schema.fbs

// Flatbuffers schema for serialized RTen models.
//
// See https://google.github.io/flatbuffers/flatbuffers_guide_writing_schema.html.
//
// Operator names and attributes align with the corresponding ONNX operators.
// See https://onnx.ai/onnx/operators/index.html.

file_identifier "RTEN";
file_extension "rten";

// Type for an operator node
enum OperatorType: ubyte {
  Add,
  ArgMin,
  ArgMax,
  AveragePool,
  BatchNormalization,
  Cast,
  Clip,
  Concat,
  ConstantOfShape,
  Conv,
  ConvTranspose,
  Cos,
  CumSum,
  Div,
  Equal,
  Erf,
  Expand,
  Flatten,
  Gather,
  Gemm,
  GlobalAveragePool,
  Greater,
  GRU,
  Identity,
  LeakyRelu,
  Less,
  LessOrEqual,
  Log,
  LogSoftmax,
  LSTM,
  MatMul,
  MaxPool,
  Mod,
  Mul,
  Pad,
  Pow,
  Range,
  ReduceMean,
  ReduceL2,
  Relu,
  Reshape,
  Resize,
  Shape,
  Sigmoid,
  Sin,
  Slice,
  Split,
  Sqrt,
  Squeeze,
  Softmax,
  Sub,
  Tanh,
  Transpose,
  Unsqueeze,
  Where,

  // New operators. These are appended here to preserve binary compatibility
  // of existing model files.
  ReduceProd,
  ReduceSum,
  ReduceMin,
  ReduceMax,
  NonZero,
  ScatterElements,
  Tile,
  Not,
  Abs,
  Max,
  Mean,
  Min,
  Sum,
  OneHot,
  Round,
  Floor,
  Ceil,
  Reciprocal,
  TopK,
  Neg,
  Exp,
  GreaterOrEqual,
  Size,
  Tan,
  Acos,
  Asin,
  Atan,
  InstanceNormalization,
  HardSigmoid,
  HardSwish,
  And,
  Or,
  Xor,
  Trilu,
  ScatterND,
  NonMaxSuppression,
  Sign,
  GatherElements,
  LayerNormalization,
  ReduceSumSquare,
  RandomUniform,
  Elu,
  RandomUniformLike,
  RandomNormal,
  RandomNormalLike,
  Softplus,
  GatherND,
  Gelu,
  Einsum,
  If,
  DequantizeLinear,
  QuantizeLinear,
  DynamicQuantizeLinear,
  MatMulInteger,
}

enum RNNDirection: ubyte {
  Forward,
  Reverse,
  Bidirectional
}

// Values for `auto_pad` field. In ONNX the default for `auto_pad` is "NOTSET".
// In this schema the default is `Same` for some operators for backwards
// compatibility. The converter will however set it to `Fixed` if the `auto_pad`
// attribute is missing in the ONNX model, preserving ONNX semantics.
enum AutoPad: ubyte {
  // nb. ONNX defines `SAME_UPPER` and `SAME_LOWER`. This corresponds to
  // `SAME_UPPER` and TensorFlow / Keras's "same".
  Same,
  NotSet
}

enum DataType: ubyte {
  Int32,
  Float,
  Int8,
  UInt8,
}

// Coordinate transform modes for Resize operator.
enum CoordTransformMode: ubyte {
  HalfPixel,
  Asymmetric,
  AlignCorners
}

// Rounding modes supported by Resize operator when `ResizeMode` is `Nearest`.
enum NearestMode: ubyte {
  Floor,
  Ceil,
  RoundPreferFloor,
  RoundPreferCeil,
}

enum ResizeMode: ubyte {
  Nearest,
  Linear
}

// Operator-specific configuration
union OperatorAttrs {
  ArgMaxAttrs, // Also used for ArgMin
  AveragePoolAttrs,
  BatchNormalizationAttrs, // Also used for InstanceNormalization
  CastAttrs,
  ConcatAttrs,
  ConstantOfShapeAttrs,
  ConvAttrs,
  ConvTransposeAttrs,
  FlattenAttrs,
  GatherAttrs, // Also used for GatherElements
  GemmAttrs,
  GRUAttrs,
  LeakyReluAttrs,
  LSTMAttrs,
  MaxPoolAttrs,
  ReduceMeanAttrs, // Also used for Reduce* ops
  ReshapeAttrs,
  ResizeAttrs,
  SplitAttrs,
  SoftmaxAttrs, // Also used for LogSoftmax
  TransposeAttrs,

