tensorflow

README

TensorFlow in Go

Construct and execute TensorFlow graphs in Go.

WARNING: The API defined in this package is not stable and can change without notice. The same goes for the package path: (github.com/tensorflow/tensorflow/tensorflow/go).

Quickstart

Refer to Installing TensorFlow for Go

Building the TensorFlow C library from source

If the "Quickstart" instructions above do not work (perhaps the release archives are not available for your operating system or architecture, or you're using a different version of CUDA/cuDNN), then the TensorFlow C library must be built from source.

Prerequisites

• bazel

• Environment to build TensorFlow from source code (Linux or macOS). If you don't need GPU support, then try the following:

  sudo apt-get install python swig python-numpy # Linux
  brew install swig                             # OS X with homebrew
  

• Protocol buffer compiler (protoc) 3.x

Build

1. Download the source code

   go get -d github.com/tensorflow/tensorflow/tensorflow/go
   

2. Build the TensorFlow C library:

   cd ${GOPATH}/src/github.com/tensorflow/tensorflow
   ./configure
   bazel build -c opt //tensorflow:libtensorflow.so
   

   This can take a while (tens of minutes, more if also building for GPU).

3. Make libtensorflow.so and libtensorflow_framework.so available to the linker. This can be done by either:

   a. Copying it to a system location, e.g.,

   sudo cp ${GOPATH}/src/github.com/tensorflow/tensorflow/bazel-bin/tensorflow/libtensorflow.so /usr/local/lib
   sudo cp ${GOPATH}/src/github.com/tensorflow/tensorflow/bazel-bin/tensorflow/libtensorflow_framework.so /usr/local/lib
   

   OR

   b. Setting environment variables:

   export LIBRARY_PATH=${GOPATH}/src/github.com/tensorflow/tensorflow/bazel-bin/tensorflow
   # Linux
   export LD_LIBRARY_PATH=${GOPATH}/src/github.com/tensorflow/tensorflow/bazel-bin/tensorflow
   # OS X
   export DYLD_LIBRARY_PATH=${GOPATH}/src/github.com/tensorflow/tensorflow/bazel-bin/tensorflow
   

4. Build and test:

   go generate github.com/tensorflow/tensorflow/tensorflow/go/op
   go test github.com/tensorflow/tensorflow/tensorflow/go
   

Generate wrapper functions for ops

Go functions corresponding to TensorFlow operations are generated in op/wrappers.go. To regenerate them:

Prerequisites:

• Protocol buffer compiler (protoc) 3.x
• The TensorFlow repository under GOPATH

go generate github.com/tensorflow/tensorflow/tensorflow/go/op

Support

Use stackoverflow and/or Github issues.

Contributions

Contributions are welcome. If making any signification changes, probably best to discuss on a Github issue before investing too much time. Github pull requests are used for contributions.

Documentation

Overview

Package tensorflow is a Go binding to TensorFlow.

The API is subject to change and may break at any time.

TensorFlow (www.tensorflow.org) is an open source software library for numerical computation using data flow graphs. This package provides functionality to build and execute such graphs and depends on TensorFlow being available. For installation instructions see https://github.com/tensorflow/tensorflow/blob/master/tensorflow/go/README.md

Example

package main

import (
	"archive/zip"
	"bufio"
	"flag"
	"fmt"
	"io"
	"io/ioutil"
	"log"
	"net/http"
	"os"
	"path/filepath"

	tf "github.com/tensorflow/tensorflow/tensorflow/go"
	"github.com/tensorflow/tensorflow/tensorflow/go/op"
)

func main() {
	// An example for using the TensorFlow Go API for image recognition
	// using a pre-trained inception model (http://arxiv.org/abs/1512.00567).
	//
	// Sample usage: <program> -dir=/tmp/modeldir -image=/path/to/some/jpeg
	//
	// The pre-trained model takes input in the form of a 4-dimensional
	// tensor with shape [ BATCH_SIZE, IMAGE_HEIGHT, IMAGE_WIDTH, 3 ],
	// where:
	// - BATCH_SIZE allows for inference of multiple images in one pass through the graph
	// - IMAGE_HEIGHT is the height of the images on which the model was trained
	// - IMAGE_WIDTH is the width of the images on which the model was trained
	// - 3 is the (R, G, B) values of the pixel colors represented as a float.
	//
	// And produces as output a vector with shape [ NUM_LABELS ].
	// output[i] is the probability that the input image was recognized as
	// having the i-th label.
	//
	// A separate file contains a list of string labels corresponding to the
	// integer indices of the output.
	//
	// This example:
	// - Loads the serialized representation of the pre-trained model into a Graph
	// - Creates a Session to execute operations on the Graph
	// - Converts an image file to a Tensor to provide as input to a Session run
	// - Executes the Session and prints out the label with the highest probability
	//
	// To convert an image file to a Tensor suitable for input to the Inception model,
	// this example:
	// - Constructs another TensorFlow graph to normalize the image into a
	//   form suitable for the model (for example, resizing the image)
	// - Creates and executes a Session to obtain a Tensor in this normalized form.
	modeldir := flag.String("dir", "", "Directory containing the trained model files. The directory will be created and the model downloaded into it if necessary")
	imagefile := flag.String("image", "", "Path of a JPEG-image to extract labels for")
	flag.Parse()
	if *modeldir == "" || *imagefile == "" {
		flag.Usage()
		return
	}
	// Load the serialized GraphDef from a file.
	modelfile, labelsfile, err := modelFiles(*modeldir)
	if err != nil {
		log.Fatal(err)
	}
	model, err := ioutil.ReadFile(modelfile)
	if err != nil {
		log.Fatal(err)
	}

