Swift for TensorFlow (2) 之训练和识别 MNIST

引入 Python：

Swift For TensorFlow supports Python interoperability.

You can import Python modules from Swift, call Python functions, and convert values between Swift and Python.

import PythonKit
print(Python.version)

MNIST 数据集识别

本文主要学习如何利用 S4TF 训练和识别 MNIST。

读取数据集

在 Swift 版本里：

先引入组件：

import TensorFlow
import PythonKit
import Foundation

let os = Python.import("os")
let np = Python.import("numpy")
let metrics = Python.import("sklearn.metrics")

在之前的文章使用 MNIST 集入门 Tensoflow 中 Python 版本读数据集：

(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data(
    path = os.path.join(os.getcwd(), "data/mnist.npz")
)

在 Swift 版本调入数据集：

if !Bool(os.path.exists("mnist.npz"))! {
    os.popen("wget https://s3.amazonaws.com/img-datasets/mnist.npz").read()
}

let mnist = np.load("mnist.npz")

Python 数据初始化：

x_train = x_train.reshape(x_train.shape[0], 28, 28, 1)
x_test = x_test.reshape(x_test.shape[0], 28, 28, 1)
input_shape = (28, 28, 1)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')

x_train /= 255
x_test /= 255

print('x_train shape: ', x_train.shape)
print('Number of images in x_train', x_train.shape[0])
print('Number of images in x_test', x_test.shape[0])

Swift 中：

let xTrain = Tensor<Float>(numpy: mnist["x_train"].astype("float32").reshape([60000, 28, 28, 1]) / 255.0)!
let xTest = Tensor<Float>(numpy: mnist["x_test"].astype("float32").reshape([10000, 28, 28, 1]) / 255.0)!

let yTrain = Tensor<Int32>(numpy: mnist["y_train"].astype("int32"))!
let yTest = Tensor<Int32>(numpy: mnist["y_test"].astype("int32"))!

模型

在 python 中：

model = Sequential()
model.add(Conv2D(28, kernel_size=(3,3), input_shape=input_shape))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten()) # Flattening the 2D arrays for fully connected layers
model.add(Dense(128, activation=tf.nn.relu))
model.add(Dropout(0.2))
model.add(Dense(10,activation=tf.nn.softmax))

在 Swift 中，我们先简单的用其中几个人 Layers：

var model = Sequential {
    Flatten<Float>()
    Dense<Float>(inputSize: 784, outputSize: 100, activation: relu)
    Dense<Float>(inputSize: 100, outputSize: 10, activation: softmax)
}

编译和优化模型

在 Python 中：

model.compile(optimizer='adam', 
              loss='sparse_categorical_crossentropy', 
              metrics=['accuracy'])
model.fit(x=x_train,y=y_train, epochs=10)

model.evaluate(x_test, y_test)

在 Swift 中：

let optimizer = Adam(for: model, learningRate: 0.02)
Context.local.learningPhase = .training

for _ in 1...5 {
    let 𝛁model = gradient(at: model) { model -> Tensor<Float> in
        let ŷ = model(xTrain)
        let loss = softmaxCrossEntropy(logits: ŷ, labels: yTrain)
//        print("Loss: \(loss)")
        return loss
    }
    optimizer.update(&model, along: 𝛁model)
}

测试

在 Python 中：

image_index = 5555
plt.imshow(x_test[image_index].reshape(28, 28),cmap='Greys')
pred = model.predict(x_test[image_index].reshape(1, 28, 28, 1))
print(pred.argmax())

在 Swift 中我们看看正确率：

// 在测试数据集上进行检查，得到模型的准度
let logits = model(xTest)
let sparseLogits = logits.makeNumpyArray().argmax(axis: -1)

let sparseLabels = yTest.makeNumpyArray()

print(Float(metrics.accuracy_score(sparseLabels, sparseLogits))!)

运行：

swift -O mnist_1.swift

正确率在： 80%左右，基本跑通 Swift For Tensorflow 流程。

总结

在接下来的学习过程中，不断了解 Python 的代码，也可以在 Swift 版本中同步学习，加深了解 TensorFlow！