 reflects by border -1 and becomes x' = 1.5, then reflects by border 1 and becomes x'' = 0.5.",
            AttributeProto::STRING,
            std::string("zeros"))
        .Attr(
            "align_corners",
            "If align_corners=1, the extrema (-1 and 1) are considered as referring to the center points of the input's corner pixels (voxels, etc.). "
            "If align_corners=0, they are instead considered as referring to the corner points of the input's corner pixels (voxels, etc.), "
            "making the sampling more resolution agnostic.",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Input(
            0,
            "X",
            "Input tensor of rank r+2 that has shape (N, C, D1, D2, ..., Dr), where N is the batch size, "
            "C is the number of channels, D1, D2, ..., Dr are the spatial dimensions.",
            "T1",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Input(
            1,
            "grid",
            "Input offset of shape (N, D1_out, D2_out, ..., Dr_out, r), where D1_out, D2_out, ..., "
            "Dr_out are the spatial dimensions of the grid and output, and r is the number of spatial dimensions. "
            "Grid specifies the sampling locations normalized by the input spatial dimensions. "
            "Therefore, it should have most values in the range of [-1, 1]. If the grid has values outside the range of [-1, 1], "
            "the corresponding outputs will be handled as defined by padding_mode. Following computer vision convention, "
            "the coordinates in the length-r location vector are listed from the innermost tensor dimension to the outermost, "
            "the opposite of regular tensor indexing.",
            "T2",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "Y",
            "Output tensor of rank r+2 that has shape (N, C, D1_out, D2_out, ..., Dr_out) of the sampled values. "
            "For integer input types, intermediate values are computed as floating point and cast to integer at the end.",
            "T1",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint(
            "T1",
            OpSchema::all_tensor_types(),
            "Constrain input `X` and output `Y` types to all tensor types.")
        .TypeConstraint(
            "T2",
            {"tensor(float16)", "tensor(float)", "tensor(double)"},
            "Constrain grid types to float tensors.")
        .SetDoc(GridSample_ver20_doc)
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) { gridSampleShapeInference(ctx); }));

static const char* AffineGrid_ver20_doc = R"DOC(
Generates a 2D or 3D flow field (sampling grid), given a batch of affine matrices theta
(https://pytorch.org/docs/stable/generated/torch.nn.functional.affine_grid.html).
An affine matrix `theta` is applied to a position tensor represented in its homogeneous expression. Here is an example in 3D:
```
[r00, r01, r02, t0]   [x]   [x']
[r10, r11, r12, t1] * [y] = [y']
[r20, r21, r22, t2]   [z]   [z']
[0,   0,   0,   1 ]   [1]   [1 ]
```
where `(x, y, z)` is the position in the original space, `(x', y', z')` is the position in the output space.
The last row is always `[0, 0, 0, 1]` and is not stored in the affine matrix. Therefore we have `theta` of shape `(N, 2, 3)` for 2D or `(N, 3, 4)` for 3D.

Input `size` is used to define grid of positions evenly spaced in the original 2D or 3D space, with dimensions ranging from `-1` to `1`.
The output `grid` contains positions in the output space.

