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

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

#include <algorithm>
#include <numeric>

#include "onnx/defs/function.h"
#include "onnx/defs/schema.h"

namespace ONNX_NAMESPACE {

static const char* SequenceEmpty_ver11_doc = R"DOC(
Construct an empty tensor sequence, with given data type.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    SequenceEmpty,
    11,
    OpSchema()
        .SetDoc(SequenceEmpty_ver11_doc)
        .Attr(
            "dtype",
            "(Optional) The data type of the tensors in the output sequence. "
            "The default type is 'float'.",
            AttributeProto::INT,
            OPTIONAL_VALUE)
        .Output(0, "output", "Empty sequence.", "S")
        .TypeConstraint("S", OpSchema::all_tensor_sequence_types(), "Constrain output types to any tensor type.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          const auto* attr_proto = ctx.getAttribute("dtype");
          auto elem_type = TensorProto::FLOAT;
          if (nullptr != attr_proto) {
            if (!attr_proto->has_i()) {
              fail_type_inference("Attribute dtype should be of integer type and specify a type.");
            }
            auto attr_value = attr_proto->i();
            elem_type = static_cast<TensorProto_DataType>(attr_value);
          }
          ctx.getOutputType(0)->mutable_sequence_type()->mutable_elem_type()->mutable_tensor_type()->set_elem_type(
              elem_type);
        }));

static const char* SequenceConstruct_ver11_doc = R"DOC(
Construct a tensor sequence containing 'inputs' tensors.
All tensors in 'inputs' must have the same data type.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    SequenceConstruct,
    11,
    OpSchema()
        .SetDoc(SequenceConstruct_ver11_doc)
        .Input(0, "inputs", "Tensors.", "T", OpSchema::Variadic)
        .Output(0, "output_sequence", "Sequence enclosing the input tensors.", "S")
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Constrain input types to any tensor type.")
        .TypeConstraint("S", OpSchema::all_tensor_sequence_types(), "Constrain output types to any tensor type.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          const size_t numInputs = ctx.getNumInputs();
          if (numInputs < 1) {
            fail_type_inference("SequenceConstruct is expected to have at least 1 input.");
          }

          std::vector<int> input_elem_types;
          input_elem_types.reserve(numInputs);
          for (size_t i = 0; i < numInputs; ++i) {
            auto input_type = ctx.getInputType(i);
            if (nullptr == input_type) {
              fail_type_inference("Input type for input at index ", i, " is null. Type info is expected.");
            }
            input_elem_types.emplace_back(input_type->tensor_type().elem_type());
          }
          if (std::adjacent_find(input_elem_types.begin(), input_elem_types.end(), std::not_equal_to<int>()) !=
              input_elem_types.end()) {
            // not all input elem types are the same.
            fail_type_inference("Element type of inputs are expected to be the same.");
          }

          auto* output_tensor_type =
              ctx.getOutputType(0)->mutable_sequence_type()->mutable_elem_type()->mutable_tensor_type();

          output_tensor_type->set_elem_type(static_cast<TensorProto_DataType>(input_elem_types[0]));

          if (!hasNInputShapes(ctx, static_cast<int>(numInputs))) {
            return;
          }

          *(output_tensor_type->mutable_shape()) = ctx.getInputType(0)->tensor_type().shape();

          for (size_t i = 1; i < numInputs; ++i) {
            const auto& input_shape = ctx.getInputType(i)->tensor_type().shape();
            UnionShapeInfo(input_shape, *output_tensor_type);
          }
        }));

