Python Julia神经网络编码速度与PyPy相同
我用Python编写了一些神经网络代码,我在Julia中重写了这些代码。纯Python代码运行时间约为7秒,而Julia和PyPy代码运行时间约为0.75秒Python Julia神经网络编码速度与PyPy相同,python,performance,julia,Python,Performance,Julia,我用Python编写了一些神经网络代码,我在Julia中重写了这些代码。纯Python代码运行时间约为7秒,而Julia和PyPy代码运行时间约为0.75秒 sigmoid(z::Float64) = 1/(1 + exp(-z)) sigmoidPrime(z::Float64) = sigmoid(z) * (1 - sigmoid(z)) ### Types ### abstract AbstractNode type Edge source::AbstractNode
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)
sigmoid(z::Float64)=1/(1+exp(-z))
sigmoidtime(z::Float64)=sigmoid(z)*(1-sigmoid(z))
###类型###
抽象节点
类型边缘
source::AbstractNode
目标::抽象节点
重量::浮动64
导数::浮点64
增广::布尔
Edge(source::AbstractNode,target::AbstractNode)=新建(source,target,randn(1,1)[1],0.0,false)
结束
键入Node这看起来更像是一个代码检查,而不是一个问题(没有任何问号),但无论如何我还是要尝试一下。唯一明显的潜在性能问题是,您正在通过评估
、计算
和反向传播
中的理解来分配数组。evaluate
中的加权和计算与for循环相比效率更高。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's来查看代码大部分时间都花在哪里——这可能会揭示一些不明显的热点,您可以进一步优化这些热点
关于与PyPy的比较,Julia和PyPy很可能都很好地处理了这段代码——达到或接近C性能——在这种情况下,您不会期望Julia比PyPy快很多,因为它们都接近最优。与C实现的性能进行比较将是非常有用的,因为它将显示Julia和PyPy在表中留下了多少性能。幸运的是,这段代码似乎很容易移植到C。这个问题可能更适合。理解在语法上非常好,在将来的构建中会优化它们吗?至于C,我学习Julia正是因为我不想学习C,:-)。重点不是你应该用C写所有的代码,而是如果你想知道为了基准测试,这段代码有多好,C是你应该比较的。如果PyPy在这段代码中的速度和C一样快,你就不能合理地期望Julia快10倍。是的,我理解,但仅仅为了测试这段代码而学习C是不值得的。理解非常好,但它们确实创建了新的数组对象——这就是语法的含义。将来可能会在obj.incomingEdges中为d写入sum(d.weight*evaluate(d.source,inputVector)
,而不分配任何数组。这是合理的。这将不得不成为一个谜。