When `align_corners=1`, consider `-1` and `1` to refer to the centers of the corner pixels (mark `v` in illustration).
```
v            v            v            v
|-------------------|------------------|
-1                  0                  1
```
When `align_corners=0`, consider `-1` and `1` to refer to the outer edge of the corner pixels.
```
    v        v         v         v
|------------------|-------------------|
-1                 0                   1
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    AffineGrid,
    20,
    OpSchema()
        .Attr(
            "align_corners",
            "if align_corners=1, consider -1 and 1 to refer to the centers of the corner pixels. "
            "if align_corners=0, consider -1 and 1 to refer to the outer edge the corner pixels.",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Input(
            0,
            "theta",
            "input batch of affine matrices with shape (N, 2, 3) for 2D or (N, 3, 4) for 3D",
            "T1",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            1,
            "size",
            "the target output image size (N, C, H, W) for 2D or (N, C, D, H, W) for 3D",
            "T2",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "grid",
            "output tensor of shape (N, H, W, 2) of 2D sample coordinates or (N, D, H, W, 3) of 3D sample coordinates.",
            "T1",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T1", OpSchema::all_float_types_ir4(), "Constrain grid types to float tensors.")
        .TypeConstraint("T2", {"tensor(int64)"}, "Constrain size's type to int64 tensors.")
        .SetDoc(AffineGrid_ver20_doc)
        .FunctionBody(R"ONNX(
        {
          # naming one: 1, one_f: 1.0, one_1d: [1], one_f_1d: [1.0]
          one = Constant <value_int: int=1> ()
          two = Constant <value_int: int=2> ()
          zero = Constant <value_int: int=0> ()
          four = Constant <value_int: int=4> ()
          one_1d = Constant <value_ints: ints = [1]> ()
          zero_1d = Constant <value_ints: ints = [0]> ()

          minus_one = Constant <value_int: int=-1> ()
          minus_one_f = CastLike (minus_one, theta)
          zero_f = CastLike (zero, theta)
          one_f = CastLike (one, theta)
          two_f = CastLike (two, theta)

          constant_align_corners = Constant <value_int: int=@align_corners> ()
          constant_align_corners_equal_zero = Equal (constant_align_corners, zero)

          size_ndim = Size (size)
          condition_is_2d = Equal (size_ndim, four)

          N, C, D, H, W = If (condition_is_2d) <
              then_branch = g1 () => (N_then, C_then, D_then, H_then, W_then) {
                  N_then, C_then, H_then, W_then = Split <num_outputs: int=4> (size)
                  D_then = Identity (one_1d)
              },
              else_branch = g2 () => (N_else, C_else, D_else, H_else, W_else) {
                  N_else, C_else, D_else, H_else, W_else = Split <num_outputs: int=5> (size)
              }
          >
          size_NCDHW = Concat <axis=0> (N, C, D, H, W)

          theta_3d = If (condition_is_2d) <
              then_branch = g3 () => (theta_then) { # theta: N by 2 by 3 => N by 3 by 4
                  # use of thetaN23 is a way to make shape inference happy when theta is N by 3 by 4.
                  gather_idx_6 = Constant <value_ints: ints = [0, 1, 2, 0, 1, 2]> ()
                  shape_23 = Constant <value_ints: ints = [2, 3]> ()
                  gather_idx_23 = Reshape (gather_idx_6, shape_23)
                  shape_N23 = Concat <axis=0>(N, shape_23)
                  gather_idx_N23 = Expand (gather_idx_23, shape_N23)
                  thetaN23 = GatherElements <axis=2> (theta, gather_idx_N23) # N by 2 by 3 => N by 3 by 2

                  r1, r2 = Split <axis: int=1, num_outputs: int=2> (thetaN23) # N by 1 by 3
                  r1_ = Squeeze (r1) # N by 3
                  r2_ = Squeeze (r2)
                  r11, r12, t1 = Split <axis: int=1, num_outputs: int=3> (r1_) # N by 1
                  r21, r22, t2 = Split <axis: int=1, num_outputs: int=3> (r2_)

                  r11_shape = Shape (r21)
                  float_zero_1d_ = ConstantOfShape (r11_shape) # N by 1
                  float_zero_1d = CastLike (float_zero_1d_, theta)
                  float_one_1d = Add (float_zero_1d, one_f) # N by 1

                  R1 = Concat <axis=1>(r11, r12, float_zero_1d, t1) # N by 4
                  R2 = Concat <axis=1>(r21, r22, float_zero_1d, t2)
                  R3 = Concat <axis=1>(float_zero_1d, float_zero_1d, float_one_1d, float_zero_1d)

                  R1_ = Unsqueeze (R1, one_1d) # N by 1 by 4
                  R2_ = Unsqueeze (R2, one_1d)
                  R3_ = Unsqueeze (R3, one_1d)
                  theta_then = Concat <axis=1> (R1_, R2_, R3_) # N by 3 by 4
                  # theta_then = Identity (theta)
              },
              else_branch = g4 () => (theta_else) {
                  theta_else = Identity (theta)
              }
          >

          two_1d = Constant <value_ints=[2]> ()
          three_1d = Constant <value_ints=[3]> ()
          five_1d = Constant <value_ints=[5]> ()
          constant_D_H_W_shape = Slice (size_NCDHW, two_1d, five_1d) # [N, C, D, H, W] => [D, H, W]
          zeros_D_H_W_ = ConstantOfShape (constant_D_H_W_shape)
          zeros_D_H_W = CastLike (zeros_D_H_W_, theta)
          ones_D_H_W = Add (zeros_D_H_W, one_f)

          D_float = CastLike (D, zero_f)
          H_float = CastLike (H, zero_f)
          W_float = CastLike (W, zero_f)
          start_d, step_d, start_h, step_h, start_w, step_w = If (constant_align_corners_equal_zero) <
              then_branch = h1 () => (start_d_then, step_d_then, start_h_then, step_h_then, start_w_then, step_w_then) { # => (float, float, float, float, float, float)
                  step_d_then = Div (two_f, D_float)
                  step_h_then = Div (two_f, H_float)
                  step_w_then = Div (two_f, W_float)

                  step_d_half = Div (step_d_then, two_f)
                  start_d_then = Add (minus_one_f, step_d_half)

                  step_h_half = Div (step_h_then, two_f)
                  start_h_then = Add (minus_one_f, step_h_half)

                  step_w_half = Div (step_w_then, two_f)
                  start_w_then = Add (minus_one_f, step_w_half)
              },
              else_branch = h2 () => (start_d_else, step_d_else, start_h_else, step_h_else, start_w_else, step_w_else) { # => (float, float, float, float, float, float)
                  D_float_nimus_one = Sub (D_float, one_f)
                  H_float_nimus_one = Sub (H_float, one_f)
                  W_float_nimus_one = Sub (W_float, one_f)
                  # avoid divide by 0
                  D_equals_one = Equal (D, one)
                  step_d_else = If (D_equals_one) <
                      then_branch = g5 () => (step_d_else_then) {
                          step_d_else_then = Identity (zero_f)
                      },
                      else_branch = g6 () => (step_d_else_else) {
                          step_d_else_else = Div (two_f, D_float_nimus_one)
                      }
                  >
                  step_h_else = Div (two_f, H_float_nimus_one)
                  step_w_else = Div (two_f, W_float_nimus_one)
                  start_d_else = Identity (minus_one_f)
                  start_h_else = Identity (minus_one_f)
                  start_w_else = Identity (minus_one_f)
              }
          >
          grid_w_steps_int = Range (zero, W, one)
          grid_w_steps_float = CastLike (grid_w_steps_int, step_w)
          grid_w_steps = Mul (grid_w_steps_float, step_w)
          grid_w_0 = Add (start_w, grid_w_steps)

          grid_h_steps_int = Range (zero, H, one)
          grid_h_steps_float = CastLike (grid_h_steps_int, step_h)
          grid_h_steps = Mul (grid_h_steps_float, step_h)
          grid_h_0 = Add (start_h, grid_h_steps)

          grid_d_steps_int = Range (zero, D, one)
          grid_d_steps_float = CastLike (grid_d_steps_int, step_d)
          grid_d_steps = Mul (grid_d_steps_float, step_d)
          grid_d_0 = Add (start_d, grid_d_steps)

          zeros_H_W_D = Transpose <perm = [1, 2, 0]> (zeros_D_H_W)
          grid_d_1 = Add (zeros_H_W_D, grid_d_0)
          grid_d = Transpose <perm = [2, 0, 1]> (grid_d_1)

          zeros_D_W_H = Transpose <perm = [0, 2, 1]> (zeros_D_H_W)
          grid_h_1 = Add (zeros_D_W_H, grid_h_0)
          grid_h = Transpose <perm = [0, 2, 1]> (grid_h_1)

          grid_w = Add (grid_w_0, zeros_D_H_W)

          grid_w_usqzed = Unsqueeze (grid_w, minus_one)
          grid_h_usqzed = Unsqueeze (grid_h, minus_one)
          grid_d_usqzed = Unsqueeze (grid_d, minus_one)
          ones_D_H_W_usqzed = Unsqueeze (ones_D_H_W, minus_one)
          original_grid = Concat <axis=-1> (grid_w_usqzed, grid_h_usqzed, grid_d_usqzed, ones_D_H_W_usqzed)