static const char* SequenceInsert_ver11_doc = R"DOC(
Outputs a tensor sequence that inserts 'tensor' into 'input_sequence' at 'position'.
'tensor' must have the same data type as 'input_sequence'.
Accepted range for 'position' is in `[-n, n]`, where `n` is the number of tensors in 'input_sequence'.
Negative value means counting positions from the back.
'position' is optional, by default it inserts 'tensor' to the back of 'input_sequence'.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    SequenceInsert,
    11,
    OpSchema()
        .SetDoc(SequenceInsert_ver11_doc)
        .Input(0, "input_sequence", "Input sequence.", "S")
        .Input(1, "tensor", "Input tensor to be inserted into the input sequence.", "T")
        .Input(
            2,
            "position",
            "Position in the sequence where the new tensor is inserted. "
            "It is optional and default is to insert to the back of the sequence. "
            "Negative value means counting positions from the back. "
            "Accepted range in `[-n, n]`, "
            "where `n` is the number of tensors in 'input_sequence'. "
            "It is an error if any of the index values are out of bounds. "
            "It must be a scalar(tensor of empty shape).",
            "I",
            OpSchema::Optional)
        .Output(0, "output_sequence", "Output sequence that contains the inserted tensor at given position.", "S")
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Constrain to any tensor type.")
        .TypeConstraint("S", OpSchema::all_tensor_sequence_types(), "Constrain to any tensor type.")
        .TypeConstraint(
            "I",
            {"tensor(int32)", "tensor(int64)"},
            "Constrain position to integral tensor. It must be a scalar(tensor of empty shape).")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          const auto input0_type = ctx.getInputType(0);
          const auto input1_type = ctx.getInputType(1);
          if (nullptr == input0_type || nullptr == input1_type) {
            fail_type_inference("Input Sequence and Tensor are expected to have type info. Current type is null.");
          }
          const auto seq_elem_type = input0_type->sequence_type().elem_type().tensor_type().elem_type();
          const auto tensor_elem_type = input1_type->tensor_type().elem_type();
          if (seq_elem_type != tensor_elem_type) {
            fail_type_inference(
                "Input Sequence and Tensor are expected to have the same elem type. Sequence=",
                seq_elem_type,
                " Tensor=",
                tensor_elem_type);
          }

          auto* output_tensor_type =
              ctx.getOutputType(0)->mutable_sequence_type()->mutable_elem_type()->mutable_tensor_type();
          output_tensor_type->set_elem_type(seq_elem_type);

          if (!hasNInputShapes(ctx, 2)) {
            return;
          }

          *(output_tensor_type->mutable_shape()) = input0_type->sequence_type().elem_type().tensor_type().shape();

          UnionShapeInfo(input1_type->tensor_type().shape(), *output_tensor_type);
        }));

static const char* SequenceAt_ver11_doc = R"DOC(
Outputs a tensor copy from the tensor at 'position' in 'input_sequence'.
Accepted range for 'position' is in `[-n, n - 1]`, where `n` is the number of tensors in 'input_sequence'.
Negative value means counting positions from the back.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    SequenceAt,
    11,
    OpSchema()
        .SetDoc(SequenceAt_ver11_doc)
        .Input(0, "input_sequence", "Input sequence.", "S")
        .Input(
            1,
            "position",
            "Position of the tensor in the sequence. "
            "Negative value means counting positions from the back. "
            "Accepted range in `[-n, n - 1]`, "
            "where `n` is the number of tensors in 'input_sequence'. "
            "It is an error if any of the index values are out of bounds. "
            "It must be a scalar(tensor of empty shape).",
            "I")
        .Output(0, "tensor", "Output tensor at the specified position in the input sequence.", "T")
        .TypeConstraint("S", OpSchema::all_tensor_sequence_types(), "Constrain to any tensor type.")
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Constrain to any tensor type.")
        .TypeConstraint(
            "I",
            {"tensor(int32)", "tensor(int64)"},
            "Constrain position to integral tensor. It must be a scalar(tensor of empty shape).")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          const auto input0_type = ctx.getInputType(0);
          if (nullptr == input0_type) {
            fail_type_inference("Input type for input at index 0 is null. Type info is expected.")
          }
          ctx.getOutputType(0)->CopyFrom(input0_type->sequence_type().elem_type());
        }));

