Comment utiliser le langage Go pour le développement du machine learning ?

go get gorgonia.org/gorgonia

package main

import "gorgonia.org/gorgonia"

func main() {
    gorgonia.NewGraph()
}

package main

import (
    "fmt"
    "gorgonia.org/datasets/mnist"
    "gorgonia.org/gorgonia"
)

func main() {
    // 下载和解压缩 mnist 数据集
    trainData, testData, err := mnist.Load(root)
    if err != nil {
        panic(err)
    }

    // 打印训练和测试数据及标签的形状
    fmt.Printf("train data shape: %v
", trainData.X.Shape())
    fmt.Printf("train labels shape: %v
", trainData.Y.Shape())
    fmt.Printf("test data shape: %v
", testData.X.Shape()) 
    fmt.Printf("test labels shape: %v
", testData.Y.Shape())
}

train data shape: (60000, 28, 28, 1)
train labels shape: (60000, 10)
test data shape: (10000, 28, 28, 1)
test labels shape: (10000, 10)

dataShape := trainData.X.Shape()
dataSize := dataShape[0]
inputSize := dataShape[1] * dataShape[2] * dataShape[3]
outputSize := testData.Y.Shape()[1]

// 构建神经网络
g := gorgonia.NewGraph()
x := gorgonia.NewTensor(g, tensor.Float32, 4, gorgonia.WithShape(dataSize, dataShape[1], dataShape[2], dataShape[3]), gorgonia.WithName("x"))
y := gorgonia.NewMatrix(g, tensor.Float32, gorgonia.WithShape(dataSize, outputSize), gorgonia.WithName("y"))

hiddenSize := 128
hidden1 := gorgonia.Must(gorgonia.NodeFromAny(g, tensor.Zero(tensor.Float32, hiddenSize), gorgonia.WithName("hidden1")))
hidden2 := gorgonia.Must(gorgonia.NodeFromAny(g, tensor.Zero(tensor.Float32, hiddenSize), gorgonia.WithName("hidden2")))

w1 := gorgonia.NewMatrix(g, tensor.Float32, gorgonia.WithShape(inputSize, hiddenSize), gorgonia.WithName("w1"))
w2 := gorgonia.NewMatrix(g, tensor.Float32, gorgonia.WithShape(hiddenSize, hiddenSize), gorgonia.WithName("w2"))
w3 := gorgonia.NewMatrix(g, tensor.Float32, gorgonia.WithShape(hiddenSize, outputSize), gorgonia.WithName("w3"))

b1 := gorgonia.NewVector(g, tensor.Float32, gorgonia.WithShape(hiddenSize), gorgonia.WithName("b1"))
b2 := gorgonia.NewVector(g, tensor.Float32, gorgonia.WithShape(hiddenSize), gorgonia.WithName("b2"))
b3 := gorgonia.NewVector(g, tensor.Float32, gorgonia.WithShape(outputSize), gorgonia.WithName("b3"))

hidden1Dot, err1 := gorgonia.Mul(x, w1)
hidden1Add, err2 := gorgonia.BroadcastAdd(hidden1Dot, b1, []byte{0})
hidden1Activate := gorgonia.Must(gorgonia.Rectify(hidden1Add))

hidden2Dot, err3 := gorgonia.Mul(hidden1Activate, w2)
hidden2Add, err4 := gorgonia.BroadcastAdd(hidden2Dot, b2, []byte{0})
hidden2Activate := gorgonia.Must(gorgonia.Rectify(hidden2Add))

yDot, err5 := gorgonia.Mul(hidden2Activate, w3)
yAdd, err6 := gorgonia.BroadcastAdd(yDot, b3, []byte{0})
ySoftMax := gorgonia.Must(gorgonia.SoftMax(yAdd))

iterations := 10
batchSize := 32
learningRate := 0.01

// 定义代价函数（交叉熵）
cost := gorgonia.Must(gorgonia.Mean(gorgonia.Must(gorgonia.Neg(gorgonia.Must(gorgonia.HadamardProd(y, gorgonia.Must(gorgonia.Log(ySoftMax)))))))

// 定义优化器
optimizer := gorgonia.NewVanillaSolver(g, gorgonia.WithLearnRate(learningRate))

// 表示模型将进行训练
vm := gorgonia.NewTapeMachine(g)

// 进行训练
for i := 0; i < iterations; i++ {
    fmt.Printf("Epoch %d
", i+1)

    for j := 0; j < dataSize; j += batchSize {
        upperBound := j + batchSize
        if upperBound > dataSize {
            upperBound = dataSize
        }
        xBatch := trainData.X.Slice(s{j, upperBound})
        yBatch := trainData.Y.Slice(s{j, upperBound})

        if err := gorgonia.Let(x, xBatch); err != nil {
            panic(err)
        }
        if err := gorgonia.Let(y, yBatch); err != nil {
            panic(err)
        }

        if err := vm.RunAll(); err != nil {
            panic(err)
        }

        if err := optimizer.Step(gorgonia.NodesToValueGrads(w1, b1, w2, b2, w3, b3)); err != nil {
            panic(err)
        }
    }

    // 测试准确率
    xTest := testData.X
    yTest := testData.Y

    if err := gorgonia.Let(x, xTest); err != nil {
        panic(err)
    }
    if err := gorgonia.Let(y, yTest); err != nil {
        panic(err)
    }

    if err := vm.RunAll(); err != nil {
        panic(err)
    }

    predict := gorgonia.Must(gorgonia.Argmax(ySoftMax, 1))
    label := gorgonia.Must(gorgonia.Argmax(yTest, 1))

    correct := 0
    for i := range label.Data().([]float32) {
        if predict.Data().([]float32)[i] == label.Data().([]float32)[i] {
            correct++
        }
    }

    fmt.Printf("Accuracy: %v
", float32(correct)/float32(len(label.Data().([]float32))))
}