          constant_shape_DHW_4 = Constant <value_ints: ints = [-1, 4]> ()
          original_grid_DHW_4 = Reshape (original_grid, constant_shape_DHW_4)
          original_grid_4_DHW_ = Transpose (original_grid_DHW_4)

          original_grid_4_DHW = CastLike (original_grid_4_DHW_, theta_3d)
          grid_N_3_DHW = MatMul (theta_3d, original_grid_4_DHW)
          grid_N_DHW_3 = Transpose <perm = [0, 2, 1]> (grid_N_3_DHW)
          N_D_H_W_3 = Concat <axis=-1> (N, D, H, W, three_1d)
          grid_3d_else_ = Reshape (grid_N_DHW_3, N_D_H_W_3)
          grid_3d = CastLike (grid_3d_else_, theta_3d)

          # grid = Identity (grid_3d)
          grid = If (condition_is_2d) <
              then_branch = g1 () => (grid_then) { # [N, D=1, H, W, 3] => [N, H, W, 2]
                  grid_squeezed = Squeeze (grid_3d, one_1d)  # [N, H, W, 3]
                  grid_then = Slice (grid_squeezed, zero_1d, two_1d, three_1d) # [N, H, W, 2]
              },
              else_branch = g2 () => (grid_else) {
                  grid_else = Identity (grid_3d)
              }
          >
        }
        )ONNX")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (!hasNInputShapes(ctx, 1)) {
            return;
          }

          checkInputRank(ctx, 1, 1);

          bool found;
          TensorShapeProto size_proto = getShapeInput(ctx, 1, found);
          if (!found) {
            return;
          }

          const auto size_length = size_proto.dim_size();
          if (size_length != 4 && size_length != 5) {
            fail_shape_inference("Length of input 'size' is ", size_length, ". It must be 4 for 2D or 5 for 5D.");
          }

          auto* output_shape = ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();
          const auto& N = size_proto.dim(0);
          *output_shape->add_dim() = N;
          // const auto& C = size_proto.dim(1); // C is not used
          if (size_length == 4) {
            // 2D case: size shape (N, C, H, W), output shape (N, C, H, W, 2)
            const auto& H = size_proto.dim(2);
            const auto& W = size_proto.dim(3);
            *output_shape->add_dim() = H;
            *output_shape->add_dim() = W;
            output_shape->add_dim()->set_dim_value(2);
          } else if (size_length == 5) {
            // 3D case: size shape (N, C, D, H, W), output shape (N, C, D, H, W, 3)
            const auto& D = size_proto.dim(2);
            const auto& H = size_proto.dim(3);
            const auto& W = size_proto.dim(4);
            *output_shape->add_dim() = D;
            *output_shape->add_dim() = H;
            *output_shape->add_dim() = W;
            output_shape->add_dim()->set_dim_value(3);
          }
        }));

ONNX_OPERATOR_SET_SCHEMA(
    Identity,
    21,
    OpSchema()
        .SetDoc("Identity operator")
        .Input(0, "input", "Input tensor", "V", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Output(0, "output", "Tensor to copy input into.", "V", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .TypeConstraint(
            "V",
            []() {
              auto t = OpSchema::all_tensor_types_ir10();
              auto s = OpSchema::all_tensor_sequence_types();
              auto o = OpSchema::all_optional_types();
              t.insert(t.end(), s.begin(), s.end());
              t.insert(t.end(), o.begin(), o.end());
              return t;
            }(),
            "Constrain input and output types to all tensor, sequence, and optional types.")
        .TypeAndShapeInferenceFunction(propagateShapeAndTypeFromFirstInput));

static const char* Compress_ver11_doc = R"DOC(
    Selects slices from an input tensor along a given axis where condition evaluates to True for each axis index.
    In case axis is not provided, input is flattened before elements are selected.
    Compress behaves like numpy.compress: https://docs.scipy.org/doc/numpy/reference/generated/numpy.compress.html
    )DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Compress,
    11,
    OpSchema()
        .SetDoc(Compress_ver11_doc)
        .Attr(
            "axis",
            "(Optional) Axis along which to take slices. If not specified, "
            "input is flattened before elements being selected. Negative value means counting dimensions "
            "from the back. Accepted range is [-r, r-1] where r = rank(input).",
            AttributeProto::INT,
            OPTIONAL_VALUE)
        .Input(0, "input", "Tensor of rank r >= 1.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "condition",
            "Rank 1 tensor of booleans to indicate which slices or data elements to be selected. "
            "Its length can be less than the input length along the axis "
            "or the flattened input size if axis is not specified. "
            "In such cases data slices or elements exceeding the condition length are discarded.",
            "T1",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "output",
            "Tensor of rank r if axis is specified. Otherwise output is a Tensor of rank 1.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Constrain input and output types to all tensor types.")
        .TypeConstraint("T1", {"tensor(bool)"}, "Constrain to boolean tensors.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          auto axisAttr = ctx.getAttribute("axis");
          if (hasInputShape(ctx, 0)) {
            const TensorShapeProto& indices_shape = ctx.getInputType(0)->tensor_type().shape();
            int r = indices_shape.dim_size();
            if (r < 1) {
              fail_shape_inference("Indices tensor must have rank >= 1");
            }
            if (axisAttr) {
              int axis = static_cast<int>(axisAttr->i());
              if (axis < -r || axis >= r) {
                fail_shape_inference("'axis' must be in [-rank(indices), rank(indices)-1]");
              }
              if (axis < 0) {
                axis += r;
              }
              TensorShapeProto* shape = ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape();
              for (int i = 0; i < indices_shape.dim_size(); i++) {
                auto* dim = shape->add_dim();
                if (i != axis) {
                  *dim = indices_shape.dim(i);
                }
              }
            }
          }
          if (!axisAttr) {
            updateOutputShape(ctx, 0, {Dim()});
          }
        }));

static const char* OneHot_ver11_doc = R"DOC(
    Produces a one-hot tensor based on inputs.
    The locations represented by the index values in the 'indices' input tensor will have 'on_value'
    and the other locations will have 'off_value' in the output tensor, where 'on_value' and 'off_value'
    are specified as part of required input argument 'values', which is a two-element tensor of format
    [off_value, on_value]. The rank of the output tensor will be one greater than the rank of the
    input tensor. The additional dimension is for one-hot representation. The additional dimension will
    be inserted at the position specified by 'axis'. If 'axis' is not specified then then additional
    dimension will be inserted as the innermost dimension, i.e. axis=-1. The size of the additional
    dimension is specified by required scalar input 'depth'. The type of the output tensor is the same
    as the type of the 'values' input. Any entries in the 'indices' input tensor with values outside
    the range [-depth, depth-1] will result in one-hot representation with all 'off_value' values in the
    output tensor.

    when axis = 0:
    output[input[i, j, k], i, j, k] = 1 for all i, j, k and 0 otherwise.

    when axis = -1:
    output[i, j, k, input[i, j, k]] = 1 for all i, j, k and 0 otherwise.

)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    OneHot,
    11,
    OpSchema()
        .SetDoc(OneHot_ver11_doc)
        .Attr(
            "axis",
            "(Optional) Axis along which one-hot representation in added. Default: axis=-1. "
            "axis=-1 means that the additional dimension will be inserted as the "
            "innermost/last dimension in the output tensor. Negative value means counting dimensions "
            "from the back. Accepted range is [-r-1, r] where r = rank(indices).",
            AttributeProto::INT,
            static_cast<int64_t>(-1))
        .Input(
            0,
            "indices",
            "Input tensor containing indices. Any entries in the 'indices' input tensor with "
            "values outside the range [-depth, depth-1] will result in one-hot representation with all "
            "'off_value' values in the output tensor."
            "In case 'indices' is of non-integer type, the values will be casted to int64 before use.",
            "T1",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            1,
            "depth",
            "Scalar or Rank 1 tensor containing exactly one element, specifying the number of classes "