static const char* SequenceErase_ver11_doc = R"DOC(
Outputs a tensor sequence that removes the tensor at 'position' from 'input_sequence'.
Accepted range for 'position' is in `[-n, n - 1]`, where `n` is the number of tensors in 'input_sequence'.
Negative value means counting positions from the back.
'position' is optional, by default it erases the last tensor from 'input_sequence'.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    SequenceErase,
    11,
    OpSchema()
        .SetDoc(SequenceErase_ver11_doc)
        .Input(0, "input_sequence", "Input sequence.", "S")
        .Input(
            1,
            "position",
            "Position of the tensor in the sequence. "
            "Negative value means counting positions from the back. "
            "Accepted range in `[-n, n - 1]`, "
            "where `n` is the number of tensors in 'input_sequence'. "
            "It is an error if any of the index values are out of bounds. "
            "It must be a scalar(tensor of empty shape).",
            "I",
            OpSchema::Optional)
        .Output(0, "output_sequence", "Output sequence that has the tensor at the specified position removed.", "S")
        .TypeConstraint("S", OpSchema::all_tensor_sequence_types(), "Constrain to any tensor type.")
        .TypeConstraint(
            "I",
            {"tensor(int32)", "tensor(int64)"},
            "Constrain position to integral tensor. It must be a scalar(tensor of empty shape).")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          const auto input0_type = ctx.getInputType(0);
          if (nullptr == input0_type) {
            fail_type_inference("Input type for input at index 0 is null. Type info is expected.")
          }
          ctx.getOutputType(0)->CopyFrom(*input0_type);
        }));

static const char* SequenceLength_ver11_doc = R"DOC(
Produces a scalar(tensor of empty shape) containing the number of tensors in 'input_sequence'.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    SequenceLength,
    11,
    OpSchema()
        .SetDoc(SequenceLength_ver11_doc)
        .Input(0, "input_sequence", "Input sequence.", "S")
        .Output(0, "length", "Length of input sequence. It must be a scalar(tensor of empty shape).", "I")
        .TypeConstraint("S", OpSchema::all_tensor_sequence_types(), "Constrain to any tensor type.")
        .TypeConstraint(
            "I",
            {"tensor(int64)"},
            "Constrain output to integral tensor. It must be a scalar(tensor of empty shape).")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          auto* output_tensor_type = ctx.getOutputType(0)->mutable_tensor_type();
          output_tensor_type->set_elem_type(TensorProto::INT64);
          output_tensor_type->mutable_shape()->Clear();
        }));

// Updated operators that consume/produce sequence of tensors.

static const char* SplitToSequence_ver11_doc =
    R"DOC(
Split a tensor into a sequence of tensors, along the specified 'axis'.
Lengths of the parts can be specified using the optional argument 'split'.
If the argument `split' is not specified, a default scalar value of 1
is used as the value of `split'.
'split' must contain only positive numbers.
'split' is either a scalar (tensor of empty shape), or a 1-D tensor.
If 'split' is a scalar, then 'input' will be split into chunks all of size 'split'
if possible. The last chunk alone may be smaller than 'split' if the 'input' size
along the given axis 'axis' is not divisible by 'split'.
If 'split' is a 1-dimensional tensor, the input tensor is split into 'size(split)' chunks,
with lengths of the parts on 'axis' specified in 'split'. In this scenario, the sum of entries
in 'split' must be equal to the dimension size of input tensor on 'axis'.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    SplitToSequence,
    11,
    OpSchema()
        .Input(0, "input", "The tensor to split", "T")
        .Input(
            1,
            "split",
            "Length of each output. "
            "It can be either a scalar(tensor of empty shape), or a 1-D tensor. All values must be >= 0. ",
            "I",
            OpSchema::Optional)
        .Output(0, "output_sequence", "One or more outputs forming a sequence of tensors after splitting", "S")
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Constrain input types to all tensor types.")
        .TypeConstraint("I", {"tensor(int32)", "tensor(int64)"}, "Constrain split size to integral tensor.")
        .TypeConstraint("S", OpSchema::all_tensor_sequence_types(), "Constrain 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].",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .Attr(
            "keepdims",
            "Keep the split dimension or not. Default 1, which means we keep split dimension. "
            "If input 'split' is specified, this attribute is ignored.",
            AttributeProto::INT,
            static_cast<int64_t>(1))
        .SetDoc(SplitToSequence_ver11_doc)
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          const auto input0_type = ctx.getInputType(0);
          if (nullptr == input0_type) {
            fail_type_inference("Input type for input at index 0 is null. Type info is expected.")
          }
          ctx.getOutputType(0)->mutable_sequence_type()->mutable_elem_type()->mutable_tensor_type()->set_elem_type(
              input0_type->tensor_type().elem_type());