	// Construct an in-memory graph from the serialized form.
	graph := tf.NewGraph()
	if err := graph.Import(model, ""); err != nil {
		log.Fatal(err)
	}

	// Create a session for inference over graph.
	session, err := tf.NewSession(graph, nil)
	if err != nil {
		log.Fatal(err)
	}
	defer session.Close()

	// Run inference on *imageFile.
	// For multiple images, session.Run() can be called in a loop (and
	// concurrently). Alternatively, images can be batched since the model
	// accepts batches of image data as input.
	tensor, err := makeTensorFromImage(*imagefile)
	if err != nil {
		log.Fatal(err)
	}
	output, err := session.Run(
		map[tf.Output]*tf.Tensor{
			graph.Operation("input").Output(0): tensor,
		},
		[]tf.Output{
			graph.Operation("output").Output(0),
		},
		nil)
	if err != nil {
		log.Fatal(err)
	}
	// output[0].Value() is a vector containing probabilities of
	// labels for each image in the "batch". The batch size was 1.
	// Find the most probably label index.
	probabilities := output[0].Value().([][]float32)[0]
	printBestLabel(probabilities, labelsfile)
}

func printBestLabel(probabilities []float32, labelsFile string) {
	bestIdx := 0
	for i, p := range probabilities {
		if p > probabilities[bestIdx] {
			bestIdx = i
		}
	}
	// Found the best match. Read the string from labelsFile, which
	// contains one line per label.
	file, err := os.Open(labelsFile)
	if err != nil {
		log.Fatal(err)
	}
	defer file.Close()
	scanner := bufio.NewScanner(file)
	var labels []string
	for scanner.Scan() {
		labels = append(labels, scanner.Text())
	}
	if err := scanner.Err(); err != nil {
		log.Printf("ERROR: failed to read %s: %v", labelsFile, err)
	}
	fmt.Printf("BEST MATCH: (%2.0f%% likely) %s\n", probabilities[bestIdx]*100.0, labels[bestIdx])
}

// Convert the image in filename to a Tensor suitable as input to the Inception model.
func makeTensorFromImage(filename string) (*tf.Tensor, error) {
	bytes, err := ioutil.ReadFile(filename)
	if err != nil {
		return nil, err
	}
	// DecodeJpeg uses a scalar String-valued tensor as input.
	tensor, err := tf.NewTensor(string(bytes))
	if err != nil {
		return nil, err
	}
	// Construct a graph to normalize the image
	graph, input, output, err := constructGraphToNormalizeImage()
	if err != nil {
		return nil, err
	}
	// Execute that graph to normalize this one image
	session, err := tf.NewSession(graph, nil)
	if err != nil {
		return nil, err
	}
	defer session.Close()
	normalized, err := session.Run(
		map[tf.Output]*tf.Tensor{input: tensor},
		[]tf.Output{output},
		nil)
	if err != nil {
		return nil, err
	}
	return normalized[0], nil
}