            "in one-hot tensor. This is also the size of the one-hot dimension (specified by 'axis' attribute) "
            "added on in the output tensor. The values in the 'indices' input tensor are expected to be "
            "in the range [-depth, depth-1]. "
            "In case 'depth' is of non-integer type, it will be casted to int64 before use.",
            "T2",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            2,
            "values",
            "Rank 1 tensor containing exactly two elements, in the format [off_value, on_value], "
            "where 'on_value' is the value used for filling locations specified in 'indices' input "
            "tensor, and 'off_value' is the value used for filling locations other than those specified "
            "in 'indices' input tensor. ",
            "T3",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "output",
            "Tensor of rank one greater than input tensor 'indices', i.e. rank(output) = rank(indices) + 1. "
            "The data type for the elements of the output tensor is the same as the type of input 'values' "
            "is used.",
            "T3",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .TypeConstraint("T1", OpSchema::all_numeric_types(), "Constrain input to only numeric types.")
        .TypeConstraint("T2", OpSchema::all_numeric_types(), "Constrain input to only numeric types.")
        .TypeConstraint("T3", OpSchema::all_tensor_types(), "Constrain to any tensor type.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          // Check that the node has three inputs.
          if (ctx.getNumInputs() != 3) {
            fail_type_inference("OneHot node must have three inputs.");
          }
          // Input 'depth' must be a scalar or a single-element vector.
          // TODO: Ideally to match spec for this input only Scalar should
          // be allowed. Making this change now can affect backward
          // compatibility for this op. Since this does not seem like a good
          // justification to update version for this op, allowing both scalar
          // and 1 element vector for now. In future when version update for
          // this op is done we should only allow scalar or change the spec to
          // allow both.
          if (hasInputShape(ctx, 1)) {
            auto& depth_shape = getInputShape(ctx, 1);
            if (depth_shape.dim_size() != 0 && depth_shape.dim_size() != 1) {
              fail_type_inference("Input 'depth' must be a scalar or rank 1 tensor.");
            }
            if (depth_shape.dim_size() == 1 && depth_shape.dim((int)0).has_dim_value() &&
                depth_shape.dim((int)0).dim_value() != 1) {
              fail_type_inference("Input 'depth' must have exactly one element.");
            }
          }
          // Input 'values' must be a two-element vector.
          if (hasInputShape(ctx, 2)) {
            auto& values_shape = getInputShape(ctx, 2);
            if (values_shape.dim_size() != 1) {
              fail_type_inference("Input 'values' must be rank 1 tensor.");
            }
            if (values_shape.dim((int)0).has_dim_value() && values_shape.dim((int)0).dim_value() != 2) {
              fail_type_inference("Input 'values' must have exactly two elements.");
            }
          }
          // Set output type to be the same as the third input, 'values'.
          propagateElemTypeFromInputToOutput(ctx, 2, 0);
          // Set the output shape, if input 0 (indices) shape is available.
          if (hasInputShape(ctx, 0)) {
            const TensorShapeProto& indices_shape = ctx.getInputType(0)->tensor_type().shape();
            int r = indices_shape.dim_size();
            if (r < 1) {
              fail_shape_inference("Indices tensor must have rank >= 1");
            }
            int out_rank = r + 1;
            int axis = static_cast<int>(getAttribute(ctx, "axis", -1));
            if (axis < -out_rank || axis >= out_rank) {
              fail_shape_inference("'axis' must be in [-rank(indices), rank(indices)-1]");
            }
            if (axis < 0) {
              axis += out_rank;
            }
            auto* output_shape = getOutputShape(ctx, 0);
            for (int i = 0; i < out_rank; ++i) {
              auto* dim = output_shape->add_dim();
              if (i < axis) {
                if (indices_shape.dim(i).has_dim_value()) {
                  dim->set_dim_value(indices_shape.dim(i).dim_value());
                } else if (indices_shape.dim(i).has_dim_param()) {
                  dim->set_dim_param(indices_shape.dim(i).dim_param());
                }
              } else if (i > axis) {
                if (indices_shape.dim(i - 1).has_dim_value()) {
                  dim->set_dim_value(indices_shape.dim(i - 1).dim_value());
                } else if (indices_shape.dim(i - 1).has_dim_param()) {
                  dim->set_dim_param(indices_shape.dim(i - 1).dim_param());
                }
              }
            }
          }
        }));

ONNX_OPERATOR_SET_SCHEMA(
    IsNaN,
    20,
    OpSchema()
        .SetDoc(R"DOC(Returns which elements of the input are NaN.)DOC")
        .Input(0, "X", "input", "T1", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
        .Output(0, "Y", "output", "T2", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
        .TypeConstraint("T1", OpSchema::all_float_types_ir9(), "Constrain input types to float tensors.")
        .TypeConstraint("T2", {"tensor(bool)"}, "Constrain output types to boolean tensors.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          updateOutputElemType(ctx, 0, TensorProto::BOOL);
          if (hasInputShape(ctx, 0)) {
            propagateShapeFromInputToOutput(ctx, 0, 0);
          }
        }));

ONNX_OPERATOR_SET_SCHEMA(
    IsInf,
    20,
    OpSchema()
        .SetDoc(R"DOC(Map infinity to true and other values to false.)DOC")
        .Input(0, "X", "input", "T1", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
        .Output(0, "Y", "output", "T2", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
        .Attr(
            "detect_positive",
            "(Optional) Whether map positive infinity to true. Default to 1 "
            "so that positive infinity induces true. Set this attribute to 0 "
            "if positive infinity should be mapped to false.",
            AttributeProto::INT,
            static_cast<int64_t>(1))
        .Attr(
            "detect_negative",
            "(Optional) Whether map negative infinity to true. Default to 1 "
            "so that negative infinity induces true. Set this attribute to 0 "
            "if negative infinity should be mapped to false.",
            AttributeProto::INT,
            static_cast<int64_t>(1))
        .TypeConstraint("T1", OpSchema::all_float_types_ir9(), "Constrain input types to float tensors.")
        .TypeConstraint("T2", {"tensor(bool)"}, "Constrain output types to boolean tensors.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          updateOutputElemType(ctx, 0, TensorProto::BOOL);
          if (hasInputShape(ctx, 0)) {
            propagateShapeFromInputToOutput(ctx, 0, 0);
          }
        }));

static const char* Where_ver16_doc = R"DOC(
Return elements, either from X or Y, depending on condition.
Where behaves like
[numpy.where](https://docs.scipy.org/doc/numpy/reference/generated/numpy.where.html)
with three parameters.

)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Where,
    16,
    OpSchema()
        .SetDoc(GET_OP_DOC_STR(std::string(Where_ver16_doc) + GenerateBroadcastingDocMul()))
        .Input(
            0,
            "condition",
            "When True (nonzero), yield X, otherwise yield Y",
            "B",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Input(
            1,
            "X",
            "values selected at indices where condition is True",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Input(
            2,
            "Y",
            "values selected at indices where condition is False",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Output(
            0,
            "output",
            "Tensor of shape equal to the broadcasted shape of condition, X, and Y.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("B", {"tensor(bool)"}, "Constrain to boolean tensors.")
        .TypeConstraint(
            "T",
            OpSchema::all_tensor_types_ir4(),
            "Constrain input and output types to all tensor types (including bfloat).")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 1, 0);
          if (hasNInputShapes(ctx, 3)) {
            std::vector<const TensorShapeProto*> shapes;
            shapes.push_back(&ctx.getInputType(0)->tensor_type().shape());
            shapes.push_back(&ctx.getInputType(1)->tensor_type().shape());
            shapes.push_back(&ctx.getInputType(2)->tensor_type().shape());
            multidirectionalBroadcastShapeInference(
