Perfect TensorFlow API Guide

November 7, 2017 · View on GitHub

Perfect TensorFlow is a Swift Class Wrapper of TensorFlow's C API to build, save, load and execute TensorFlow models on Server Side Swift.

⚠️WARNING⚠️: The API is currently experimental and is not covered by TensorFlow API stability guarantees.

The Perfect TensorFlow Computer Vision Server example demonstrates use of this API to classify images using a pre-trained Inception architecture convolutional neural network. It demonstrates:

Graph construction: using the OperationBuilder class to construct a graph to decode, resize and normalize a JPEG image. Model loading: Using Graph.import() to load a pre-trained Inception model. Graph execution: Using a Session to execute the graphs and find the best label for an image.

Please install Perfect TensorFlow library before using.

Classes

There are quite a few classes in Perfect-TensorFlow. Relationship between these classes can be described as the workflow below:

TFLib: TensorFlow C API DLL Low levelled library
  |
TensorFlow: Runtime Library
  |
  |------ Shape: dimension info of a tensor
  |         |
  |       Tensor: multi-dimensional array
  |         |
 Graph -- OperationBuilder: construct an operation from tensors
  |         |
  |       Operation: a graph node that performs computation on tensors
  |         |
  |       Output: a symbolic handle to a tensor produced by an Operation
  |         |
  |------ Session: driver for graph execution.
            |
          Runner: Run Operations and evaluate tensors

Content

Class Name	Description
TFLib	TensorFlow C API low levelled DLL library - suggested for internal use only.
TensorFlow	namespace describing the TensorFlow runtime.
Shape	The possibly partially known shape of a tensor produced by an operation; create from an Int64 array
Tensor	A typed multi-dimensional array; can be created either from a scalar or a typed array.
OperationBuilder	A builder for Operations in a Graph.
Operation	A Graph node that performs computation on Tensors.
Output	A symbolic handle to a tensor produced by an Operation.
Graph	A data flow graph representing a TensorFlow computation.
Session Runner	Run Operations and evaluate Tensors.

Protocol Buffers

Perfect TensorFlow depends heavily on Google Protocol Buffers, which means that you can easily save / load most of the objects from / into persistent files or database records in form of either binary bytes or JSON string.

Take Graph as a typical example, you can easily load a pre-trained model from a third party into your Swift program, as the digest source code in Perfect TensorFlow Demo of Computer Vision:

import PerfectLib
import PerfectTensorFlow

typealias TF = TensorFlow

// load a third party graph protocol buffer file:
let fModel = File("/tmp/tensorflow_inception_graph.pb")
try fModel.open(.read)
let modelBytes = try fModel.readSomeBytes(count: fModel.size)
fModel.close()

// import buffer file into Graph Object:
let def = try TF.GraphDef(Data(bytes: modelBytes))
let g = try TF.Graph()
try g.import(definition: def)

⚠️NOTE⚠️ You can get the data buffer by let data = try def.serializedData() or let json = try def.jsonString() vice versa.

TensorFlow Runtime

Class TensorFlow is actually the namespace that represents the TensorFlow runtime.

To align with other language implementation such as Python or Java, make an alias of class TensorFlow is a good idea:

/// load the TensorFlow C API DLL library
import TensorFlowAPI
import PerfectTensorFlow
public typealias TF = TensorFlow

⚠️NOTE⚠️ Prior to use ANY ACTUAL FUNCTIONS of Perfect TensorFlow framework, TF.Open() must be called first:

// this action will load all api functions defined 
// in /usr/local/lib/libtensorflow.so
try TF.Open()

Please also note that you can active the library with a specific path, alternatively, especially in case of different versions or CPU/GPU library adjustment required:

// this action will load the library with the path
try TF.Open("/path/to/DLL/of/libtensorflow.so")

Shape

In Perfect TensorFlow definition, Shape is a struct that holding dimensions for a future tensor:

  /// Tensor Shape Object
  public struct Shape {
    public var dimensions = [Int64]()
  }//end class

Tensor

A Tensor is defined as a typed multi-dimensional array.

Generally you don't have to call the initiator of Tensor object directly, but using two static methods will be more efficient: Scalar for one dimensional constant and Array for a multi-dimensional array.