          if (!hasInputShape(ctx, 0)) {
            return;
          }

          const auto& inputShape = input0_type->tensor_type().shape();

          int r = inputShape.dim_size();
          int axis = static_cast<int>(getAttribute(ctx, "axis", 0));
          if (axis < -r || axis > r - 1) {
            fail_shape_inference("Invalid value of attribute 'axis'. Rank=", r, " Value=", axis);
          }
          if (axis < 0) {
            axis += r;
          }

          size_t num_inputs = ctx.getNumInputs();
          int64_t splitSize = 1;
          int64_t keepdims = 1;
          if (num_inputs == 1) {
            // input split is omitted, default to split by 1.
            auto attr_proto = ctx.getAttribute("keepdims");
            if (attr_proto) {
              keepdims = attr_proto->i();
            }
          } else {
            splitSize = [&]() -> int64_t {
              // Need input split shape info and initializer data to infer split sizes.
              if (!hasInputShape(ctx, 1)) {
                return -1;
              }
              const TensorProto* splitInitializer = ctx.getInputData(1);
              if (nullptr == splitInitializer || !splitInitializer->has_data_type()) {
                return -1;
              }

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

              if (splitSizes.size() == 0) {
                fail_shape_inference("Input 'split' can not be empty.");
              }

              const auto& splitDim = inputShape.dim(axis);
              if (!splitDim.has_dim_value()) {
                // Unable to verify nor infer exact split dimension size.
                return -1;
              }

              int64_t splitDimValue = splitDim.dim_value();
              const auto& splitShape = getInputShape(ctx, 1);
              if (splitShape.dim_size() == 0) {
                // split is scalar
                if (splitDimValue % splitSizes[0] == 0) {
                  // all output chunks have the same shape, assign that to output sequence shape.
                  return splitSizes[0];
                }
                return -1;
              } else {
                // split is 1-D tensor
                int64_t splitSizesSum = std::accumulate(splitSizes.begin(), splitSizes.end(), (int64_t)0);
                if (splitDimValue != splitSizesSum) {
                  fail_shape_inference(
                      "Sum of split values not equal to 'input' dim size on 'axis'. 'axis' dim size=",
                      splitDimValue,
                      " sum of split values=",
                      splitSizesSum);
                }
                if (std::adjacent_find(splitSizes.begin(), splitSizes.end(), std::not_equal_to<int64_t>()) ==
                    splitSizes.end()) {
                  // all split sizes are the same.
                  return splitSizes[0];
                }
                return -1;
              }
            }();
          }

          if (keepdims) {
            auto* outputShape = ctx.getOutputType(0)
                                    ->mutable_sequence_type()
                                    ->mutable_elem_type()
                                    ->mutable_tensor_type()
                                    ->mutable_shape();
            *outputShape = inputShape;
            auto* dim = outputShape->mutable_dim(axis);
            // Tensors in sequence could not have different shapes explicitly.
            // Only assign dim_value when all chunks have the same shape.
            if (splitSize > 0) {
              dim->set_dim_value(splitSize);
            } else {
              dim->clear_dim_value();
              dim->clear_dim_param();
            }
          } else {
            TensorShapeProto* outputShape = ctx.getOutputType(0)
                                                ->mutable_sequence_type()
                                                ->mutable_elem_type()
                                                ->mutable_tensor_type()
                                                ->mutable_shape();
            for (int i = 0; i < inputShape.dim_size(); ++i) {
              if (i != axis) {
                auto* dim = outputShape->add_dim();
                dim->CopyFrom(inputShape.dim(i));
              }
            }
          }
        }));

static const char* ConcatFromSequence_ver11_doc = R"DOC(
Concatenate a sequence 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.
By default 'new_axis' is 0, the behavior is similar to numpy.concatenate.
When 'new_axis' is 1, the behavior is similar to numpy.stack.
)DOC";