                shapes, *ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape());
          }
        }));

ONNX_OPERATOR_SET_SCHEMA(
    NonZero,
    13,
    OpSchema()
        .SetDoc(NonZero_ver9_doc)
        .Input(0, "X", "input", "T", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
        .Output(0, "Y", "output", "tensor(int64)", OpSchema::Single, true, 1, OpSchema::NonDifferentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain to all tensor types.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          updateOutputElemType(ctx, 0, TensorProto::INT64);
          TensorShapeProto output_shape;
          auto* dim = output_shape.add_dim();
          if (hasInputShape(ctx, 0)) {
            const TensorShapeProto& input_shape = getInputShape(ctx, 0);
            dim->set_dim_value(input_shape.dim_size());
          }
          output_shape.add_dim();
          updateOutputShape(ctx, 0, output_shape);
        }));

static const char* ReverseSequence_ver10_doc = R"DOC(
Reverse batch of sequences having different lengths specified by `sequence_lens`.

For each slice i iterating on batch axis, the operator reverses the first sequence_lens[i] elements on time axis,
and copies elements whose index's beyond sequence_lens[i] to the output. So the output slice i contains reversed
sequences on the first sequence_lens[i] elements, then have original values copied for the other elements.

Example 1:
  input = [[0.0, 4.0, 8.0,  12.0],
           [1.0, 5.0, 9.0,  13.0],
           [2.0, 6.0, 10.0, 14.0],
           [3.0, 7.0, 11.0, 15.0]]
  sequence_lens = [4, 3, 2, 1]
  time_axis = 0
  batch_axis = 1

  output = [[3.0, 6.0, 9.0,  12.0],
            [2.0, 5.0, 8.0,  13.0],
            [1.0, 4.0, 10.0, 14.0],
            [0.0, 7.0, 11.0, 15.0]]

Example 2:
  input = [[0.0,  1.0,  2.0,  3.0 ],
           [4.0,  5.0,  6.0,  7.0 ],
           [8.0,  9.0,  10.0, 11.0],
           [12.0, 13.0, 14.0, 15.0]]
  sequence_lens = [1, 2, 3, 4]
  time_axis = 1
  batch_axis = 0

  output = [[0.0,  1.0,  2.0,  3.0 ],
            [5.0,  4.0,  6.0,  7.0 ],
            [10.0, 9.0,  8.0,  11.0],
            [15.0, 14.0, 13.0, 12.0]]
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    ReverseSequence,
    10,
    OpSchema()
        .SetDoc(ReverseSequence_ver10_doc)
        .Attr(
            "time_axis",
            "(Optional) Specify which axis is time axis. Must be one of 0 (default), or 1.",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Attr(
            "batch_axis",
            "(Optional) Specify which axis is batch axis. Must be one of 1 (default), or 0.",
            AttributeProto::INT,
            static_cast<int64_t>(1))
        .Input(0, "input", "Tensor of rank r >= 2.", "T", OpSchema::Single)
        .Input(
            1,
            "sequence_lens",
            "Tensor specifying lengths of the sequences in a batch. It has shape `[batch_size]`.",
            "tensor(int64)",
            OpSchema::Single)
        .Output(0, "Y", "Tensor with same shape of input.", "T", OpSchema::Single)
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Input and output types can be of any tensor type.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (!hasNInputShapes(ctx, 2)) {
            return;
          }

          auto& first_input_shape = getInputShape(ctx, 0);
          if (first_input_shape.dim_size() < 2) {
            fail_shape_inference("'input' must have rank >= 2");
          }
          auto& seq_len_input_shape = getInputShape(ctx, 1);
          if (seq_len_input_shape.dim_size() != 1) {
            fail_shape_inference("'sequence_lens' must have rank of 1");
          }

          propagateShapeFromInputToOutput(ctx, 0, 0);
        }));

static const char* Unique_ver11_doc = R"DOC(
Find the unique elements of a tensor. When an optional attribute 'axis' is provided, unique subtensors sliced along the 'axis' are returned.
Otherwise the input tensor is flattened and unique values of the flattened tensor are returned.

This operator returns the unique values or sliced unique subtensors of the input tensor and three optional outputs.
The first output tensor 'Y' contains all unique values or subtensors of the input.
The second optional output tensor 'indices' contains indices of 'Y' elements' first occurrence in 'X'.
The third optional output tensor 'inverse_indices' contains, for elements of 'X', its corresponding indices in 'Y'.
The fourth optional output tensor 'counts' contains the count of each element of 'Y' in the input.

Outputs are either sorted in ascending order or optionally in the order of the first occurrence of the values in the input.

https://docs.scipy.org/doc/numpy/reference/generated/numpy.unique.html

Example 1:
```
input_X = [2, 1, 1, 3, 4, 3]
attribute_sorted = 0
attribute_axis = None
output_Y = [2, 1, 3, 4]
output_indices = [0, 1, 3, 4]
output_inverse_indices = [0, 1, 1, 2, 3, 2]
output_counts = [1, 2, 2, 1]
```

Example 2:
```
input_X = [[1, 3], [2, 3]]
attribute_sorted = 1
attribute_axis = None
output_Y = [1, 2, 3]
output_indices = [0, 2, 1]
output_inverse_indices = [0, 2, 1, 2]
output_counts = [1, 1, 2]
```

Example 3:
```
input_X = [[1, 0, 0], [1, 0, 0], [2, 3, 4]]
attribute_sorted = 1
attribute_axis = 0
output_Y = [[1, 0, 0], [2, 3, 4]]
output_indices = [0, 2]
output_inverse_indices = [0, 0, 1]
output_counts = [2, 1]
```

Example 4:
```
input_x = [[[1., 1.], [0., 1.], [2., 1.], [0., 1.]],
            [[1., 1.], [0., 1.], [2., 1.], [0., 1.]]]
attribute_sorted = 1
attribute_axis = 1
```

intermediate data are presented below for better understanding:
there are 4 subtensors sliced along axis 1 of input_x (shape = (2, 4, 2)):
```
A: [[1, 1], [1, 1]],
   [[0, 1], [0, 1]],
   [[2, 1], [2, 1]],
   [[0, 1], [0, 1]].
```

there are 3 unique subtensors:
```
[[1, 1], [1, 1]],
[[0, 1], [0, 1]],
[[2, 1], [2, 1]].
```

sorted unique subtensors:
```
B: [[0, 1], [0, 1]],
   [[1, 1], [1, 1]],
   [[2, 1], [2, 1]].
```

output_Y is constructed from B:
```
[[[0. 1.], [1. 1.], [2. 1.]],
 [[0. 1.], [1. 1.], [2. 1.]]]
```

output_indices is to map from B to A:
```
[1, 0, 2]
```

output_inverse_indices is to map from A to B:
```
[1, 0, 2, 0]
```

output_counts:
```
[2, 1, 1]
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Unique,
    11,
    OpSchema()
        .SetDoc(Unique_ver11_doc)
        .Attr(
            "sorted",
            "(Optional) Whether to sort the unique elements in ascending order before returning as output. "
            "Must be one of 0, or 1 (default).",
            AttributeProto::INT,
            static_cast<int64_t>(1))
        .Attr(
            "axis",
            "(Optional) The dimension to apply unique. If not specified, the unique elements of the "
            "flattened input are returned. Negative value means counting dimensions "
            "from the back. Accepted range is [-r, r-1] where r = rank(input).",
            AttributeProto::INT,
            OPTIONAL_VALUE)
        .Input(
            0,
            "X",
            "A N-D input tensor that is to be processed.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "Y",
            "A tensor of the same type as 'X' "
            "containing all the unique values or subtensors sliced along a provided 'axis' in 'X', either sorted "
            "or maintained in the same order they occur in input 'X'",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            1,
            "indices",
            "A 1-D INT64 tensor "
            "containing indices of 'Y' elements' first occurrence in 'X'. "
            "When 'axis' is provided, it contains indices to subtensors in input 'X' on the 'axis'. "
            "When 'axis' is not provided, it contains indices to values in the flattened input tensor. ",
            "tensor(int64)",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            2,
            "inverse_indices",
            "A 1-D INT64 tensor "
            "containing, for elements of 'X', its corresponding indices in 'Y'. "
            "When 'axis' is provided, it contains indices to subtensors in output 'Y' on the 'axis'. "
            "When 'axis' is not provided, it contains indices to values in output 'Y'. ",
            "tensor(int64)",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            3,
            "counts",
            "A 1-D INT64 tensor containing "
            "the count of each element "
            "of 'Y' in input 'X'",
            "tensor(int64)",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Input can be of any tensor type.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          // Type inference
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          const TypeProto* xTensorProto = ctx.getInputType(0);
          TypeProto* yTensorProto = ctx.getOutputType(0);
          TypeProto* indicesTensorProto = nullptr;
          TypeProto* inverseIndicesTensorProto = nullptr;
          TypeProto* countsTensorProto = nullptr;