// The inception model takes as input the image described by a Tensor in a very
// specific normalized format (a particular image size, shape of the input tensor,
// normalized pixel values etc.).
//
// This function constructs a graph of TensorFlow operations which takes as
// input a JPEG-encoded string and returns a tensor suitable as input to the
// inception model.
func constructGraphToNormalizeImage() (graph *tf.Graph, input, output tf.Output, err error) {
	// Some constants specific to the pre-trained model at:
	// https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip
	//
	// - The model was trained after with images scaled to 224x224 pixels.
	// - The colors, represented as R, G, B in 1-byte each were converted to
	//   float using (value - Mean)/Scale.
	const (
		H, W  = 224, 224
		Mean  = float32(117)
		Scale = float32(1)
	)
	// - input is a String-Tensor, where the string the JPEG-encoded image.
	// - The inception model takes a 4D tensor of shape
	//   [BatchSize, Height, Width, Colors=3], where each pixel is
	//   represented as a triplet of floats
	// - Apply normalization on each pixel and use ExpandDims to make
	//   this single image be a "batch" of size 1 for ResizeBilinear.
	s := op.NewScope()
	input = op.Placeholder(s, tf.String)
	output = op.Div(s,
		op.Sub(s,
			op.ResizeBilinear(s,
				op.ExpandDims(s,
					op.Cast(s,
						op.DecodeJpeg(s, input, op.DecodeJpegChannels(3)), tf.Float),
					op.Const(s.SubScope("make_batch"), int32(0))),
				op.Const(s.SubScope("size"), []int32{H, W})),
			op.Const(s.SubScope("mean"), Mean)),
		op.Const(s.SubScope("scale"), Scale))
	graph, err = s.Finalize()
	return graph, input, output, err
}

func modelFiles(dir string) (modelfile, labelsfile string, err error) {
	const URL = "https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip"
	var (
		model   = filepath.Join(dir, "tensorflow_inception_graph.pb")
		labels  = filepath.Join(dir, "imagenet_comp_graph_label_strings.txt")
		zipfile = filepath.Join(dir, "inception5h.zip")
	)
	if filesExist(model, labels) == nil {
		return model, labels, nil
	}
	log.Println("Did not find model in", dir, "downloading from", URL)
	if err := os.MkdirAll(dir, 0755); err != nil {
		return "", "", err
	}
	if err := download(URL, zipfile); err != nil {
		return "", "", fmt.Errorf("failed to download %v - %v", URL, err)
	}
	if err := unzip(dir, zipfile); err != nil {
		return "", "", fmt.Errorf("failed to extract contents from model archive: %v", err)
	}
	os.Remove(zipfile)
	return model, labels, filesExist(model, labels)
}

func filesExist(files ...string) error {
	for _, f := range files {
		if _, err := os.Stat(f); err != nil {
			return fmt.Errorf("unable to stat %s: %v", f, err)
		}
	}
	return nil
}

func download(URL, filename string) error {
	resp, err := http.Get(URL)
	if err != nil {
		return err
	}
	defer resp.Body.Close()
	file, err := os.OpenFile(filename, os.O_RDWR|os.O_CREATE, 0644)
	if err != nil {
		return err
	}
	defer file.Close()
	_, err = io.Copy(file, resp.Body)
	return err
}

func unzip(dir, zipfile string) error {
	r, err := zip.OpenReader(zipfile)
	if err != nil {
		return err
	}
	defer r.Close()
	for _, f := range r.File {
		src, err := f.Open()
		if err != nil {
			return err
		}
		log.Println("Extracting", f.Name)
		dst, err := os.OpenFile(filepath.Join(dir, f.Name), os.O_WRONLY|os.O_CREATE, 0644)
		if err != nil {
			return err
		}
		if _, err := io.Copy(dst, src); err != nil {
			return err
		}
		dst.Close()
	}
	return nil
}

