Julia神经网络编码与PyPy的速度相同

Question

所以我在Python中有一些神经网络代码,我在Julia中重写了这些代码.直接的Python代码在大约7秒内运行,而Julia和PyPy代码在大约0.75秒内运行

sigmoid(z::Float64) = 1/(1 + exp(-z))
sigmoidPrime(z::Float64) = sigmoid(z) * (1 - sigmoid(z))

### Types ###

abstract AbstractNode

type Edge
    source::AbstractNode
    target::AbstractNode
    weight::Float64
    derivative::Float64
    augmented::Bool

    Edge(source::AbstractNode, target::AbstractNode) = new(source, target, randn(1,1)[1], 0.0, false)
end

type Node <: AbstractNode
    incomingEdges::Vector{Edge}
    outgoingEdges::Vector{Edge}
    activation::Float64
    activationPrime::Float64

    Node() = new([], [], -1.0, -1.0)
end

type InputNode <: AbstractNode
    index::Int
    incomingEdges::Vector{Edge}
    outgoingEdges::Vector{Edge}
    activation::Float64

    InputNode(index::Int) = new(index, [], [], -1.0)
end

type BiasNode <: AbstractNode
    incomingEdges::Vector{Edge}
    outgoingEdges::Vector{Edge}
    activation::Float64

    BiasNode() = new([], [], 1.0)
end

type Network
    inputNodes::Vector{InputNode}
    hiddenNodes::Vector{Node}
    outputNodes::Vector{Node}

    function Network(sizes::Array, bias::Bool=true)
        inputNodes = [InputNode(i) for i in 1:sizes[1]];
        hiddenNodes = [Node() for _ in 1:sizes[2]];
        outputNodes = [Node() for _ in 1:sizes[3]];

        for inputNode in inputNodes
            for node in hiddenNodes
                edge = Edge(inputNode, node);
                push!(inputNode.outgoingEdges, edge)
                push!(node.incomingEdges, edge)
            end
        end

        for node in hiddenNodes
            for outputNode in outputNodes
                edge = Edge(node, outputNode);
                push!(node.outgoingEdges, edge)
                push!(outputNode.incomingEdges, edge)
            end
        end

        if bias == true
            biasNode = BiasNode()
            for node in hiddenNodes
                edge = Edge(biasNode, node);
                push!(biasNode.outgoingEdges, edge)
                push!(node.incomingEdges, edge)
            end
        end

        new(inputNodes, hiddenNodes, outputNodes)
    end
end


### Methods ###

function evaluate(obj::Node, inputVector::Array)
    if obj.activation > -0.5
        return obj.activation
    else
        weightedSum = sum([d.weight * evaluate(d.source, inputVector) for d in obj.incomingEdges])
        obj.activation = sigmoid(weightedSum)
        obj.activationPrime = sigmoidPrime(weightedSum)

        return obj.activation
    end
end

function evaluate(obj::InputNode, inputVector::Array)
    obj.activation = inputVector[obj.index]
    return obj.activation
end

function evaluate(obj::BiasNode, inputVector::Array)
    obj.activation = 1.0
    return obj.activation
end

function updateWeights(obj::AbstractNode, learningRate::Float64)
    for d in obj.incomingEdges
        if d.augmented == false
            d.augmented = true
            d.weight -= learningRate * d.derivative
            updateWeights(d.source, learningRate)
            d.derivative = 0.0
        end
    end
end

function compute(obj::Network, inputVector::Array)
    output = [evaluate(node, inputVector) for node in obj.outputNodes]
    for node in obj.outputNodes
        clear(node)
    end
    return output
end

function clear(obj::AbstractNode)
    for d in obj.incomingEdges
        obj.activation = -1.0
        obj.activationPrime = -1.0
        d.augmented = false
        clear(d.source)
    end
end

function propagateDerivatives(obj::AbstractNode, error::Float64)
    for d in obj.incomingEdges
        if d.augmented == false
            d.augmented = true
            d.derivative += error * obj.activationPrime * d.source.activation
            propagateDerivatives(d.source, error * d.weight * obj.activationPrime)
        end
    end
end

function backpropagation(obj::Network, example::Array)
    output = [evaluate(node, example[1]) for node in obj.outputNodes]
    error = output - example[2]
    for (node, err) in zip(obj.outputNodes, error)
        propagateDerivatives(node, err)
    end

    for node in obj.outputNodes
        clear(node)
    end
end

function train(obj::Network, labeledExamples::Array, learningRate::Float64=0.7, iterations::Int=10000)
    for _ in 1:iterations
        for ex in labeledExamples
            backpropagation(obj, ex)
        end

        for node in obj.outputNodes
            updateWeights(node, learningRate)
        end

        for node in obj.outputNodes
            clear(node)
        end
    end
end


labeledExamples = Array[Array[[0,0,0], [0]],
                        Array[[0,0,1], [1]],
                        Array[[0,1,0], [0]],
                        Array[[0,1,1], [1]],
                        Array[[1,0,0], [0]],
                        Array[[1,0,1], [1]],
                        Array[[1,1,0], [1]],
                        Array[[1,1,1], [0]]];

neuralnetwork = Network([3,4,1])
@time train(neuralnetwork, labeledExamples)

我没有提供Python代码,因为我不确定它是否有必要(但是如果你真的想要它我会的话),我当然不希望花大量的时间来完全理解这段代码,我只是基本上寻找与正确的Julia实现相关的明显/系统的低效率(与算法本身相反).

我这样做的动机是以这种方式设计神经网络比矢量化算法和使用Numpy要自然得多,但当然所有循环和跳过类结构的人在Python中都很慢.

因此,这似乎是移植到Julia的自然选择,看看我是否无法获得一些主要的加速,虽然比直接Python快一个数量级的速度很酷,我真正希望的是一个数量级的速度通过PyPy(我在网上找到的一些基准测试似乎表明这是一个合理的期望).

注意:这必须在Julia 0.3中运行才能工作

Answer 1

这似乎更像是一个代码审查而不是一个问题(没有任何问号),但无论如何我都会采取行动.唯一明显的潜在性能问题是你通过理解来分配数组evaluate,compute和backpropagation.evaluate作为for循环,加权和计算将更有效.对于其他两种方法,您可能希望使用预先分配的数组而不是理解.您可以使用Julia的内置分析器来查看代码在大部分时间内花费的时间 - 这可能会显示一些您可以进一步优化的非明显热点.

关于与PyPy的比较,Julia和PyPy很可能在这个代码上做得很好 - 在C性能或接近C性能 - 在这种情况下你不会指望Julia比PyPy快得多,因为它们都很接近最佳.与C实现的性能相比,它将提供非常丰富的信息,因为它将显示Julia和PyPy在桌面上留下了多少性能.幸运的是,这段代码似乎很容易移植到C.