ONNX_OPERATOR_SET_SCHEMA(
    ConcatFromSequence,
    11,
    OpSchema()
        .Attr(
            "axis",
            "Which axis to concat on. Accepted range in `[-r, r - 1]`, "
            "where `r` is the rank of input tensors. "
            "When `new_axis` is 1, accepted range is `[-r - 1, r]`. ",
            AttributeProto::INT)
        .Attr(
            "new_axis",
            "Insert and concatenate on a new axis or not, "
            "default 0 means do not insert new axis.",
            AttributeProto::INT,
            static_cast<int64_t>(0))
        .SetDoc(ConcatFromSequence_ver11_doc)
        .Input(0, "input_sequence", "Sequence of tensors for concatenation", "S")
        .Output(0, "concat_result", "Concatenated tensor", "T")
        .TypeConstraint("S", OpSchema::all_tensor_sequence_types(), "Constrain input types to any tensor type.")
        .TypeConstraint("T", OpSchema::all_tensor_types(), "Constrain output types to any tensor type.")
        .TypeAndShapeInferenceFunction([](InferenceContext& ctx) {
          const auto input0_type = ctx.getInputType(0);
          if (nullptr == input0_type) {
            fail_type_inference("Input type for input at index 0 is null. Type info is expected.")
          }
          auto elem_type = input0_type->sequence_type().elem_type().tensor_type().elem_type();
          ctx.getOutputType(0)->mutable_tensor_type()->set_elem_type(elem_type);

          if (!hasInputShape(ctx, 0)) {
            return;
          }

          auto axis_attr = ctx.getAttribute("axis");
          if (!axis_attr) {
            fail_shape_inference("Required attribute axis is missing");
          }
          int axis = static_cast<int>(axis_attr->i());

          int new_axis = 0;
          auto new_axis_attr = ctx.getAttribute("new_axis");
          if (new_axis_attr) {
            new_axis = static_cast<int>(new_axis_attr->i());
          }

          const auto& input_shape = ctx.getInputType(0)->sequence_type().elem_type().tensor_type().shape();
          auto rank = input_shape.dim_size();
          if (1 != new_axis && 0 != new_axis) {
            fail_shape_inference("new_axis must be either 0 or 1");
          }

          auto upper_bound = 1 == new_axis ? rank : rank - 1;
          auto lower_bound = 1 == new_axis ? -rank - 1 : -rank;

          if (axis < lower_bound || axis > upper_bound) {
            fail_shape_inference(
                "Invalid value of attribute 'axis'. Accepted range=[",
                lower_bound,
                ", ",
                upper_bound,
                "], Value=",
                axis);
          }

          if (axis < 0) {
            axis += (upper_bound + 1);
          }

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

          for (int i = 0; i <= upper_bound; ++i) {
            output_shape->add_dim();
            if (i != axis) {
              output_shape->mutable_dim(i)->CopyFrom(input_shape.dim((i > axis && new_axis) ? i - 1 : i));
            }
          }
        }));

static const char* SequenceMap_ver17_doc = R"DOC(
Applies a sub-graph to each sample in the input sequence(s).

Inputs can be either tensors or sequences, with the exception of the first input which must
be a sequence. The length of the first input sequence will determine the number of samples in the
outputs. Any other sequence inputs should have the same number of samples. The number of inputs
and outputs, should match the one of the subgraph.

For each i-th element in the output, a sample will be extracted from the input sequence(s) at
the i-th position and the sub-graph will be applied to it.
The outputs will contain the outputs of the sub-graph for each sample, in the same order as in
the input.