Index

• func Version() string
• type Consumer
  • func (p Consumer) DataType() DataType
  • func (p Consumer) Producer() Output
• type Context
  • func NewContext(options *ContextOptions) (*Context, error)
  • func (c *Context) ListDevices() ([]Device, error)
• type ContextOptions
• type DataType
• type Device
  • func (d Device) String() string
• type Graph
  • func NewGraph() *Graph
  • func (g *Graph) AddGradients(prefix string, y []Output, x []Output, dx []Output) ([]Output, error)
  • func (g *Graph) AddOperation(args OpSpec) (*Operation, error)
  • func (g *Graph) Import(def []byte, prefix string) error
  • func (g *Graph) ImportWithOptions(def []byte, options GraphImportOptions) error
  • func (g *Graph) Operation(name string) *Operation
  • func (g *Graph) Operations() []Operation
  • func (g *Graph) WriteTo(w io.Writer) (int64, error)
• type GraphImportOptions
• type Input
• type OpSpec
• type Operation
  • func (op *Operation) Attr(name string) (interface{}, error)
  • func (op *Operation) Device() string
  • func (op *Operation) Name() string
  • func (op *Operation) NumInputs() int
  • func (op *Operation) NumOutputs() int
  • func (op *Operation) Output(i int) Output
  • func (op *Operation) OutputListSize(output string) (int, error)
  • func (op *Operation) Type() string
• type Output
  • func (p Output) Consumers() []Consumer
  • func (p Output) DataType() DataType
  • func (p Output) Shape() Shape
• type OutputList
• type PartialRun
  • func (pr *PartialRun) Run(feeds map[Output]*Tensor, fetches []Output, targets []*Operation) ([]*Tensor, error)
• type SavedModel
  • func LoadSavedModel(exportDir string, tags []string, options *SessionOptions) (*SavedModel, error)
• type Session
  • func NewSession(graph *Graph, options *SessionOptions) (*Session, error)
  • func (s *Session) Close() error
  • func (s *Session) ListDevices() ([]Device, error)
  • func (s *Session) NewPartialRun(feeds, fetches []Output, targets []*Operation) (*PartialRun, error)
  • func (s *Session) Run(feeds map[Output]*Tensor, fetches []Output, targets []*Operation) ([]*Tensor, error)
• type SessionOptions
• type Shape
  • func MakeShape(shape ...int64) Shape
  • func ScalarShape() Shape
  • func (s Shape) IsFullySpecified() bool
  • func (s Shape) NumDimensions() int
  • func (s Shape) Size(dim int) int64
  • func (s Shape) String() string
  • func (s Shape) ToSlice() ([]int64, error)
• type Signature
• type Tensor
  • func NewTensor(value interface{}) (*Tensor, error)
  • func ReadTensor(dataType DataType, shape []int64, r io.Reader) (*Tensor, error)
  • func (t *Tensor) DataType() DataType
  • func (t *Tensor) Shape() []int64
  • func (t *Tensor) Value() interface{}
  • func (t *Tensor) WriteContentsTo(w io.Writer) (int64, error)
• type TensorHandle
  • func NewTensorHandle(t *Tensor) (*TensorHandle, error)
  • func (th *TensorHandle) BackingDeviceName() (string, error)
  • func (th *TensorHandle) CopyToDevice(c *Context, deviceName string) (*TensorHandle, error)
  • func (th *TensorHandle) DataType() DataType
  • func (th *TensorHandle) DeviceName() (string, error)
  • func (th *TensorHandle) Shape() ([]int64, error)
  • func (th *TensorHandle) ToTensor() (*Tensor, error)
• type TensorInfo

Examples

• Package
• PartialRun

Functions

func Version

func Version() string

Version returns a string describing the version of the underlying TensorFlow runtime.

Types

type Consumer

type Consumer struct {
	// Op is the Operation that is consuming the output of another operation.
	Op *Operation

	// Index is the index of the input within Op that the output of another
	// operation is connected to.
	Index int
}

Consumer identifies a specific input of an operation that consumes the output of another operation.

func (Consumer) DataType

func (p Consumer) DataType() DataType

DataType returns the type of the input.

func (Consumer) Producer

func (p Consumer) Producer() Output

Producer returns the Output that is connected to this Consumer.

type Context

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

Context for executing operations eagerly.

A Context allows operations to be executed immediately. It encapsulates information such as the available devices, resource manager etc. It also allows the user to configure execution using a ConfigProto, as they can configure a Session when executing a Graph.

func NewContext

func NewContext(options *ContextOptions) (*Context, error)

NewContext creates a new context for eager execution. options may be nil to use the default options.

func (*Context) ListDevices

func (c *Context) ListDevices() ([]Device, error)

ListDevices returns the list of devices associated with a Context.

type ContextOptions

type ContextOptions struct {
	// Config is a binary-serialized representation of the
	// tensorflow.ConfigProto protocol message
	// (https://www.tensorflow.org/code/tensorflow/core/protobuf/config.proto).
	Config []byte

	// Sets the default execution mode
	Async bool
}

ContextOptions contains configuration information for a session

type DataType

type DataType C.TF_DataType

DataType holds the type for a scalar value. E.g., one slot in a tensor.

const (
	Float      DataType = C.TF_FLOAT
	Double     DataType = C.TF_DOUBLE
	Int32      DataType = C.TF_INT32
	Uint32     DataType = C.TF_UINT32
	Uint8      DataType = C.TF_UINT8
	Int16      DataType = C.TF_INT16
	Int8       DataType = C.TF_INT8
	String     DataType = C.TF_STRING
	Complex64  DataType = C.TF_COMPLEX64
	Complex    DataType = C.TF_COMPLEX
	Int64      DataType = C.TF_INT64
	Uint64     DataType = C.TF_UINT64
	Bool       DataType = C.TF_BOOL
	Qint8      DataType = C.TF_QINT8
	Quint8     DataType = C.TF_QUINT8
	Qint32     DataType = C.TF_QINT32
	Bfloat16   DataType = C.TF_BFLOAT16
	Qint16     DataType = C.TF_QINT16
	Quint16    DataType = C.TF_QUINT16
	Uint16     DataType = C.TF_UINT16
	Complex128 DataType = C.TF_COMPLEX128
	Half       DataType = C.TF_HALF
)

Types of scalar values in the TensorFlow type system.