  // New operator attrs. These are appended here to preserve binary compatibility
  // of existing model files.
  ModAttrs,
  ScatterElementsAttrs,
  OneHotAttrs,
  TopKAttrs,
  HardSigmoidAttrs,
  TriluAttrs,
  ScatterNDAttrs,
  NonMaxSuppressionAttrs,
  LayerNormalizationAttrs,
  RandomUniformAttrs,
  EluAttrs,
  RandomUniformLikeAttrs,
  RandomNormalAttrs,
  RandomNormalLikeAttrs,
  GatherNDAttrs,
  GeluAttrs,
  EinsumAttrs,
  IfAttrs,
  PadAttrs,
  DequantizeLinearAttrs,
  QuantizeLinearAttrs,
}

table ArgMaxAttrs {
  axis:int;
  keep_dims:bool;
}

table AveragePoolAttrs {
  kernel_size:[uint] (required);
  auto_pad:AutoPad;

  // Padding for spatial axes as [top, left, bottom, right]
  pads:[uint];

  strides:[uint];

  count_include_pad:bool;
}

table BatchNormalizationAttrs {
  epsilon:float;
}

table CastAttrs {
  to:DataType;
}

table ConcatAttrs {
  axis:int;
}

union Scalar {
  IntScalar,
  FloatScalar
}

table IntScalar {
  value:int32;
}

table FloatScalar {
  value:float32;
}

table ConstantOfShapeAttrs {
  value:Scalar (required);
}

table ConvAttrs {
  auto_pad:AutoPad;

  // Padding for spatial axes as [left, right] or [top, left, bottom, right]
  pads:[uint];

  groups:uint;
  strides:[uint];
  dilations:[uint];
}

table ConvTransposeAttrs {
  strides:[uint];

  // Defaults to `NotSet` for backwards compatibility.
  auto_pad:AutoPad = NotSet;

  // Padding for spatial axes as [left, right] or [top, left, bottom, right]
  pads:[uint];
}

table DequantizeLinearAttrs {
  axis:int;
}

table EinsumAttrs {
  equation:string;
}

table EluAttrs {
  alpha:float;
}

table FlattenAttrs {
  axis:int;
}

table LayerNormalizationAttrs {
  axis:int;
  epsilon:float;
}

table GatherAttrs {
  axis:int;
}

table GatherNDAttrs {
  batch_dims:int;
}

// Reserved for supporting `approximate` attr in future.
table GeluAttrs {}

table GemmAttrs {
  alpha:float;
  beta:float;
  transpose_a:bool;
  transpose_b:bool;
}

table GRUAttrs {
  direction:RNNDirection;
  hidden_size:uint;
  linear_before_reset:bool;
}

table HardSigmoidAttrs {
  alpha:float;
  beta:float;
}

table IfAttrs {
  then_branch:Graph;
  else_branch:Graph;
}

table LeakyReluAttrs {
  alpha:float;
}

table LSTMAttrs {
  direction:RNNDirection;
  hidden_size:uint;
}

table MaxPoolAttrs {
  kernel_size:[uint] (required);
  auto_pad:AutoPad;

  // Padding for spatial axes as [top, left, bottom, right]
  pads:[uint];

  strides:[uint];
}

table ModAttrs {
  fmod:bool;
}

enum NMSBoxOrder: ubyte {
  TopLeftBottomRight,
  CenterWidthHeight,
}

table NonMaxSuppressionAttrs {
  box_order:NMSBoxOrder;
}

table OneHotAttrs {
  axis:int;
}

enum PadMode: ubyte {
  Constant,
  Reflect,
}

table PadAttrs {
  mode:PadMode;
}

table QuantizeLinearAttrs {
  axis:int;
  output_dtype:DataType = null;
}

table RandomNormalAttrs {
  mean:float;
  scale:float;
  seed:float = null;
  shape:[uint];
}

table RandomNormalLikeAttrs {
  mean:float;
  scale:float;
  seed:float = null;
}

table RandomUniformAttrs {
  shape:[uint];
  high:float;
  low:float;
  seed:float = null;
}