This operator assumes that processing each sample is independent and could executed in parallel
or in any order. Users cannot expect any specific ordering in which each subgraph is computed.)DOC";

void SequenceMapInferenceFunction(InferenceContext& ctx) {
  auto num_inputs = ctx.getNumInputs();
  assert(num_inputs > 0);

  auto num_outputs = ctx.getNumOutputs();
  assert(num_outputs > 0);

  std::vector<TypeProto> tmp_type_protos(num_inputs);
  std::vector<const TypeProto*> subgraph_input_types;
  subgraph_input_types.reserve(num_inputs);
  for (size_t inputIndex = 0; inputIndex < num_inputs; inputIndex++) {
    auto input_type = ctx.getInputType(inputIndex);
    if (input_type == nullptr) {
      fail_type_inference("Input ", inputIndex, " expected to have type info");
    }
    if (input_type->value_case() == TypeProto::kSequenceType) {
      tmp_type_protos[inputIndex].CopyFrom(input_type->sequence_type().elem_type());
      subgraph_input_types.push_back(&tmp_type_protos[inputIndex]);
    } else {
      if (inputIndex == 0)
        fail_type_inference("Input ", inputIndex, " expected to be a sequence type");
      subgraph_input_types.push_back(input_type);
    }
  }

  GraphInferencer* graphInferencer = ctx.getGraphAttributeInferencer("body");
  if (!graphInferencer)
    fail_type_inference("Graph attribute inferencer for \"body\" not available");

  std::vector<const TensorProto*> input_data(num_inputs, nullptr);
  std::vector<const TypeProto*> subgraph_output_types =
      graphInferencer->doInferencing(subgraph_input_types, input_data);

  // if empty(), assume inferencing was skipped
  if (!subgraph_output_types.empty()) {
    if (subgraph_output_types.size() != num_outputs) {
      fail_type_inference(
          "Graph attribute inferencing returned type information for ",
          subgraph_output_types.size(),
          " outputs. Expected ",
          num_outputs);
    }

    for (size_t outputIndex = 0; outputIndex < num_outputs; outputIndex++) {
      auto* subgraph_output_type = subgraph_output_types[outputIndex];
      ctx.getOutputType(outputIndex)->mutable_sequence_type()->mutable_elem_type()->CopyFrom(*subgraph_output_type);
    }
  }
}

bool BuildSequenceMapBodyFunc(
    const FunctionBodyBuildContext& ctx,
    const OpSchema& schema,
    FunctionProto& functionProto) {
  schema.BuildFunction(functionProto);

  // variadic input/outputs will be expanded
  functionProto.clear_input();
  functionProto.clear_output();

  auto body_attr = ctx.getAttribute("body");
  if (!body_attr || !body_attr->has_g())
    ONNX_THROW_EX(std::invalid_argument("Invalid ``body`` argument. Expected a graph"));
  const GraphProto& body = body_attr->g();

  auto g_inputs = body.input();
  int ninputs = g_inputs.size();
  if (ninputs < 1)
    ONNX_THROW_EX(std::invalid_argument("Expected 1 or more inputs."));

  auto g_outputs = body.output();
  int noutputs = g_outputs.size();
  if (noutputs < 1)
    ONNX_THROW_EX(std::invalid_argument("Expected 1 or more outputs."));

  if (!ctx.hasInput(0))
    ONNX_THROW_EX(std::invalid_argument(MakeString("Input 0 expected but not provided")));

  const auto* first_input_type = ctx.getInputType(0);
  assert(first_input_type);
  if (!first_input_type->has_sequence_type())
    ONNX_THROW_EX(std::invalid_argument("Expected a sequence type for input 0"));

  auto schema_inputs = schema.inputs();
  auto input_0_name = schema_inputs[0].GetName();
  auto input_1_name = schema_inputs[1].GetName(); // variadic input

