Documentation
¶
Overview ¶
Package onnx provides ONNX model import/export functionality.
ONNX (Open Neural Network Exchange) is an open format for representing deep learning models. This package implements a hand-written protobuf parser for .onnx files without external dependencies.
Key components:
- ModelProto: Top-level ONNX model structure with metadata and graph
- GraphProto: Computation graph with nodes, inputs, outputs, and initializers
- NodeProto: Single operation in the graph (e.g., Conv, MatMul, Relu)
- TensorProto: Weight/initializer tensor with data and shape
- ValueInfoProto: Input/output tensor type information
Supported data types:
- float32, float64 (primary ML types)
- int8, int16, int32, int64 (integer types)
- uint8, uint16, uint32, uint64 (unsigned types)
- bool (boolean type)
Example usage:
// Parse ONNX file
model, err := onnx.ParseFile("resnet50.onnx")
if err != nil {
log.Fatal(err)
}
// Inspect model
fmt.Printf("Model: %s (version %d)\n", model.ProducerName, model.ModelVersion)
fmt.Printf("Graph: %s with %d nodes\n", model.Graph.Name, len(model.Graph.Nodes))
// Iterate over operations
for _, node := range model.Graph.Nodes {
fmt.Printf("Op: %s (type: %s)\n", node.Name, node.OpType)
}
Index ¶
- Constants
- func ListSupportedOps() []string
- type AttributeProto
- type DimensionProto
- type GraphProto
- type LoadOptions
- type Model
- func (m *Model) Forward(input *tensor.RawTensor) (*tensor.RawTensor, error)
- func (m *Model) ForwardNamed(inputs map[string]*tensor.RawTensor) (map[string]*tensor.RawTensor, error)
- func (m *Model) InputNames() []string
- func (m *Model) Metadata() map[string]string
- func (m *Model) OpsetVersion() int64
- func (m *Model) OutputNames() []string
- type ModelInfo
- type ModelProto
- type NodeProto
- type OperatorSetID
- type StringStringEntry
- type TensorProto
- type TensorShapeProto
- type TensorTypeProto
- type TypeProto
- type ValueInfoProto
Constants ¶
const ( TensorProtoUndefined = 0 TensorProtoFloat = 1 // float32 TensorProtoUint8 = 2 // uint8 TensorProtoInt8 = 3 // int8 TensorProtoUint16 = 4 // uint16 TensorProtoInt16 = 5 // int16 TensorProtoInt32 = 6 // int32 TensorProtoInt64 = 7 // int64 TensorProtoString = 8 // string TensorProtoBool = 9 // bool TensorProtoFloat16 = 10 // float16 TensorProtoDouble = 11 // float64 TensorProtoUint32 = 12 // uint32 TensorProtoUint64 = 13 // uint64 TensorProtoComplex64 = 14 // complex64 TensorProtoComplex128 = 15 // complex128 TensorProtoBfloat16 = 16 // bfloat16 )
ONNX data types (TensorProto.DataType).
const ( AttributeProtoUndefined = 0 AttributeProtoFloat = 1 // FLOAT AttributeProtoInt = 2 // INT AttributeProtoString = 3 // STRING AttributeProtoTensor = 4 // TENSOR AttributeProtoGraph = 5 // GRAPH AttributeProtoFloats = 6 // FLOATS AttributeProtoInts = 7 // INTS AttributeProtoStrings = 8 // STRINGS AttributeProtoTensors = 9 // TENSORS AttributeProtoGraphs = 10 // GRAPHS )
ONNX attribute types (AttributeProto.Type).
Variables ¶
This section is empty.
Functions ¶
func ListSupportedOps ¶
func ListSupportedOps() []string
ListSupportedOps returns all supported ONNX operators.
Types ¶
type AttributeProto ¶
type AttributeProto struct {
Name string // Attribute name
Type int32 // Attribute type
F float32 // FLOAT value
I int64 // INT value
S []byte // STRING value
T *TensorProto // TENSOR value
G *GraphProto // GRAPH value
Floats []float32 // FLOATS array
Ints []int64 // INTS array
Strings [][]byte // STRINGS array
Tensors []TensorProto // TENSORS array
Graphs []GraphProto // GRAPHS array
DocString string // Description
}
AttributeProto represents node attributes.
type DimensionProto ¶
type DimensionProto struct {
DimValue int64 // Static dimension value (e.g., 224 for image size)
DimParam string // Dynamic dimension name (e.g., "batch_size")
}
DimensionProto describes a single dimension.
type GraphProto ¶
type GraphProto struct {
Name string // Graph name
Nodes []NodeProto // Operation nodes
Inputs []ValueInfoProto // Graph inputs
Outputs []ValueInfoProto // Graph outputs
Initializers []TensorProto // Weight tensors
DocString string // Graph description
ValueInfo []ValueInfoProto // Intermediate tensor info
}
GraphProto represents the computation graph.
type LoadOptions ¶
type LoadOptions struct {
// StrictMode fails on unsupported operators (default: false = skip with warning).