          // 'indices', 'inverse_indices', and 'counts' are 1-D tensors of
          // unknown dimension.
          // Shape inference will happen even in case of empty optional outputs,
          // graph-level shape inference should not propagate the shape downstream for empty optional outputs.
          auto num_outputs = ctx.getNumOutputs();
          if (num_outputs >= 2) {
            indicesTensorProto = ctx.getOutputType(1);
            updateOutputElemType(ctx, 1, TensorProto::INT64);
            indicesTensorProto->mutable_tensor_type()->mutable_shape()->add_dim();
          }

          if (num_outputs >= 3) {
            inverseIndicesTensorProto = ctx.getOutputType(2);
            updateOutputElemType(ctx, 2, TensorProto::INT64);
            inverseIndicesTensorProto->mutable_tensor_type()->mutable_shape()->add_dim();
          }

          if (num_outputs >= 4) {
            countsTensorProto = ctx.getOutputType(3);
            updateOutputElemType(ctx, 3, TensorProto::INT64);
            countsTensorProto->mutable_tensor_type()->mutable_shape()->add_dim();
          }

          auto axisAttr = ctx.getAttribute("axis");
          if (!axisAttr) {
            // 'axis' is not provided. Input 'X' is flattened.
            // 'Y' is a 1-D tensor of unknown dimension.
            yTensorProto->mutable_tensor_type()->mutable_shape()->add_dim();
          } else {
            // 'axis' is provided.
            int axis = static_cast<int>(axisAttr->i());
            if (!xTensorProto->tensor_type().has_shape()) {
              return;
            }
            const TensorShapeProto& input_shape = xTensorProto->tensor_type().shape();
            int rank = input_shape.dim_size();
            if (axis < 0)
              axis += rank;
            if (axis < 0 || axis >= rank) {
              fail_shape_inference("Invalid value for attribute axis");
            }
            // 'Y' has the same shape as 'X' except in the 'axis' dimension
            // which is unknown.
            for (int i = 0; i < input_shape.dim_size(); i++) {
              auto* dim = yTensorProto->mutable_tensor_type()->mutable_shape()->add_dim();
              if (i != axis) {
                *dim = input_shape.dim(i);
              }
            }
          }
        }));

static const char* GatherND_ver13_doc = R"DOC(
Given `data` tensor of rank `r` >= 1, `indices` tensor of rank `q` >= 1, and `batch_dims` integer `b`, this operator gathers
slices of `data` into an output tensor of rank `q + r - indices_shape[-1] - 1 - b`.

`indices` is an q-dimensional integer tensor, best thought of as a `(q-1)`-dimensional tensor of index-tuples into `data`,
where each element defines a slice of `data`

`batch_dims` (denoted as `b`) is an integer indicating the number of batch dimensions, i.e the leading `b` number of dimensions of
`data` tensor and `indices` are representing the batches, and the gather starts from the `b+1` dimension.

Some salient points about the inputs' rank and shape:

1) r >= 1 and q >= 1 are to be honored. There is no dependency condition to be met between ranks `r` and `q`

2) The first `b` dimensions of the shape of `indices` tensor and `data` tensor must be equal.

3) b < min(q, r) is to be honored.

4) The `indices_shape[-1]` should have a value between 1 (inclusive) and rank `r-b` (inclusive)

5) All values in `indices` are expected to be within bounds [-s, s-1] along axis of size `s` (i.e.) `-data_shape[i] <= indices[...,i] <= data_shape[i] - 1`.
   It is an error if any of the index values are out of bounds.

The output is computed as follows:

The output tensor is obtained by mapping each index-tuple in the `indices` tensor to the corresponding slice of the input `data`.

1) If `indices_shape[-1] > r-b` => error condition

2) If `indices_shape[-1] == r-b`, since the rank of `indices` is `q`, `indices` can be thought of as `N` `(q-b-1)`-dimensional tensors
   containing 1-D tensors of dimension `r-b`, where `N` is an integer equals to the product of 1 and all the elements in the batch dimensions
   of the indices_shape. Let us think of each such `r-b` ranked tensor as `indices_slice`. Each *scalar value* corresponding to `data[0:b-1,indices_slice]`
   is filled into the corresponding location of the `(q-b-1)`-dimensional tensor to form the `output` tensor (Example 1 below)

3) If `indices_shape[-1] < r-b`, since the rank of `indices` is `q`, `indices` can be thought of as `N` `(q-b-1)`-dimensional tensor
   containing 1-D tensors of dimension `< r-b`. Let us think of each such tensors as `indices_slice`. Each *tensor slice* corresponding
   to `data[0:b-1, indices_slice , :]` is filled into the corresponding location of the `(q-b-1)`-dimensional tensor
   to form the `output` tensor (Examples 2, 3, 4 and 5 below)

This operator is the inverse of `ScatterND`.