  *functionProto.add_input() = input_0_name;
  for (int i = 1; i < ninputs; i++) {
    if (!ctx.hasInput(i))
      ONNX_THROW_EX(std::invalid_argument(MakeString("Input ", i, " expected but not provided")));
    *functionProto.add_input() = MakeString(input_1_name, "_", i);
  }

  auto schema_outputs = schema.outputs();
  auto output_0_name = schema_outputs[0].GetName();
  for (int i = 0; i < noutputs; i++) {
    if (!ctx.hasOutput(i))
      ONNX_THROW_EX(std::invalid_argument(MakeString("Output ", i, " expected but not provided")));
    *functionProto.add_output() = MakeString(output_0_name, "_", i);
  }

  // Loop body subgraph
  std::string loopbody_graph_name("SequenceMap_loop_body");
  GraphProto loopbody_graph;
  loopbody_graph.set_name(loopbody_graph_name);
  {
    TypeProto int64_type;
    int64_type.mutable_tensor_type()->set_elem_type(TensorProto_DataType_INT64);
    int64_type.mutable_tensor_type()->mutable_shape()->Clear();

    TypeProto bool_type;
    bool_type.mutable_tensor_type()->set_elem_type(TensorProto_DataType_BOOL);
    bool_type.mutable_tensor_type()->mutable_shape()->Clear();

    ValueInfoProto iter_count;
    std::string iter_count_name = MakeString(loopbody_graph_name, "_itercount");
    iter_count.set_name(iter_count_name);
    *iter_count.mutable_type() = int64_type;
    *loopbody_graph.add_input() = iter_count;

    ValueInfoProto cond_in;
    std::string cond_in_name = MakeString(loopbody_graph_name, "_cond_in");
    cond_in.set_name(cond_in_name);
    *cond_in.mutable_type() = bool_type;
    *loopbody_graph.add_input() = cond_in;

    ValueInfoProto cond_out;
    std::string cond_out_name = MakeString(loopbody_graph_name, "_cond_out");
    cond_out.set_name(cond_out_name);
    *cond_out.mutable_type() = bool_type;
    *loopbody_graph.add_output() = cond_out;

    NodeProto cond_identity;
    cond_identity.set_domain(ONNX_DOMAIN);
    cond_identity.set_op_type("Identity");
    cond_identity.add_input(cond_in_name);
    cond_identity.add_output(cond_out_name);
    *loopbody_graph.add_node() = cond_identity;

    for (int inputIndex = 0; inputIndex < ninputs; inputIndex++) {
      const auto* input_type = ctx.getInputType(inputIndex);
      if (input_type && input_type->has_sequence_type()) {
        // If it's a sequence input, extract ``iter_count`` element
        NodeProto seq_at_node;
        seq_at_node.set_domain(ONNX_DOMAIN);
        seq_at_node.set_op_type("SequenceAt");
        seq_at_node.add_input(functionProto.input(inputIndex));
        seq_at_node.add_input(iter_count_name);
        seq_at_node.add_output(g_inputs.Get(inputIndex).name());
        *loopbody_graph.add_node() = seq_at_node;
      } else {
        // If not a sequence, simply connect
        NodeProto identity;
        identity.set_domain(ONNX_DOMAIN);
        identity.set_op_type("Identity");
        identity.add_input(functionProto.input(inputIndex));
        identity.add_output(g_inputs.Get(inputIndex).name());
        *loopbody_graph.add_node() = identity;
      }
    }

    for (const auto& item : body.node())
      *loopbody_graph.add_node() = item;
    for (const auto& item : body.value_info())
      *loopbody_graph.add_value_info() = item;
    for (const auto& item : body.initializer())
      *loopbody_graph.add_initializer() = item;
    for (const auto& item : body.sparse_initializer())
      *loopbody_graph.add_sparse_initializer() = item;

    for (int outputIndex = 0; outputIndex < noutputs; outputIndex++) {
      const auto& body_out_i = body.output(outputIndex);
      assert(body_out_i.type().has_tensor_type());
      std::string prefix = MakeString(loopbody_graph_name, "_", body_out_i.name());
      std::string loopbody_in_name = MakeString(prefix, "_in");

