domain

Documentation

Index

• Constants
• func Transform(ctx context.Context, signature *Signature, input *gtly.Object, ...) (common.Storable, error)
• type Feed
• type Input
• type Output
• type Signature
• type Transformer

Constants

const (
	//DefaultSignatureKey signature key
	DefaultSignatureKey = "serving_default"
)

Functions

func Transform

func Transform(ctx context.Context, signature *Signature, input *gtly.Object, output interface{}) (common.Storable, error)

The default Transformer will create an object that when marshalled will return as close to the the TensorFlow SavedModel signature as possible; with the caveat that any output will be keyed by the output tensor's name.

Types

type Feed

type Feed struct {
	Name  string
	Value interface{}
}

Feed represents data feed

type Input

type Input struct {
	Name  string
	Index int // Position of Tensor in model input.

	Placeholder tf.Output // TODO refactor out this usage in service/domain.Signature is different from its usage in service/request.Request

	Vocab     bool // false if embedded vocabulary should be ignored
	Auxiliary bool // true if this input isn't part of the model

	Type reflect.Type
}

type Output

type Output struct {
	Name         string
	DataType     string
	DataTypeKind reflect.Kind
	Index        int
	*tf.Operation
}

Output represents model output

type Signature

type Signature struct {
	Method string

	Inputs  []Input
	Outputs []Output

	Output Output // Deprecated: Use Outputs[0] if there is only 1 output
}

Signature represents Tensorflow SavedModel function Signature. Contains information required to extract vocabularies, unmarshal requests, and validate request inputs. TODO document and address issues if reloaded model IO changes.

type Transformer

type Transformer func(ctx context.Context, signature *Signature, input *gtly.Object, output interface{}) (common.Storable, error)

Transformer is an adapter module used when the output of the TensorFlow model wants to be modified server side. signature is the Signature of the relevant model, determined by request context. input is the request body object as a *gtly.Object output is the TensorFlow SavedModel prediction output.