type Device

type Device struct {
	Name, Type       string
	MemoryLimitBytes int64
}

Device structure contains information about a device associated with a session, as returned by ListDevices()

func (Device) String

func (d Device) String() string

String describes d and implements fmt.Stringer.

type Graph

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

Graph represents a computation graph. Graphs may be shared between sessions.

func NewGraph

func NewGraph() *Graph

NewGraph returns a new Graph.

func (*Graph) AddGradients

func (g *Graph) AddGradients(prefix string, y []Output, x []Output, dx []Output) ([]Output, error)

AddGradients adds operations to compute the partial derivatives of the sum of tensors in y with respect to tensors in x, i.e., d(y[0] + y[1] + ...) / d x[0], d(y[0] + y[1] + ... ) / d x[1] etc.

prefix, if non-empty, is the name prefix used for all operations added to the graph to compute these gradients.

func (*Graph) AddOperation

func (g *Graph) AddOperation(args OpSpec) (*Operation, error)

AddOperation adds an operation to g.

func (*Graph) Import

func (g *Graph) Import(def []byte, prefix string) error

Import imports the nodes and edges from a serialized representation of another Graph into g.

Names of imported nodes will be prefixed with prefix.

func (*Graph) ImportWithOptions

func (g *Graph) ImportWithOptions(def []byte, options GraphImportOptions) error

ImportWithOptions imports the nodes and edges from a serialized representation of another Graph into g.

Multiple options can be specified for the newly imported nodes.

func (*Graph) Operation

func (g *Graph) Operation(name string) *Operation

Operation returns the Operation named name in the Graph, or nil if no such operation is present.

func (*Graph) Operations

func (g *Graph) Operations() []Operation

Operations returns a list of all operations in the graph

func (*Graph) WriteTo

func (g *Graph) WriteTo(w io.Writer) (int64, error)

WriteTo writes out a serialized representation of g to w.

Implements the io.WriterTo interface.

type GraphImportOptions

type GraphImportOptions struct {
	// Node prefix
	Prefix string

	// Execution device
	Device string
}

The GraphImportOptions struct holds parameters for the ImportWithOptions function.

type Input

type Input interface {
	// contains filtered or unexported methods
}

Input is the interface for specifying inputs to an operation being added to a Graph.

Operations can have multiple inputs, each of which could be either a tensor produced by another operation (an Output object), or a list of tensors produced by other operations (an OutputList). Thus, this interface is implemented by both Output and OutputList.

See OpSpec.Input for more information.

type OpSpec

type OpSpec struct {
	// Type of the operation (e.g., "Add", "MatMul").
	Type string

	// Name by which the added operation will be referred to in the Graph.
	// If omitted, defaults to Type.
	Name string

	// Inputs to this operation, which in turn must be outputs
	// of other operations already added to the Graph.
	//
	// An operation may have multiple inputs with individual inputs being
	// either a single tensor produced by another operation or a list of
	// tensors produced by multiple operations. For example, the "Concat"
	// operation takes two inputs: (1) the dimension along which to
	// concatenate and (2) a list of tensors to concatenate. Thus, for
	// Concat, len(Input) must be 2, with the first element being an Output
	// and the second being an OutputList.
	Input []Input

	// Map from attribute name to its value that will be attached to this
	// operation.
	Attrs map[string]interface{}

	// Operations that must be executed before executing the operation
	// being added.
	ControlDependencies []*Operation

	// The device on which the operation should be executed.
	// If omitted, an appropriate device will automatically be selected.
	//
	// For example, if set of "/device:GPU:0", then the operation will
	// execute on GPU #0.
	Device string
}

OpSpec is the specification of an Operation to be added to a Graph (using Graph.AddOperation).

type Operation

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

Operation that has been added to the graph.

func (*Operation) Attr

func (op *Operation) Attr(name string) (interface{}, error)

Attr returns the value of an attribute on op. It returns an error if the attribute does not exist.

func (*Operation) Device

func (op *Operation) Device() string

Device returns a specification of the device on which this operation will be executed, or the empty string if there is no such specification.

func (*Operation) Name

func (op *Operation) Name() string

Name returns the name of the operation.

func (*Operation) NumInputs

func (op *Operation) NumInputs() int

NumInputs returns the number of inputs of op.

func (*Operation) NumOutputs

func (op *Operation) NumOutputs() int

NumOutputs returns the number of outputs of op.

func (*Operation) Output

func (op *Operation) Output(i int) Output

Output returns the i-th output of op.

func (*Operation) OutputListSize

func (op *Operation) OutputListSize(output string) (int, error)

OutputListSize returns the size of the list of Outputs that is produced by a named output of op.