Scalar

Scalar Tensor is actually a one dimensional array (or zero dimensional in an other term). To create such a tensor, just pass the variable to the static method Scalar<T>(value: T):

// create an Int32 scalar tensor; 
// Please note that Int will be Int64 actually
let x = try TF.Tensor.Scalar(Int32(100))

// create a Float scalar tensor;
// Please note that constant without Float type cast will be Double indeed
let y = try TF.Tensor.Scalar(Float(1.1))

// create a String scalar tensor:
let s = try TF.Tensor.Scalar("Hello, TensorFlow! 🇨🇳🇨🇦")

To get the value of Tensor as a scalar, use asScalar<T>() method:

let x: Int32 = try tensor.asScalar()

Array

Array Tensor requires shape information to create; Also please note that the array has to be flatten to generate the expected array:

/* here is the matrix to input:
| 1 2 |
| 3 4 |
*/
// flatten the matrix first
let m:[Float] = [[1, 2], [3, 4]].flatMap { \$0 }

// then turn it into a tensor with shape info.
let tensor = try TF.Tensor.Array(dimenisons: [2,2], value: m)

To get the value of Tensor as an array, use asArray<T>() method:

let y: [Float] = try tensor.asArray()

⚠️NOTES⚠️ If the tensor is a string array, please directly call the strings property, which will actually return an array of Data:

// datastr is actually [Data]
let datastr = tensor.strings

// translate it into [String]
let str = datastr.map { \$0.string }

Matrix

Since Perfect-TensorFlow v1.2.1, however, you can apply a multi-dimensional array to a tensor as a matrix without considering the shape or dimensions. The equivalent example of above will look like:

let M = try TF.Tensor.Matrix([[1, 2], [3, 4]])

⚠️NOTES⚠️ Element in a Matrix must be number!

Raw Data of a Tensor

To access raw data of a tensor, you can use either data property, or withDataPointer() method with better performance but tricky pointer operations - and property of bytesCount and type will be useful in this case:

public class Tensor {
	/// get a buffer copy from the tensor value
	public var data: [Int8]
	
	/// get total size of memory in bytes
    public var bytesCount: Int
    
    /// check data type of the value / element of value array
    public var `type`: DataType? 
    
    /// perform data pointer operations unsafely.
    public func withDataPointer<R>(body: (UnsafeMutableRawPointer) throws -> R) throws -> R
}

Shape of a Tensor

To get the dimension of a tensor, use dim property:

public class Tensor {
	public var dim: [Int64]
}

Operation and Its Builder

An operation is a graph node that performs computation on Tensors.

To create an operation in a graph, check the workflow below:

export an OperationBuilder object instance from a graph
set the operation name, type and device
set the tensors of the operation
add input / output
set other attributes of the operation
call build() method and return an expected operation.

Here is an example:

// make a tensor, e.g., Tensor.Scalar(100)
let tensor = try TF.Tensor ...

let g = try TF.Graph()

// export an instance from the graph and set the name / type
let operation = try g.opBuilder(name: "Const_0", type: `Const`)

	// set attributes
	.set(attributes: ["value": tensor, "dtype": tensor.type])
	
	// build the operation
	.build()

OperationBuilder

OperationBuilder is defined as below:

public class OperationBuilder {
	 /// constructor; using of graph.opBuilder() is more recommended.
    public init(graph: Graph, name: String, `type`: String) throws
    
    /// add an input to the operation
    public func add(input: Output) -> OperationBuilder 

    /// add an input array to the operation
    public func add(inputs: [Output]) throws -> OperationBuilder 

    /// add a control to the builder
    public func add(control: Operation) -> OperationBuilder 

    /// build the operation
    public func build() throws -> Operation 

    /// set the device
    public func `set`(device: String) -> OperationBuilder 
    
    /// set attributes for the operation to build.
	 /// parameter attributes: a dictionary of attributes of the operation, 
	 /// key for the attribute name. 
	 /// Valid attributes include Int64, [Int64], Float, [Float], 
	 /// Bool, [Bool], DataType, [DataType], String, [String], 
	 /// Shape, [Shape], Tensor, [Tensor], 
	 /// TensorProto, [TensorProto], Data
    public func `set`(attributes: [String: Any] = [:])

Operation

Once built by a OperationBuilder, all preset attributes and values of an operation can be read by the following properties and methods below:

Instance Interface	Type	Description
fun attribute(forKey: String)	AttrValue	lookup an attribute in the current operation
var NodeDefinition	NodeDef	get node definition
var name	String	get operation name
var type	String	get operation type
var device	String	get operation device
var numberOfInputs	Int	get number of inputs of the operation
var numberOfOutputs	Int	get number of outputs of the operation
func sizeOfInputList(argument: String)	Int	get size of the input list array by an argument string
func sizeOfOutputList(argument: String)	Int	get size of the output list array by an argument string
var controlInputs	[Operation]	get control inputs as an operation array
func asInput(_ index:Int)	Input	generate an Input from the current operation
func asOutput(_ index:Int)	Output	generate an Output from the current operation

Besides, Operation Class also have three different static methods:

Static Method	Description
func Consumers(output: Output) -> [Input]	get the consumers of an output
func TypeOf(input: Input) -> DataType?	get type of the input
func TypeOf(output: Output) -> DataType?	get type of the output

⚠️NOTES⚠️ More convenient methods to create an operation from a tensor in a graph can be found in the Chapter Graph

Output

Output is a symbolic handle using in TensorFlow and can be generate from Operation, you can simply treat an Output as "an Operation with an index number": let output = op.asOutput(0)

Graph

Graph is a data flow graph representing a TensorFlow computation. There are several different approaches to create a Graph object:

public class Graph {

	/// create a blank graph, just `let x = try TF.Graph()`
	public init () throws 
	
	/// create a graph from a reference pointer, 
	/// usually take place in array operation.
	public init(handle: OpaquePointer) 
}

Besides, you can also import a pre-trained model into the graph, where definition is a GraphDef Protobuf which can be read / write from a binary file:

let g = try TF.Graph()
try g.import(definition: def)

Operations in a Graph

You can get all operations in a graph by accessing property operations, or using searchOperation(forName) to search an operation by name:

// search an operation in a graph instance:
let placeholder = try graph.searchOperation(forName: "placeholder")

// retrieve all operations into an array:
let list = graph.operations
// list is [Operation] now.

Common Operations

There is also a rich and growing set of methods for construct operations from different tensors in a graph, check the table below:

Name	Description	Example
const	create an operation from a constant tensor	`let x = try graph.const(tensor: t, name: "Const_0")`
placeholder	create a placeholder operation	`let feed = try graph.placeholder(name: "feed")`
binaryOp	any binary operations between two outputs	`func binaryOp(_` type`: String, _ in1: Output, _ in2: Output, name: String = "", index: Int = 0) throws -> Output`
scalar	create an Int32 scalar	`let ten = try graph.scalar(10, name: "ten")`
scalar	create an Float32 scalar	`let point = try graph.scalar(0.1, name: "point")`
constantArray	create an array tensor and attach it to the graph	`let size = try g.constantArray(name: "size", value: [1024,768])`
add	add two outputs	`let sum = try graph.add(left:lOutput, right:rOutput, name: "sum")`
add	add two operations (indices are automatically assuming to zero)	`let sum = try graph.add(left:lOp, right:rOp, name: "sumOp")`
neg	get the negative operation	`let neg = try graph.neg(op, name: "MyNeg")`
matMul	Matrix Multiply	`let m = try graph.matMul(l: aOp, r: bOp, name: "m0", transposeA:false, transposeB: true)` will perform an A x B^ multiplication
div	y Output divided by x Output	`let z = try graph.div(x: x, y: y, name: "Div0")`
sub	y Output subtracted by x Output	`let z = try graph.sub(x: x, y: y, name: "Div0")`
decodeJpeg	decode a JPEG picture	`let decoded = try graph.decodeJpeg(content: input, channels: 3)`
resizeBilinear	resize a bilinear image	`let resizes = try g.resizeBilinear(images: images, size: size)`
cast	cast an output to a specific data type	`let cast = try g.cast(value: jpeg, dtype: TF.DataType.dtFloat)`
expandDims	expand dimensions	`let images = try g.expandDims(input: cast, dim: batch)`
addGradients	partial derivatives	`let dy = try g.addGradients(y: y, x: x, dx: [dx])`

A Full demo of almost all the express operation creations above can be found on the source code of Perfect TensorFlow Computer Vision Demo

⚠️NOTE⚠️ To load a graph from a saved model, use graph.load() method and will get a session object, see Load A Session

Session

A session object represents a round of execution of all computations defined in a graph.