StrictMode bool
// CustomOps provides custom operator handlers.
CustomOps map[string]operators.OpHandler
}
LoadOptions configures model loading behavior.
func DefaultLoadOptions ¶
func DefaultLoadOptions() LoadOptions
DefaultLoadOptions returns default loading options.
type Model ¶
type Model struct {
// contains filtered or unexported fields
}
Model represents a loaded ONNX model ready for inference. It executes the computation graph using the provided backend.
func Load ¶
Load loads an ONNX model from file and prepares it for inference. The backend is used for tensor operations during inference.
Example:
model, err := onnx.Load("resnet50.onnx", backend)
if err != nil {
log.Fatal(err)
}
output, err := model.Forward(input)
func LoadFromBytes ¶
LoadFromBytes loads an ONNX model from bytes.
func LoadFromProto ¶
func LoadFromProto(proto *ModelProto, backend tensor.Backend, opt LoadOptions) (*Model, error)
LoadFromProto loads a model from parsed ModelProto.
func (*Model) Forward ¶
Forward runs inference with a single input tensor. For models with multiple inputs, use ForwardNamed.
func (*Model) ForwardNamed ¶
func (m *Model) ForwardNamed(inputs map[string]*tensor.RawTensor) (map[string]*tensor.RawTensor, error)
ForwardNamed runs inference with named inputs. Returns a map of output name to tensor.
func (*Model) InputNames ¶
InputNames returns the names of model inputs.
func (*Model) OpsetVersion ¶
OpsetVersion returns the ONNX opset version.
func (*Model) OutputNames ¶
OutputNames returns the names of model outputs.
type ModelInfo ¶
type ModelInfo struct {
IRVersion int64
OpsetVersion int64
ProducerName string
ProducerVersion string
InputNames []string
OutputNames []string
NodeCount int
WeightCount int
}
ModelInfo contains basic information about an ONNX model without fully loading it.
func GetModelInfo ¶
GetModelInfo extracts basic info from an ONNX file.
type ModelProto ¶
type ModelProto struct {
IRVersion int64 // IR version (e.g., 7, 8, 9)
OpsetImport []OperatorSetID // Opset version(s)
ProducerName string // Framework name (e.g., "pytorch", "tf")
ProducerVersion string // Framework version
Domain string // Model domain
ModelVersion int64 // Model version number
DocString string // Model description
Graph *GraphProto // Computation graph
MetadataProps []StringStringEntry // Key-value metadata
}
ModelProto represents an ONNX model.
func ParseFile ¶
func ParseFile(path string) (*ModelProto, error)
ParseFile parses an ONNX model from file.
type NodeProto ¶
type NodeProto struct {
Name string // Node name (optional)
OpType string // Operation type (e.g., "Conv", "MatMul", "Relu")
Inputs []string // Input tensor names
Outputs []string // Output tensor names
Attributes []AttributeProto // Operation attributes
Domain string // Custom domain (empty for default)
DocString string // Node description
}
NodeProto represents a single operation.
type OperatorSetID ¶
type OperatorSetID struct {
Domain string // Operator domain (empty for default)
Version int64 // Opset version number
}
OperatorSetID identifies opset version.
type StringStringEntry ¶
StringStringEntry represents key-value metadata.
type TensorProto ¶
type TensorProto struct {
Name string // Tensor name
DataType int32 // Element data type
Dims []int64 // Tensor shape
RawData []byte // Raw binary data (most common)
FloatData []float32 // Float32 data (legacy)
Int32Data []int32 // Int32 data (legacy)
Int64Data []int64 // Int64 data (legacy)
DocString string // Tensor description
}
TensorProto represents a tensor (weights/initializers).
type TensorShapeProto ¶
type TensorShapeProto struct {
Dims []DimensionProto // Dimensions
}
TensorShapeProto describes tensor dimensions.
type TensorTypeProto ¶
type TensorTypeProto struct {
ElemType int32 // Element data type
Shape *TensorShapeProto // Tensor shape
}
TensorTypeProto describes tensor shape and element type.
type TypeProto ¶
type TypeProto struct {
TensorType *TensorTypeProto // Tensor type (most common)
}
TypeProto describes tensor type.
type ValueInfoProto ¶
type ValueInfoProto struct {
Name string // Tensor name
Type *TypeProto // Tensor type information
DocString string // Description
}
ValueInfoProto describes input/output tensor specifications.
Source Files
Directories