table RandomUniformLikeAttrs {
  high:float;
  low:float;
  seed:float = null;
}

table ReduceMeanAttrs {
  axes:[int];
  keep_dims:bool;
}

table ReshapeAttrs {
  allow_zero:bool;
}

table ResizeAttrs {
  mode:ResizeMode;
  coord_mode:CoordTransformMode;
  nearest_mode:NearestMode;
}

enum ScatterReduction: ubyte {
  None,
  Add,
  Mul,
  Min,
  Max,
}

table ScatterElementsAttrs {
  axis:int;
  reduction:ScatterReduction;
}

table ScatterNDAttrs {
  reduction:ScatterReduction;
}

table SoftmaxAttrs {
  axis:int;
}

table SplitAttrs {
  axis:int;
}

table TopKAttrs {
  axis:int;
  largest:bool;
  sorted:bool;
}

table TransposeAttrs {
  perm:[uint];
}

table TriluAttrs {
  upper:bool;
}

// Node in the dataflow graph
union NodeKind {
  OperatorNode,
  ConstantNode,
  ValueNode
}

// Graph node that computes an output tensor given one or more inputs and
// operator configuration.
table OperatorNode {
  type:OperatorType;
  attrs:OperatorAttrs;

  // Indexes of input nodes. Negative values indicate missing optional inputs.
  inputs:[int];

  // Indexes of output nodes. Negative values indicate unused outputs.
  outputs:[int];
}

// Data for constants stored inline in a model.
union ConstantData {
  FloatData,
  Int32Data,
  Int8Data,
  UInt8Data,
}

table FloatData {
  data: [float32] (required);
}

table Int8Data {
  data: [byte] (required);
}

table Int32Data {
  data: [int32] (required);
}

table UInt8Data {
  data: [ubyte] (required);
}

enum ConstantDataType: ushort {
  Int32, // Signed 32-bit int
  Float32, // IEEE-754 32-bit float
  Int8,
  UInt8,
}

// Graph node for a constant tensor value, whose data is part of the model.
table ConstantNode {
  shape:[uint] (required);

  // Tensor data embedded within the model file.
  data:ConstantData;

  // Data type. This is null in older models.
  dtype:ConstantDataType = null;

  // Offset of tensor data from the start of the tensor data segment in the
  // model file. Null if the tensor data is stored inline.
  data_offset:uint64 = null;
}

// Dimension of a ValueNode's shape. This can be either a fixed value or a
// symbolic name.
//
// This is logically a union, but uses a table due to
// https://github.com/google/flatbuffers/issues/5024.
table Dim {
  value:uint;
  name:string;
}

// Graph node for a dynamic tensor value, such as a model input or operator
// output.
table ValueNode {
  // Expected shape of the tensor at runtime.
  shape:[Dim];
  // Expected data type of the tensor at runtime.
  dtype:DataType = null;
}

table Node {
  // Identifier for external referencing and debugging
  name:string;
  data:NodeKind;
}

// Graph describing a sequence of operations in a model.
//
// This is analagous to the body of a function or closure in a programming
// language.
table Graph {
  // Nodes are sorted in topological order. This means that if a node references
  // other nodes (eg. an operator node referencing inputs), the referents must
  // come earlier in the list.
  nodes:[Node];

  // IDs of input nodes
  inputs:[uint];

  // IDs of output nodes
  outputs:[uint];

  // IDs of nodes which capture their values from the enclosing scope when
  // this graph is run as a subgraph.
  captures:[uint];
}

table Metadata {
  // SHA-256 hash of the ONNX model that was used as the source for this RTen
  // model.
  onnx_hash:string;

  // A short description of what this model does.
  description:string;

  // Identifier for the license used in this model.
  //
  // This should be an SPDX (https://spdx.org/licenses/) identifier for openly
  // licensed models.
  license:string;

  // Commit ID for the code that produced this model.
  commit:string;

  // URL of repository where the model's code is hosted (eg. GitHub).
  code_repository:string;

  // URL of repository where the model is hosted (eg. Hugging Face).
  model_repository:string;

  // Identifier for the training run that produced this model.
  run_id:string;

  // URL of logs etc. for the training run that produced this model.
  run_url:string;
}

table Model {
  schema_version:int;
  graph:Graph (required);
  metadata:Metadata;
}

root_type Model;