An Operation has multiple named outputs, each of which produces either a single tensor or a list of tensors. This method returns the size of the list of tensors for a specific output of the operation, identified by its name.

func (*Operation) Type

func (op *Operation) Type() string

Type returns the name of the operator used by this operation.

type Output

type Output struct {
	// Op is the Operation that produces this Output.
	Op *Operation

	// Index specifies the index of the output within the Operation.
	Index int
}

Output represents one of the outputs of an operation in the graph. Has a DataType (and eventually a Shape). May be passed as an input argument to a function for adding operations to a graph, or to a Session's Run() method to fetch that output as a tensor.

func (Output) Consumers

func (p Output) Consumers() []Consumer

Consumers returns the inputs that consume this output.

func (Output) DataType

func (p Output) DataType() DataType

DataType returns the type of elements in the tensor produced by p.

func (Output) Shape

func (p Output) Shape() Shape

Shape returns the (possibly incomplete) shape of the tensor produced p.

type OutputList

type OutputList []Output

OutputList represents a list of Outputs that can be provided as input to another operation.

type PartialRun

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

PartialRun enables incremental evaluation of graphs.

PartialRun allows the caller to pause the evaluation of a graph, run arbitrary code that depends on the intermediate computation of the graph, and then resume graph execution. The results of the arbitrary code can be fed into the graph when resuming execution. In contrast, Session.Run executes the graph to compute the requested fetches using the provided feeds and discards all intermediate state (e.g., value of intermediate tensors) when it returns.

For example, consider a graph for unsupervised training of a neural network model. PartialRun can be used to pause execution after the forward pass of the network, let the caller actuate the output (e.g., play a game, actuate a robot etc.), determine the error/loss and then feed this calculated loss when resuming the backward pass of the graph.

Example

var (
	// Create a graph: a + 2 + 3 + b.
	//
	// Skipping error handling for brevity of this example.
	// The 'op' package can be used to make graph construction code
	// with error handling more succinct.
	g        = NewGraph()
	a, _     = Placeholder(g, "a", Int32)
	b, _     = Placeholder(g, "b", Int32)
	two, _   = Const(g, "Two", int32(2))
	three, _ = Const(g, "Three", int32(3))

	plus2, _ = Add(g, "plus2", a, two)       // a + 2
	plus3, _ = Add(g, "plus3", plus2, three) // (a + 2) + 3
	plusB, _ = Add(g, "plusB", plus3, b)     // ((a + 2) + 3) + b

)
sess, err := NewSession(g, nil)
if err != nil {
	panic(err)
}
defer sess.Close()

// All the feeds, fetches and targets for subsequent PartialRun.Run
// calls must be provided at setup.
pr, err := sess.NewPartialRun(
	[]Output{a, b},
	[]Output{plus2, plusB},
	[]*Operation{plus3.Op},
)
if err != nil {
	panic(err)
}

// Feed 'a=1', fetch 'plus2', and compute (but do not fetch) 'plus3'.
// Imagine this to be the forward pass of unsupervised neural network
// training of a robot.
val, _ := NewTensor(int32(1))
fetches, err := pr.Run(
	map[Output]*Tensor{a: val},
	[]Output{plus2},
	nil)
if err != nil {
	panic(err)
}
v1 := fetches[0].Value().(int32)

// Now, feed 'b=4', fetch 'plusB=a+2+3+b'
// Imagine this to be the result of actuating the robot to determine
// the error produced by the current state of the neural network.
val, _ = NewTensor(int32(4))
fetches, err = pr.Run(
	map[Output]*Tensor{b: val},
	[]Output{plusB},
	nil)
if err != nil {
	panic(err)
}
v2 := fetches[0].Value().(int32)

fmt.Println(v1, v2)

Output:

3 10

func (*PartialRun) Run

func (pr *PartialRun) Run(feeds map[Output]*Tensor, fetches []Output, targets []*Operation) ([]*Tensor, error)

Run resumes execution of the graph to compute the requested fetches and targets with the provided feeds.

type SavedModel

type SavedModel struct {
	Session    *Session
	Graph      *Graph
	Signatures map[string]Signature
}

SavedModel represents the contents of loaded SavedModel. TODO(jhseu): Add and document metagraphdef when we pregenerate protobufs.

func LoadSavedModel

func LoadSavedModel(exportDir string, tags []string, options *SessionOptions) (*SavedModel, error)

LoadSavedModel creates a new SavedModel from a model previously exported to a directory on disk.

Exported models contain a set of graphs and, optionally, variable values. Tags in the model identify a single graph. LoadSavedModel initializes a session with the identified graph and with variables initialized to from the checkpoints on disk.

The tensorflow package currently does not have the ability to export a model to a directory from Go. This function thus currently targets loading models exported in other languages, such as using tf.saved_model.builder in Python. See: https://www.tensorflow.org/code/tensorflow/python/saved_model/

type Session

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

Session drives a TensorFlow graph computation.

