モデルトレーニングのウォークスルー

import TensorFlow
import PythonKit

// This cell is here to display the plots in a Jupyter Notebook.
// Do not copy it into another environment.
%include "EnableIPythonDisplay.swift"
print(IPythonDisplay.shell.enable_matplotlib("inline"))

('inline', 'module://ipykernel.pylab.backend_inline')

let plt = Python.import("matplotlib.pyplot")

import Foundation
import FoundationNetworking
func download(from sourceString: String, to destinationString: String) {
    let source = URL(string: sourceString)!
    let destination = URL(fileURLWithPath: destinationString)
    let data = try! Data.init(contentsOf: source)
    try! data.write(to: destination)
}

let trainDataFilename = "iris_training.csv"
download(from: "http://download.tensorflow.org/data/iris_training.csv", to: trainDataFilename)

let f = Python.open(trainDataFilename)
for _ in 0..<5 {
    print(Python.next(f).strip())
}
print(f.close())

120,4,setosa,versicolor,virginica
6.4,2.8,5.6,2.2,2
5.0,2.3,3.3,1.0,1
4.9,2.5,4.5,1.7,2
4.9,3.1,1.5,0.1,0
None

let featureNames = ["sepal_length", "sepal_width", "petal_length", "petal_width"]
let labelName = "species"
let columnNames = featureNames + [labelName]

print("Features: \(featureNames)")
print("Label: \(labelName)")

Features: ["sepal_length", "sepal_width", "petal_length", "petal_width"]
Label: species

let classNames = ["Iris setosa", "Iris versicolor", "Iris virginica"]

let batchSize = 32

/// A batch of examples from the iris dataset.
struct IrisBatch {
    /// [batchSize, featureCount] tensor of features.
    let features: Tensor<Float>

    /// [batchSize] tensor of labels.
    let labels: Tensor<Int32>
}

/// Conform `IrisBatch` to `Collatable` so that we can load it into a `TrainingEpoch`.
extension IrisBatch: Collatable {
    public init<BatchSamples: Collection>(collating samples: BatchSamples)
        where BatchSamples.Element == Self {
        /// `IrisBatch`es are collated by stacking their feature and label tensors
        /// along the batch axis to produce a single feature and label tensor
        features = Tensor<Float>(stacking: samples.map{$0.features})
        labels = Tensor<Int32>(stacking: samples.map{$0.labels})
    }
}

/// Initialize an `IrisBatch` dataset from a CSV file.
func loadIrisDatasetFromCSV(
        contentsOf: String, hasHeader: Bool, featureColumns: [Int], labelColumns: [Int]) -> [IrisBatch] {
        let np = Python.import("numpy")

        let featuresNp = np.loadtxt(
            contentsOf,
            delimiter: ",",
            skiprows: hasHeader ? 1 : 0,
            usecols: featureColumns,
            dtype: Float.numpyScalarTypes.first!)
        guard let featuresTensor = Tensor<Float>(numpy: featuresNp) else {
            // This should never happen, because we construct featuresNp in such a
            // way that it should be convertible to tensor.
            fatalError("np.loadtxt result can't be converted to Tensor")
        }

        let labelsNp = np.loadtxt(
            contentsOf,
            delimiter: ",",
            skiprows: hasHeader ? 1 : 0,
            usecols: labelColumns,
            dtype: Int32.numpyScalarTypes.first!)
        guard let labelsTensor = Tensor<Int32>(numpy: labelsNp) else {
            // This should never happen, because we construct labelsNp in such a
            // way that it should be convertible to tensor.
            fatalError("np.loadtxt result can't be converted to Tensor")
        }

        return zip(featuresTensor.unstacked(), labelsTensor.unstacked()).map{IrisBatch(features: $0.0, labels: $0.1)}

    }

let trainingDataset: [IrisBatch] = loadIrisDatasetFromCSV(contentsOf: trainDataFilename, 
                                                  hasHeader: true, 
                                                  featureColumns: [0, 1, 2, 3], 
                                                  labelColumns: [4])

let trainingEpochs: TrainingEpochs = TrainingEpochs(samples: trainingDataset, batchSize: batchSize)

let firstTrainEpoch = trainingEpochs.next()!
let firstTrainBatch = firstTrainEpoch.first!.collated
let firstTrainFeatures = firstTrainBatch.features
let firstTrainLabels = firstTrainBatch.labels

print("First batch of features: \(firstTrainFeatures)")
print("firstTrainFeatures.shape: \(firstTrainFeatures.shape)")
print("First batch of labels: \(firstTrainLabels)")
print("firstTrainLabels.shape: \(firstTrainLabels.shape)")

