menoh

README

Menoh Go

Golang binding for Menoh

Requirements

• Go 1.10+
• Menoh 1.0.3+

Install

After install Menoh, then

$ go get -u github.com/pfnet-research/go-menoh

Linux/Mac

Add a path to library to LD_LIBRARY_PATH environment variable. Menoh libraries are installed to /usr/local/lib on default.

$ export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH

Windows

Add a path to DLLs distributed by Menoh to local Path environment.

\path\to\menoh\bin
  |- libiomp5md.dll
  |- menoh.dll
  |- mkldnn.dll
  |- mklml.dll

set PATH=\path\to\menoh\bin;%PATH%

Usage

• example/vgg16 is a tutorial for this package.
• example/mnist is an example using MNIST dataset and model.

Development

Test

Download ONNX file, using in menoh-rust test, before testing.

$ wget https://github.com/pfnet-research/menoh-rs/releases/download/assets/MLP.onnx -P test_data
$ go test ./...

Additionally go-menoh follows gofmt with simplify option (-s), go vet and golint.

Note

• At first name of this repository is "menoh-go", and renamed to "go-menoh" to follow recommended naming rule

License

MIT License (see LICENSE file).

Documentation

Overview

Package menoh provides a wrapper "runner" for Menoh library to execute ONNX model. It consits the following structures.

Runner is a wrapper to manage an ONNX model and run with input variables.

Config is information about ONNX model, required for making a runner.

Tensor represents number array for in/out variable, similar to ONNX's Tensor.

Index

• type Config
• type FloatTensor
  • func (t *FloatTensor) FloatArray() ([]float32, error)
  • func (t *FloatTensor) Shape() []int32
  • func (t *FloatTensor) Size() int
  • func (t *FloatTensor) WriteFloat(i int, f float32) error
• type InputConfig
• type ModelData
  • func NewModelDataFromBytes(data []byte) (*ModelData, error)
  • func NewModelDataFromPath(path string) (*ModelData, error)
  • func NewRawModelData() (*ModelData, error)
  • func (m *ModelData) AddTensorParameter(name string, param Tensor) error
• type OutputConfig
• type Runner
  • func NewRunner(conf Config) (*Runner, error)
  • func NewRunnerWithModelData(modelData *ModelData, conf Config) (*Runner, error)
  • func (r *Runner) GetInput(name string) (Tensor, error)
  • func (r *Runner) GetOutput(name string) (Tensor, error)
  • func (r *Runner) Outputs() map[string]Tensor
  • func (r *Runner) Run(inputs map[string]Tensor) error
  • func (r *Runner) RunWithTensor(name string, t Tensor) error
  • func (r *Runner) Stop()
• type Tensor
• type TypeBackend
  • func (t TypeBackend) String() string
• type TypeDtype

Types

type Config

type Config struct {
	ONNXModelPath string         // path of ONNX file
	Backend       TypeBackend    // backend type, like menoh.TypeMKLDNN
	BackendConfig string         // backend configuration
	Inputs        []InputConfig  // list of input configuration
	Outputs       []OutputConfig // list of output configuration
}

Config is setup information to build Menoh model.

type FloatTensor

type FloatTensor struct {
	Dims  []int32
	Array []float32
}

FloatTensor represents float32 Tessor.

func (*FloatTensor) FloatArray

func (t *FloatTensor) FloatArray() ([]float32, error)

FloatArray returns float32 array.

func (*FloatTensor) Shape

func (t *FloatTensor) Shape() []int32

Shape returns shape of array.

func (*FloatTensor) Size

func (t *FloatTensor) Size() int

Size returns array size.

func (*FloatTensor) WriteFloat

func (t *FloatTensor) WriteFloat(i int, f float32) error

WriteFloat puts float value to i-th index of array.

type InputConfig

type InputConfig struct {
	Name  string    // layer name
	Dtype TypeDtype // data type
	Dims  []int32   // list of dimension size
}

InputConfig is input variable information to pass to the model.

type ModelData

type ModelData struct {
	external.ModelData
}

ModelData is a wrap of menoh.ModelData.

func NewModelDataFromBytes

func NewModelDataFromBytes(data []byte) (*ModelData, error)

NewModelDataFromBytes return ModelData from binary data.

func NewModelDataFromPath

func NewModelDataFromPath(path string) (*ModelData, error)

NewModelDataFromPath returns ModelData using ONNX file placed on the path.

func NewRawModelData

func NewRawModelData() (*ModelData, error)

NewRawModelData return empty ModelData to be setup outer.

func (*ModelData) AddTensorParameter

func (m *ModelData) AddTensorParameter(name string, param Tensor) error

AddTensorParameter adds tensor to named parameter.

type OutputConfig

type OutputConfig struct {
	Name         string    // layer name
	Dtype        TypeDtype // data type
	FromInternal bool      // if the output comes from operator variable such as weight, set true
}

OutputConfig is output variable information to get from the model.

type Runner

type Runner struct {
	// contains filtered or unexported fields
}

Runner setups Menoh model with profiling and executes with input variables. A runner supports to call Run (or RunWithTensor) method repeatedly until stopping.

func NewRunner

func NewRunner(conf Config) (*Runner, error)

NewRunner returns Runner using configuration, the runner setup Menoh model and ready for execution. Require to call Stop function after the process is done.

func NewRunnerWithModelData

func NewRunnerWithModelData(modelData *ModelData, conf Config) (*Runner, error)

NewRunnerWithModelData returns Runner using configuration and ONNX model. The ONNX model is passed on memory, not use conf.ONNXModelPath. Spec of a returned runner is same as NewRunner, see docs of the function.

func (*Runner) GetInput

func (r *Runner) GetInput(name string) (Tensor, error)

GetInput returns a Tensor attached to the target model.

func (*Runner) GetOutput

func (r *Runner) GetOutput(name string) (Tensor, error)

GetOutput returns the target variable.

func (*Runner) Outputs

func (r *Runner) Outputs() map[string]Tensor

Outputs all variables set by the configuration.

func (*Runner) Run

func (r *Runner) Run(inputs map[string]Tensor) error

Run with the inputs which are set name and tensor as key-value. If nothing to input, set nil.

func (*Runner) RunWithTensor

func (r *Runner) RunWithTensor(name string, t Tensor) error

RunWithTensor inputs the tenser with tne name, and runs.

func (*Runner) Stop

func (r *Runner) Stop()

Stop the runner.

type Tensor

type Tensor interface {

	// Size returns array size.
	Size() int

	// Shape returns shape of array dimension.
	Shape() []int32

	// FloatArray returns float32 array. Returns an error when the array
	// cannot cast to the type. It is possible that returned array is copied
	// or split-off from attached array in Menoh model. When updating values,
	// WriteFloat method.
	FloatArray() ([]float32, error)

	// WriteFloat puts float value to i-th index of array.
	WriteFloat(int, float32) error
	// contains filtered or unexported methods
}

Tensor is base unit of matrix data to pass Menoh model.

type TypeBackend

type TypeBackend int

TypeBackend is a type of backend, like MKL-DNN.

const (

	// TypeMKLDNN is a TypeBackend of MKL-DNN
	TypeMKLDNN TypeBackend
)

func (TypeBackend) String

func (t TypeBackend) String() string

type TypeDtype

type TypeDtype int

TypeDtype is a type of data.

const (

	// TypeFloat is a TypeDtype of Float.
	TypeFloat TypeDtype
)