When a Session is created with a given target, a new Session object is bound to the universe of resources specified by that target. Those resources are available to this session to perform computation described in the GraphDef. After creating the session with a graph, the caller uses the Run() API to perform the computation and potentially fetch outputs as Tensors. A Session allows concurrent calls to Run().

func NewSession

func NewSession(graph *Graph, options *SessionOptions) (*Session, error)

NewSession creates a new execution session with the associated graph. options may be nil to use the default options.

func (*Session) Close

func (s *Session) Close() error

Close a session. This contacts any other processes associated with this session, if applicable. Blocks until all previous calls to Run have returned.

func (*Session) ListDevices

func (s *Session) ListDevices() ([]Device, error)

ListDevices returns the list of devices associated with a Session.

func (*Session) NewPartialRun

func (s *Session) NewPartialRun(feeds, fetches []Output, targets []*Operation) (*PartialRun, error)

NewPartialRun sets up the graph for incremental evaluation.

All values of feeds, fetches and targets that may be provided to Run calls on the returned PartialRun need to be provided to NewPartialRun.

See documentation for the PartialRun type.

func (*Session) Run

func (s *Session) Run(feeds map[Output]*Tensor, fetches []Output, targets []*Operation) ([]*Tensor, error)

Run the graph with the associated session starting with the supplied feeds to compute the value of the requested fetches. Runs, but does not return Tensors for operations specified in targets.

On success, returns the fetched Tensors in the same order as supplied in the fetches argument. If fetches is set to nil, the returned Tensor fetches is empty.

type SessionOptions

type SessionOptions struct {
	// Target indicates the TensorFlow runtime to connect to.
	//
	// If 'target' is empty or unspecified, the local TensorFlow runtime
	// implementation will be used.  Otherwise, the TensorFlow engine
	// defined by 'target' will be used to perform all computations.
	//
	// "target" can be either a single entry or a comma separated list
	// of entries. Each entry is a resolvable address of one of the
	// following formats:
	//   local
	//   ip:port
	//   host:port
	//   ... other system-specific formats to identify tasks and jobs ...
	//
	// NOTE: at the moment 'local' maps to an in-process service-based
	// runtime.
	//
	// Upon creation, a single session affines itself to one of the
	// remote processes, with possible load balancing choices when the
	// "target" resolves to a list of possible processes.
	//
	// If the session disconnects from the remote process during its
	// lifetime, session calls may fail immediately.
	Target string

	// Config is a binary-serialized representation of the
	// tensorflow.ConfigProto protocol message
	// (https://www.tensorflow.org/code/tensorflow/core/protobuf/config.proto).
	Config []byte
}

SessionOptions contains configuration information for a session.

type Shape

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

Shape represents the (possibly partially known) shape of a tensor that will be produced by an operation.

The zero-value of a Shape represents a shape with an unknown number of dimensions.

func MakeShape

func MakeShape(shape ...int64) Shape

MakeShape returns a Shape with the provided size of each dimension.

A value of -1 implies that the size of the corresponding dimension is not known.

func ScalarShape

func ScalarShape() Shape

ScalarShape returns a Shape representing a scalar.

func (Shape) IsFullySpecified

func (s Shape) IsFullySpecified() bool

IsFullySpecified returns true iff the size of all the dimensions of s are known.

func (Shape) NumDimensions

func (s Shape) NumDimensions() int

NumDimensions returns the number of dimensions represented by s, or -1 if unknown.

func (Shape) Size

func (s Shape) Size(dim int) int64

Size returns the size of the dim-th dimension of the shape, or -1 if it is unknown.

REQUIRES: 0 <= dim < s.NumDimensions()

func (Shape) String

func (s Shape) String() string

func (Shape) ToSlice

func (s Shape) ToSlice() ([]int64, error)

ToSlice returns the (possibly partially known) shape represented by s as a slice, or an error if the number of dimensions is not known.

type Signature

type Signature struct {
	Inputs, Outputs map[string]TensorInfo
	MethodName      string
}

A Signature defines the signature of a computation supported by a TensorFlow graph.

For example, a model with two loss computations, sharing a single input, might have the following signature_def map.