      ValueInfoProto tmp;
      *tmp.mutable_type()->mutable_sequence_type()->mutable_elem_type()->mutable_tensor_type() =
          body_out_i.type().tensor_type();
      tmp.set_name(loopbody_in_name);
      *loopbody_graph.add_input() = tmp;

      std::string loopbody_out_name = MakeString(prefix, "_out");
      tmp.set_name(loopbody_out_name);
      *loopbody_graph.add_output() = tmp;

      NodeProto seq_insert_node;
      seq_insert_node.set_domain(ONNX_DOMAIN);
      seq_insert_node.set_op_type("SequenceInsert");
      seq_insert_node.add_input(loopbody_in_name);
      seq_insert_node.add_input(body_out_i.name());
      seq_insert_node.add_output(loopbody_out_name);
      *loopbody_graph.add_node() = seq_insert_node;
    }
  }

  std::vector<FunctionBodyHelper::NodeDef> nodes;

  // TODO: figure out a way to prevent name collisions?
  auto first_input_name = functionProto.input(0);
  std::string prefix = MakeString("SequenceMap_", first_input_name);
  std::string seqlen = MakeString(prefix, "_seqlen");
  nodes.push_back({{seqlen}, "SequenceLength", {first_input_name}});

  std::string cond_bool = MakeString(prefix, "_cond");
  nodes.push_back(FunctionBodyHelper::Const<bool>(cond_bool, true));

  std::vector<std::string> loop_node_inputs = {seqlen, cond_bool};
  std::vector<std::string> loop_node_outputs;
  for (int outputIndex = 0; outputIndex < noutputs; outputIndex++) {
    auto output_name = functionProto.output(outputIndex);
    std::string out_prefix = MakeString("SequenceMap_", output_name);

    std::string seqempty_name = MakeString(out_prefix, "_seqempty");
    int64_t dtype = g_outputs.Get(outputIndex).type().tensor_type().elem_type();
    nodes.push_back({{seqempty_name}, "SequenceEmpty", {}, {MakeAttribute("dtype", dtype)}});
    loop_node_inputs.push_back(seqempty_name);
    loop_node_outputs.push_back(output_name);
  }

  nodes.push_back({loop_node_outputs, "Loop", loop_node_inputs, {MakeAttribute("body", loopbody_graph)}});

  auto func_nodes = FunctionBodyHelper::BuildNodes(nodes);
  for (const auto& node : func_nodes) {
    auto new_node = functionProto.add_node();
    new_node->CopyFrom(node);
  }

  return true;
}

ONNX_OPERATOR_SET_SCHEMA(
    SequenceMap,
    17,
    OpSchema()
        .SetDoc(SequenceMap_ver17_doc)
        .Attr(
            "body",
            "The graph to be run for each sample in the sequence(s). "
            "It should have as many inputs and outputs as inputs and "
            "outputs to the SequenceMap function.",
            AttributeProto::GRAPH)
        .Input(0, "input_sequence", "Input sequence.", "S")
        .Input(1, "additional_inputs", "Additional inputs to the graph", "V", OpSchema::Variadic, false, 0)
        .Output(0, "out_sequence", "Output sequence(s)", "S", OpSchema::Variadic, false)
        .TypeConstraint("S", OpSchema::all_tensor_sequence_types(), "Constrain input types to any sequence type.")
        .TypeConstraint(
            "V",
            []() {
              auto t = OpSchema::all_tensor_types();
              auto s = OpSchema::all_tensor_sequence_types();
              t.insert(t.end(), s.begin(), s.end());
              return t;
            }(),
            "Constrain to any tensor or sequence type.")
        .SetContextDependentFunctionBodyBuilder(BuildSequenceMapBodyFunc)
        .TypeAndShapeInferenceFunction(SequenceMapInferenceFunction));

} // namespace ONNX_NAMESPACE