First batch of features: [[5.1, 2.5, 3.0, 1.1],
 [6.4, 3.2, 4.5, 1.5],
 [4.9, 3.1, 1.5, 0.1],
 [5.0, 2.0, 3.5, 1.0],
 [6.3, 2.5, 5.0, 1.9],
 [6.7, 3.1, 5.6, 2.4],
 [4.9, 3.1, 1.5, 0.1],
 [7.7, 2.8, 6.7, 2.0],
 [6.7, 3.0, 5.0, 1.7],
 [7.2, 3.6, 6.1, 2.5],
 [4.8, 3.0, 1.4, 0.1],
 [5.2, 3.4, 1.4, 0.2],
 [5.0, 3.5, 1.3, 0.3],
 [4.9, 3.1, 1.5, 0.1],
 [5.0, 3.5, 1.6, 0.6],
 [6.7, 3.3, 5.7, 2.1],
 [7.7, 3.8, 6.7, 2.2],
 [6.2, 3.4, 5.4, 2.3],
 [4.8, 3.4, 1.6, 0.2],
 [6.0, 2.9, 4.5, 1.5],
 [5.0, 3.0, 1.6, 0.2],
 [6.3, 3.4, 5.6, 2.4],
 [5.1, 3.8, 1.9, 0.4],
 [4.8, 3.1, 1.6, 0.2],
 [7.6, 3.0, 6.6, 2.1],
 [5.7, 3.0, 4.2, 1.2],
 [6.3, 3.3, 6.0, 2.5],
 [5.6, 2.5, 3.9, 1.1],
 [5.0, 3.4, 1.6, 0.4],
 [6.1, 3.0, 4.9, 1.8],
 [5.0, 3.3, 1.4, 0.2],
 [6.3, 3.3, 4.7, 1.6]]
firstTrainFeatures.shape: [32, 4]
First batch of labels: [1, 1, 0, 1, 2, 2, 0, 2, 1, 2, 0, 0, 0, 0, 0, 2, 2, 2, 0, 1, 0, 2, 0, 0, 2, 1, 2, 1, 0, 2, 0, 1]
firstTrainLabels.shape: [32]

let firstTrainFeaturesTransposed = firstTrainFeatures.transposed()
let petalLengths = firstTrainFeaturesTransposed[2].scalars
let sepalLengths = firstTrainFeaturesTransposed[0].scalars

plt.scatter(petalLengths, sepalLengths, c: firstTrainLabels.array.scalars)
plt.xlabel("Petal length")
plt.ylabel("Sepal length")
plt.show()

Use `print()` to show values.

import TensorFlow

let hiddenSize: Int = 10
struct IrisModel: Layer {
    var layer1 = Dense<Float>(inputSize: 4, outputSize: hiddenSize, activation: relu)
    var layer2 = Dense<Float>(inputSize: hiddenSize, outputSize: hiddenSize, activation: relu)
    var layer3 = Dense<Float>(inputSize: hiddenSize, outputSize: 3)

    @differentiable
    func callAsFunction(_ input: Tensor<Float>) -> Tensor<Float> {
        return input.sequenced(through: layer1, layer2, layer3)
    }
}

var model = IrisModel()

// Apply the model to a batch of features.
let firstTrainPredictions = model(firstTrainFeatures)
print(firstTrainPredictions[0..<5])

[[  1.1514063,  -0.7520321,  -0.6730235],
 [  1.4915676,  -0.9158071,  -0.9957161],
 [  1.0549936,  -0.7799266,   -0.410466],
 [  1.1725322, -0.69009197,  -0.8345413],
 [  1.4870572,  -0.8644099,  -1.0958937]]

print(softmax(firstTrainPredictions[0..<5]))

[[  0.7631462,  0.11375094, 0.123102814],
 [  0.8523791, 0.076757915,  0.07086295],
 [  0.7191151,  0.11478964,  0.16609532],
 [ 0.77540654,  0.12039323,  0.10420021],
 [  0.8541314,  0.08133837, 0.064530246]]

print("Prediction: \(firstTrainPredictions.argmax(squeezingAxis: 1))")
print("    Labels: \(firstTrainLabels)")

Prediction: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
    Labels: [1, 1, 0, 1, 2, 2, 0, 2, 1, 2, 0, 0, 0, 0, 0, 2, 2, 2, 0, 1, 0, 2, 0, 0, 2, 1, 2, 1, 0, 2, 0, 1]

let untrainedLogits = model(firstTrainFeatures)
let untrainedLoss = softmaxCrossEntropy(logits: untrainedLogits, labels: firstTrainLabels)
print("Loss test: \(untrainedLoss)")

Loss test: 1.7598655

let optimizer = SGD(for: model, learningRate: 0.01)

let (loss, grads) = valueWithGradient(at: model) { model -> Tensor<Float> in
    let logits = model(firstTrainFeatures)
    return softmaxCrossEntropy(logits: logits, labels: firstTrainLabels)
}
print("Current loss: \(loss)")

Current loss: 1.7598655

optimizer.update(&model, along: grads)

let logitsAfterOneStep = model(firstTrainFeatures)
let lossAfterOneStep = softmaxCrossEntropy(logits: logitsAfterOneStep, labels: firstTrainLabels)
print("Next loss: \(lossAfterOneStep)")