**Example 1**

```
batch_dims = 0
data    = [[0,1],[2,3]]   # data_shape    = [2, 2]
indices = [[0,0],[1,1]]   # indices_shape = [2, 2]
output  = [0,3]           # output_shape  = [2]
```

**Example 2**

```
batch_dims = 0
data    = [[0,1],[2,3]]  # data_shape    = [2, 2]
indices = [[1],[0]]      # indices_shape = [2, 1]
output  = [[2,3],[0,1]]  # output_shape  = [2, 2]
```

**Example 3**

```
batch_dims = 0
data    = [[[0,1],[2,3]],[[4,5],[6,7]]] # data_shape    = [2, 2, 2]
indices = [[0,1],[1,0]]                 # indices_shape = [2, 2]
output  = [[2,3],[4,5]]                 # output_shape  = [2, 2]
```

**Example 4**

```
batch_dims = 0
data    = [[[0,1],[2,3]],[[4,5],[6,7]]] # data_shape    = [2, 2, 2]
indices = [[[0,1]],[[1,0]]]             # indices_shape = [2, 1, 2]
output  = [[[2,3]],[[4,5]]]             # output_shape  = [2, 1, 2]
```

**Example 5**

```
batch_dims = 1
data    = [[[0,1],[2,3]],[[4,5],[6,7]]] # data_shape    = [2, 2, 2]
indices = [[1],[0]]                     # indices_shape = [2, 1]
output  = [[2,3],[4,5]]                 # output_shape  = [2, 2]
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    GatherND,
    13,
    OpSchema()
        .SetDoc(GatherND_ver13_doc)
        .Attr(
            "batch_dims",
            "The number of batch dimensions. The gather of indexing starts from dimension of data[batch_dims:]",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Input(0, "data", "Tensor of rank r >= 1.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Input(
            1,
            "indices",
            "Tensor of rank q >= 1. All index values are expected to be within bounds [-s, s-1] "
            "along axis of size s. It is an error if any of the index values are out of bounds.",
            "tensor(int64)",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "output",
            "Tensor of rank q + r - indices_shape[-1] - 1.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to any tensor type.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          // Type inference
          propagateElemTypeFromInputToOutput(ctx, 0, 0);

          // Shape inference
          if (!hasNInputShapes(ctx, 2)) {
            // cannot proceed with shape or rank inference
            return;
          }

          const auto& data_shape = ctx.getInputType(0)->tensor_type().shape();
          const auto data_rank = data_shape.dim_size();

          const auto& indices_shape = ctx.getInputType(1)->tensor_type().shape();
          const auto indices_rank = indices_shape.dim_size();

          int64_t batch_dims_data = getAttribute(ctx, "batch_dims", 0);
          if (data_rank < 1 || indices_rank < 1) {
            fail_shape_inference(
                "Both `data` and `indices` input tensors in GatherND op "
                "need to have rank larger than 0.");
          }

          // cannot ascertain if the input shapes are valid if shape of
          // `indices` is missing last dimension value so return at this point
          if (!indices_shape.dim(indices_rank - 1).has_dim_value()) {
            return;
          }

          const auto last_index_dimension = indices_shape.dim(indices_rank - 1).dim_value() + batch_dims_data;

          if (last_index_dimension > data_rank) {
            fail_shape_inference(
                "Last dimension of `indices` input tensor in GatherND op "
                "must not be larger than the rank of `data` tensor");
          }

          for (int i = 0; i < indices_rank - 1; ++i) {
            *ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->add_dim() = indices_shape.dim(i);
          }

          for (int i = static_cast<int>(last_index_dimension); i < data_rank; ++i) {
            *ctx.getOutputType(0)->mutable_tensor_type()->mutable_shape()->add_dim() = data_shape.dim(i);
          }
        }));

static const char* Pad_ver19_doc = R"DOC(
Given a tensor containing the data to be padded (`data`), a tensor containing the number of start and end pad values for axis (`pads`), (optionally) a `mode`, and (optionally) `constant_value`,
a padded tensor (`output`) is generated.

The three supported `modes` are (similar to corresponding modes supported by `numpy.pad`):

1) `constant`(default) - pads with a given constant value as specified by `constant_value` (which defaults to 0, empty string, or False)

2) `reflect` - pads with the reflection of the vector mirrored on the first and last values of the vector along each axis

3) `edge` - pads with the edge values of array

4) `wrap` - wrap-around padding as if the data tensor forms a torus


Example 1 (`constant` mode):

Insert 0 pads to the beginning of the second dimension.

```
data = [
    [1.0, 1.2],
    [2.3, 3.4],
    [4.5, 5.7],
]

pads = [0, 2, 0, 0]

mode = 'constant'

constant_value = 0.0

output = [
    [0.0, 0.0, 1.0, 1.2],
    [0.0, 0.0, 2.3, 3.4],
    [0.0, 0.0, 4.5, 5.7],
]
```

Example 2 (`reflect` mode):

```
data = [
    [1.0, 1.2],
    [2.3, 3.4],
    [4.5, 5.7],
]

pads = [0, 2, 0, 0]

mode = 'reflect'

output = [
    [1.0, 1.2, 1.0, 1.2],
    [2.3, 3.4, 2.3, 3.4],
    [4.5, 5.7, 4.5, 5.7],
]
```

Example 3 (`edge` mode):

```
data = [
    [1.0, 1.2],
    [2.3, 3.4],
    [4.5, 5.7],
]

pads = [0, 2, 0, 0]

mode = 'edge'

output = [
    [1.0, 1.0, 1.0, 1.2],
    [2.3, 2.3, 2.3, 3.4],
    [4.5, 4.5, 4.5, 5.7],
]
```

Example 4 (`wrap` mode):

```
data = [
    [1.0, 1.2],
    [2.3, 3.4],
    [4.5, 5.7],
]

pads = [2, 1, 1, 1]

mode = 'wrap'