The best practice to create a session is using a chain of operations of Runner object. Here is an example:

let results = try graph.runner()
	.feed("input", tensor: image)
	.fetch("output")
	.run()

Let's break it down a bit. The first step is to get a session runner instance: let r = try graph.runner().

Then you can feed some inputs into the runner:

class Runner {
	/// feed the session by using an output as input with a tensor
	public func feed(_ output: Output, tensor: Tensor) -> Runner
	
	/// feed the session by an operation with a tensor
	public func feed(_ operation: Operation, index: Int = 0, tensor: Tensor) -> Runner 
	
	/// feed the session by an operation name with a tensor, 
	/// in this case the graph will search for the operation by name first
	public func feed(_ operation: String, index: Int = 0, tensor: Tensor) throws -> Runner
}

Similar to feed, you can also indicate the session runner to fetch something out after the running:

class Runner {
	/// fetch the session to a specific output
	public func fetch(_ output: Output) -> Runner
	
	/// fetch the session to a specific operation. 
	/// if index = 0, than it will be equivalent to fetch(output)
	public func fetch(_ operation: Operation, index: Int = 0) -> Runner
	
	/// fetch the session to a specific operation by its name
	public func fetch(_ operation: String, index: Int = 0) throws -> Runner
}

Then you can add target operations to the session as well:

class Runner {
	/// add an operation to the target
	public func addTarget(_ operation: Operation) -> Runner 
	
	/// add an operation to the target by its name
	public func addTarget(_ operation: String) throws -> Runner
}

The final step of runner creation is to call the run() method and the result is an array of tensor:

class Runner {
	/// Execute the graph fragments necessary to compute all requested fetches.
	public func run() throws -> [Tensor]
}

Since you may want to feed / fetch / add targets to the session with more than one time, you can chain everything up like this:

let r = try graph
	.feed(op1, tensor: t1).feed(op2, tensor: t2)....
	.fetch(out1).fetch(out2).fetch(out3)....
	.addTarget("opA").addTarget("opB").addTarget("opC")...
	.run()

Load A Session Runner

To import a session object from a previously saved model, firstly create a blank graph, then use import method to load it back:

let g = try TF.Graph()

// the meta signature info defined in a saved model
let metaBuf = try TF.Buffer()

// load the session runner
let runner = try g.load(
	exportDir: "/path/to/saved/model",
	tags: ["tag1", "tag2", ...],
	metaGraphDef: metaBuf)

In this case, you can check the meta data stored in this previously saved model, if available, for more information:

if let data = metaBuf.data {
	let meta = try TF.MetaGraphDef(serializedData: data)
	let signature_def = meta.signatureDef["some signatures ..."] 
	let input_name = signature_def.inputs["name of inputs ..."]?.name
	let output_name = signature_def.outputs["name of outputs"]?.name
	...
}

If ready, you can call runner.run() as the previous chapter.

Device List of a Session

Since TensorFlow 1.3.0+, you can access device information by calling session.devices property, which will return a tuple array:

let dev = try g.runner().session.devices
print(dev)
// sample output:
// ["/job:localhost/replica:0/task:0/cpu:0": (type: "CPU", memory: 268435456)]

Functions of a Graph

Since TensorFlow 1.4.0+, a new feature named Function has been introduced as a nested class of Graph.

A Function object can be either built by graph.toFunction() or imported by its definition protocol buffer:

// build a function from the current graph
let function = try graph.toFunction("funcName", 
	operations:[], inputs:[] outputs: [operation.asOutput(0)], 
	outputNames: [], description: "myFunc")

// get the function definition buffer:
guard let def = function.definition else {
	// something wrong
}

// import a function from its definition buffer
let function2 = try TF.Graph.Function(importDefinition: def)
// now function == function2

The purpose of Function is to copy its gradient function from one graph to another:

let function = try graph1.toFunction(...)
try graph.copy(function: function)

You can also get and set a function object's attribute:

// assuming value is a TF.AttrValue protocol buffer
try function.setAttributeFor("foo_attr", value: value)
let value2 = try function.getAttributeFor("foo_attr")
// now value2 should be the same as value