Next loss: 1.5318773

let epochCount = 500
var trainAccuracyResults: [Float] = []
var trainLossResults: [Float] = []

func accuracy(predictions: Tensor<Int32>, truths: Tensor<Int32>) -> Float {
    return Tensor<Float>(predictions .== truths).mean().scalarized()
}

for (epochIndex, epoch) in trainingEpochs.prefix(epochCount).enumerated() {
    var epochLoss: Float = 0
    var epochAccuracy: Float = 0
    var batchCount: Int = 0
    for batchSamples in epoch {
        let batch = batchSamples.collated
        let (loss, grad) = valueWithGradient(at: model) { (model: IrisModel) -> Tensor<Float> in
            let logits = model(batch.features)
            return softmaxCrossEntropy(logits: logits, labels: batch.labels)
        }
        optimizer.update(&model, along: grad)

        let logits = model(batch.features)
        epochAccuracy += accuracy(predictions: logits.argmax(squeezingAxis: 1), truths: batch.labels)
        epochLoss += loss.scalarized()
        batchCount += 1
    }
    epochAccuracy /= Float(batchCount)
    epochLoss /= Float(batchCount)
    trainAccuracyResults.append(epochAccuracy)
    trainLossResults.append(epochLoss)
    if epochIndex % 50 == 0 {
        print("Epoch \(epochIndex): Loss: \(epochLoss), Accuracy: \(epochAccuracy)")
    }
}

Epoch 0: Loss: 1.475254, Accuracy: 0.34375
Epoch 50: Loss: 0.91668004, Accuracy: 0.6458333
Epoch 100: Loss: 0.68662673, Accuracy: 0.6979167
Epoch 150: Loss: 0.540665, Accuracy: 0.6979167
Epoch 200: Loss: 0.46283028, Accuracy: 0.6979167
Epoch 250: Loss: 0.4134724, Accuracy: 0.8229167
Epoch 300: Loss: 0.35054502, Accuracy: 0.8958333
Epoch 350: Loss: 0.2731444, Accuracy: 0.9375
Epoch 400: Loss: 0.23622067, Accuracy: 0.96875
Epoch 450: Loss: 0.18956228, Accuracy: 0.96875

plt.figure(figsize: [12, 8])

let accuracyAxes = plt.subplot(2, 1, 1)
accuracyAxes.set_ylabel("Accuracy")
accuracyAxes.plot(trainAccuracyResults)

let lossAxes = plt.subplot(2, 1, 2)
lossAxes.set_ylabel("Loss")
lossAxes.set_xlabel("Epoch")
lossAxes.plot(trainLossResults)

plt.show()

Use `print()` to show values.

let testDataFilename = "iris_test.csv"
download(from: "http://download.tensorflow.org/data/iris_test.csv", to: testDataFilename)

let testDataset = loadIrisDatasetFromCSV(
    contentsOf: testDataFilename, hasHeader: true,
    featureColumns: [0, 1, 2, 3], labelColumns: [4]).inBatches(of: batchSize)

// NOTE: Only a single batch will run in the loop since the batchSize we're using is larger than the test set size
for batchSamples in testDataset {
    let batch = batchSamples.collated
    let logits = model(batch.features)
    let predictions = logits.argmax(squeezingAxis: 1)
    print("Test batch accuracy: \(accuracy(predictions: predictions, truths: batch.labels))")
}

Test batch accuracy: 0.96666664

let firstTestBatch = testDataset.first!.collated
let firstTestBatchLogits = model(firstTestBatch.features)
let firstTestBatchPredictions = firstTestBatchLogits.argmax(squeezingAxis: 1)

print(firstTestBatchPredictions)
print(firstTestBatch.labels)

[1, 2, 0, 1, 1, 1, 0, 2, 1, 2, 2, 0, 2, 1, 1, 0, 1, 0, 0, 2, 0, 1, 2, 2, 1, 1, 0, 1, 2, 1]
[1, 2, 0, 1, 1, 1, 0, 2, 1, 2, 2, 0, 2, 1, 1, 0, 1, 0, 0, 2, 0, 1, 2, 1, 1, 1, 0, 1, 2, 1]

let unlabeledDataset: Tensor<Float> =
    [[5.1, 3.3, 1.7, 0.5],
     [5.9, 3.0, 4.2, 1.5],
     [6.9, 3.1, 5.4, 2.1]]

let unlabeledDatasetPredictions = model(unlabeledDataset)

for i in 0..<unlabeledDatasetPredictions.shape[0] {
    let logits = unlabeledDatasetPredictions[i]
    let classIdx = logits.argmax().scalar!
    print("Example \(i) prediction: \(classNames[Int(classIdx)]) (\(softmax(logits)))")
}

Example 0 prediction: Iris setosa ([   0.98731947,   0.012679046, 1.4035809e-06])
Example 1 prediction: Iris versicolor ([0.005065103,  0.85957265,  0.13536224])
Example 2 prediction: Iris virginica ([2.9613977e-05,     0.2637373,    0.73623306])