output = [
    [3.4, 2.3, 3.4, 2.3],
    [5.7, 4.5, 5.7, 4.5],
    [1.2, 1.0, 1.2, 1.0],
    [3.4, 2.3, 3.4, 2.3],
    [5.7, 4.5, 5.7, 4.5],
    [1.2, 1.0, 1.2, 1.0],
]
```
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Pad,
    21,
    OpSchema().FillUsing(PadDocGenerator(
        Pad_ver19_doc,
        "Supported modes: `constant`(default), `reflect`, `edge`, `wrap`",
        OpSchema::all_tensor_types_ir10(),
        "Constrain input and output types to all tensor types up to IRv10.")));

static const char* Trilu_ver14_doc = R"DOC(
Given a 2-D matrix or batches of 2-D matrices, returns the upper or lower triangular part of the tensor(s).
The attribute "upper" determines whether the upper or lower part is retained. If set to true,
the upper triangular matrix is retained. Lower triangular matrix is retained otherwise.
Default value for the "upper" attribute is true.
Trilu takes one input tensor of shape [*, N, M], where * is zero or more batch dimensions. The upper triangular part consists
of the elements on and above the given diagonal (k). The lower triangular part consists of elements on and below the diagonal.
All other elements in the matrix are set to zero.
If k = 0, the triangular part on and above/below the main diagonal is retained.
If upper is set to true, a positive k retains the upper triangular matrix excluding the main diagonal and (k-1) diagonals above it.
A negative k value retains the main diagonal and |k| diagonals below it.
If upper is set to false, a positive k retains the lower triangular matrix including the main diagonal and k diagonals above it.
A negative k value excludes the main diagonal and (|k|-1) diagonals below it.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    Trilu,
    14,
    OpSchema()
        .SetDoc(Trilu_ver14_doc)
        .Attr(
            "upper",
            "Boolean. Indicates whether upper or lower part of matrix is retained. Default is true.",
            AttributeProto::INT,
            static_cast<int64_t>(1))
        .Input(
            0,
            "input",
            "Input tensor of rank 2 or higher.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Input(
            1,
            "k",
            "A 0-D tensor containing a single value corresponding to the number diagonals above or below the main diagonal to exclude or include. "
            "Default value is 0 if it's not specified.",
            "tensor(int64)",
            OpSchema::Optional,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(
            0,
            "output",
            "Output tensor of the same type and shape as the input tensor.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to all tensor types.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          // Type inference
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          // Shape inference needs the input data shape
          if (hasInputShape(ctx, 0)) {
            const TensorShapeProto& input_shape = ctx.getInputType(0)->tensor_type().shape();
            const int rank = static_cast<int>(input_shape.dim_size());
            if (rank < 2) {
              fail_shape_inference("Input rank must be >= 2.")
            }
            propagateShapeFromInputToOutput(ctx, 0, 0);
          }
        }));

static const char* CenterCropPad_ver18_doc = R"DOC(
Center crop or pad an input to given dimensions.

The crop/pad dimensions can be specified for a subset of the `axes`. Non-specified dimensions will not be
cropped or padded.

If the input dimensions are bigger than the crop shape, a centered cropping window is extracted from the input.
If the input dimensions are smaller than the crop shape, the input is padded on each side equally,
so that the input is centered in the output.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    CenterCropPad,
    18,
    OpSchema()
        .SetDoc(CenterCropPad_ver18_doc)
        .Input(
            0,
            "input_data",
            "Input to extract the centered crop from.",
            "T",
            OpSchema::Single,
            true,
            1,
            OpSchema::Differentiable)
        .Input(
            1,
            "shape",
            "1-D tensor representing the cropping window dimensions.",
            "Tind",
            OpSchema::Single,
            true,
            1,
            OpSchema::NonDifferentiable)
        .Output(0, "output_data", "Output data.", "T", OpSchema::Single, true, 1, OpSchema::Differentiable)
        .Attr(
            "axes",
            "If provided, it specifies a subset of axes that 'shape' refer to. "
            "If not provided, all axes are assumed [0, 1, ..., r-1], where r = rank(data). "
            "Negative value means counting dimensions from the back. Accepted range is [-r, r-1], where r = rank(data). "
            "Behavior is undefined if an axis is repeated.",
            AttributeProto::INTS,
            OPTIONAL_VALUE)
        .TypeConstraint("T", OpSchema::all_tensor_types_ir4(), "Constrain input and output types to all tensor types.")
        .TypeConstraint("Tind", {"tensor(int32)", "tensor(int64)"}, "Constrain indices to integer types")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          if (ctx.getNumInputs() != 2) {
            fail_type_inference("CenterCropPad op must have 2 inputs.");
          }
          propagateElemTypeFromInputToOutput(ctx, 0, 0);
          if (!hasNInputShapes(ctx, 1)) {
            return;
          }
          // Shape Inference if shape is initializer
          const TensorProto* cropShapeInitializer = ctx.getInputData(1);
          if (!cropShapeInitializer) {
            return;
          }

          // don't know data_type - can't proceed
          if (!cropShapeInitializer->has_data_type())
            return;

          const auto& input_shape = ctx.getInputType(0)->tensor_type().shape();
          const int64_t input_rank = input_shape.dim_size();

          std::vector<int64_t> shape;
          if (cropShapeInitializer->data_type() == TensorProto::INT64) {
            const auto& data = ParseData<int64_t>(cropShapeInitializer);
            shape.insert(shape.end(), data.begin(), data.end());
          } else if (cropShapeInitializer->data_type() == TensorProto::INT32) {
            const auto& data = ParseData<int32_t>(cropShapeInitializer);
            shape.insert(shape.end(), data.begin(), data.end());
          } else {
            // unaccepted data type
            fail_shape_inference("`shape` only supports `int32_t` or `int64_t` inputs");
          }

          auto axes_attr = ctx.getAttribute("axes");
          std::vector<int64_t> axes;
          if (axes_attr) {
            axes = RetrieveValues<int64_t>(*axes_attr);
            checkAxesRange(axes, input_rank);
            adjustNegativeAxes(axes, input_rank);
            checkDuplicateAxes(axes, input_rank);
          } else {
            axes.resize(input_rank);
            std::iota(axes.begin(), axes.end(), 0);
          }

          if (shape.size() != axes.size()) {
            fail_shape_inference(
                "Number of elements of input 'shape' (",
                shape.size(),
                ") does not match the number of axes (",
                axes.size(),
                ").");
          }

          // Populating default dims
          std::vector<TensorShapeProto_Dimension*> out_dims(input_rank);
          auto* output_shape = getOutputShape(ctx, 0);
          for (int i = 0; i < input_rank; ++i) {
            out_dims[i] = output_shape->add_dim();
            const auto& input_dim = input_shape.dim(i);
            if (input_dim.has_dim_value()) {
              out_dims[i]->set_dim_value(input_dim.dim_value());
            } else if (input_dim.has_dim_param()) {
              out_dims[i]->set_dim_param(input_dim.dim_param());
            }
          }
          int j = 0;
          for (int axis : axes) {
            out_dims[axis]->set_dim_value(shape[j++]);
          }
        })
        .SetContextDependentFunctionBodyBuilder([](const FunctionBodyBuildContext& ctx,
                                                   const OpSchema& schema,
                                                   FunctionProto& functionProto) {
          FunctionBuilder builder(functionProto);
          builder.Const("k2", std::vector<int64_t>{2});

          auto axes_attr = ctx.getAttribute("axes");
          if (axes_attr) { // axes provided, need to work on a subset of dimensions
            builder.Add("axes_input = Constant <value_ints : ints = @axes>()");
            builder.Add("x_shape_alldims = Shape (input_data)").Add("x_shape = Gather (x_shape_alldims, axes_input)");
          } else { // axes not provided, assuming all dims
            builder.Add("x_shape = Shape (input_data)");
          }

          // First: Pad step
          builder.Add("padded_sh = Max(x_shape, shape)")
              .Add("pad_amount = Sub(padded_sh, x_shape)")
              .Add("pad_amount_left = Div(pad_amount, k2)")
              .Add("pad_amount_right = Sub(pad_amount, pad_amount_left)")
              .Add("pads = Concat <axis = 0> (pad_amount_left, pad_amount_right)");
          if (axes_attr)
            builder.Add("padded_input = Pad (input_data, pads, , axes_input)");
          else
            builder.Add("padded_input = Pad (input_data, pads)");

          // Second: Slice step
          if (axes_attr) {
            builder.Add("x_shape_alldims2 = Shape (padded_input)")
                .Add("x_shape2 = Gather (x_shape_alldims2, axes_input)");
          } else {
            builder.Add("x_shape2 = Shape (padded_input)");
          }

          builder.Add("sh_diff = Sub (x_shape2, shape)")
              .Add("start_dims = Div (sh_diff, k2)")
              .Add("end_dims = Add (start_dims, shape)");
          if (axes_attr)
            builder.Add("output_data = Slice (padded_input, start_dims, end_dims, axes_input)");
          else
            builder.Add("output_data = Slice (padded_input, start_dims, end_dims)");

          schema.BuildFunction(functionProto);
          return true;
        }));

} // namespace ONNX_NAMESPACE