Note that across the two Signatures "loss_A" and "loss_B", the input key, output key, and method_name are identical, and will be used by system(s) that implement or rely upon this particular loss method. The output tensor names differ, demonstrating how different outputs can exist for the same method.

signature_def {
  key: "loss_A"
  value {
    inputs {
      key: "input"
      value {
        name: "input:0"
        dtype: DT_STRING
        tensor_shape: ...
      }
    }
    outputs {
      key: "loss_output"
      value {
        name: "loss_output_A:0"
        dtype: DT_FLOAT
        tensor_shape: ...
      }
    }
  }
  ...
  method_name: "some/package/compute_loss"
}

signature_def {
  key: "loss_B"
  value {
    inputs {
      key: "input"
      value {
        name: "input:0"
        dtype: DT_STRING
        tensor_shape: ...
      }
    }
    outputs {
      key: "loss_output"
      value {
        name: "loss_output_B:0"
        dtype: DT_FLOAT
        tensor_shape: ...
      }
    }
  }
  ...
  method_name: "some/package/compute_loss"
}

type Tensor

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

Tensor holds a multi-dimensional array of elements of a single data type.

func NewTensor

func NewTensor(value interface{}) (*Tensor, error)

NewTensor converts from a Go value to a Tensor. Valid values are scalars, slices, and arrays. Every element of a slice must have the same length so that the resulting Tensor has a valid shape.

func ReadTensor

func ReadTensor(dataType DataType, shape []int64, r io.Reader) (*Tensor, error)

ReadTensor constructs a Tensor with the provided type and shape from the serialized tensor contents in r.

See also WriteContentsTo.

func (*Tensor) DataType

func (t *Tensor) DataType() DataType

DataType returns the scalar datatype of the Tensor.

func (*Tensor) Shape

func (t *Tensor) Shape() []int64

Shape returns the shape of the Tensor.

func (*Tensor) Value

func (t *Tensor) Value() interface{}

Value converts the Tensor to a Go value. For now, not all Tensor types are supported, and this function may panic if it encounters an unsupported DataType.

The type of the output depends on the Tensor type and dimensions. For example: Tensor(int64, 0): int64 Tensor(float64, 3): [][][]float64

func (*Tensor) WriteContentsTo

func (t *Tensor) WriteContentsTo(w io.Writer) (int64, error)

WriteContentsTo writes the serialized contents of t to w.

Returns the number of bytes written. See ReadTensor for reconstructing a Tensor from the serialized form.

WARNING: WriteContentsTo is not comprehensive and will fail if t.DataType() is non-numeric (e.g., String). See https://github.com/tensorflow/tensorflow/issues/6003.

type TensorHandle

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

TensorHandle is a handle to a tensor on a device.

A Tensor referenced by a TensorHandle may be on any device, whereas a Tensor always resides in the host CPU's memory.

A Tensor referenced by a TensorHandle may not have been computed yet. For example, a TensorHandle might reference the output of an operation that has not finished executing. Because of this, various methods, such as Shape() may block until the tensor has been instantiated.

This allows multiple operations to be performed on tensors on a device (e.g. a GPU) without sending these values back to the host CPU in between every operation.

func NewTensorHandle

func NewTensorHandle(t *Tensor) (*TensorHandle, error)

NewTensorHandle creates a new tensor handle from a tensor.

func (*TensorHandle) BackingDeviceName

func (th *TensorHandle) BackingDeviceName() (string, error)

BackingDeviceName returns the name of the device in whose memory the tensor handle resides. This function will block till the operation that produces `h` has completed.

WARNING: The implementation currently returns the same as DeviceName(). After TensoFlow 1.13's C library is released, this implementation will be updated to return what the documentation says!

func (*TensorHandle) CopyToDevice

func (th *TensorHandle) CopyToDevice(c *Context, deviceName string) (*TensorHandle, error)

CopyToDevice creates a new TensorHandle with the same contents as this TensorHandle but placed in the memory of the device 'deviceName'. If source and destination are the same device, then this creates a new handle that shares the underlying buffer. Otherwise, it currently requires at least one of the source or destination devices to be CPU (i.e., for the source or destination tensor to be placed in host memory).

func (*TensorHandle) DataType

func (th *TensorHandle) DataType() DataType

DataType returns the TensorHandle's datatype.

func (*TensorHandle) DeviceName

func (th *TensorHandle) DeviceName() (string, error)

DeviceName returns the name of the device of the operation that produced the TensorHandle. If the handle was produced by a copy, it returns the destination device of the copy. Note that returned device name is not always the device holding the tensor handle's memory. If you want the latter, use BackingDeviceName. This function will block till the operation that produces th has completed.

func (*TensorHandle) Shape

func (th *TensorHandle) Shape() ([]int64, error)

Shape returns the shape of the Tensor referenced by th.

func (*TensorHandle) ToTensor

func (th *TensorHandle) ToTensor() (*Tensor, error)

ToTensor returns the Tensor referenced by th. It may block if this tensor is not yet computed.

type TensorInfo

type TensorInfo struct {
	Name  string
	DType DataType
	Shape Shape
}

A TensorInfo contains the information about a Tensor necessary for feeding or retrieval.