---

Load MNIST Data Set

載入60000筆訓練數據與10000筆測試數據。

(train_feature, train_label), (test_feature, test_label) = mnist.load_data()

Data Preprocessing

Reshape

將28x28特徵值Raw Data(圖片)轉換為32位元浮點數一維數據。

train_feature_vector = train_feature.reshape(len(train_feature), 784).astype('float32')
test_feature_vector = test_feature.reshape(len(test_feature), 784).astype('float32')

Feature Normalization

對特徵值進行正規化處理，也就是將數據按比例縮放至[0, 1]區間，且不改變其原始分佈，以收斂速度與預測精準度。

train_feature_normal = train_feature_vector / 255
test_feature_normal = test_feature_vector / 255

One-Hot Encoding

對離散型資料標籤進行獨熱編碼處理轉換為布林陣列，便於進行矩陣運算。

train_label_onehot = np_utils.to_categorical(train_label)
test_label_onehot = np_utils.to_categorical(test_label)

Model Definition

定義循序模型之結構、訓練方法、準確率評估

model = Sequential()

Layer Definition

定義輸入層、隱藏層、輸出層 :

• Units : 784 -> 256 -> 10
• 常態分佈亂數初始化weight＆bias
• 隱藏層活化函數使用ReLU
• 輸出層活化函數使用Softmax

model.add( Dense(units=256, input_dim=784, init='normal', activation='relu') )
model.add( Dense(units=10, init='normal', activation='softmax') )

Training Definition

定義訓練方法 :

• 損失函數為 CrossEntropy Loss
• 優化器使用 Adam
• 驗證數據分割比例為0.2(將6萬筆訓練數據進一步分割為4.8萬筆訓練數據和1.2萬筆驗證數據)
• 訓練週期(epoch)為10
• 每批次樣本數為200(因此一個訓練週期為4.8萬/200=240批次)

model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit(x=train_feature_normal, y=train_label_onehot, validation_split=0.2, epochs=10, batch_size=200, verbose=2)

Accuracy Evaluation

accuracy = model.evaluate(test_feature_normal, test_label_onehot)
print('\n[Accuracy] = ', accuracy[1])

Save & Load Model

# save
model.save("mdl_mlp_mnist.h5")
# load
model = load_model("mdl_mlp_mnist.h5")

Full Code

#!/usr/bin/env python3
import numpy as np
import matplotlib.pyplot as plt
from keras.utils import np_utils
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense
from keras.models import load_model
np.random.seed(1234)  # for reproducibility


def showPredict(imgs, lbls, predictions):
    plt.gcf().set_size_inches(10, 10)
    for i in range(0, 10):
        fig = plt.subplot(2, 5, i + 1)
        fig.imshow(imgs[i], cmap='binary')

        title = 'prediction = ' + str(predictions[i])
        if predictions[i] != lbls[i]:
            title += '(X)'

        title += '\nlabel = ' + str(lbls[i])
        fig.set_title(title, fontsize=10)
        fig.set_xticks([])
        fig.set_yticks([])
    
    plt.show()


def mdlTrain(train_feature, train_label, test_feature, test_label):
    # model definition
    model = Sequential()

    # input:784, hidden:256, output:10
    model.add( Dense(units=256, input_dim=784, init='normal', activation='relu') )
    model.add( Dense(units=10, init='normal', activation='softmax') )

    # training definition
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
    model.fit(x=train_feature, y=train_label, validation_split=0.2, epochs=10, batch_size=200, verbose=2)

    # accuracy evaluation
    accuracy = model.evaluate(test_feature, test_label)
    print('\n[Accuracy] = ', accuracy[1])

    return model


# load mnist data
(train_feature, train_label), (test_feature, test_label) = mnist.load_data()

# data preprocessing
# reshape
train_feature_vector = train_feature.reshape(len(train_feature), 784).astype('float32')
test_feature_vector = test_feature.reshape(len(test_feature), 784).astype('float32')

# feature normalization 
train_feature_normal = train_feature_vector / 255
test_feature_normal = test_feature_vector / 255

# one-hot encoding
train_label_onehot = np_utils.to_categorical(train_label)
test_label_onehot = np_utils.to_categorical(test_label)

action = input("1: Model Testing\n2: Model Training\n")
if action == "1":
    print("Load mdl_mlp_mnist.h5")
    model = load_model("mdl_mlp_mnist.h5")
    prediction = model.predict_classes(test_feature_normal)
    showPredict(test_feature, test_label, prediction)
    del model
else:
    print("===== Start training =====")
    model = mdlTrain(train_feature_normal, train_label_onehot, test_feature_normal, test_label_onehot)
    model.save("mdl_mlp_mnist.h5")
    print("===== Model has been saved =====")
    prediction = model.predict_classes(test_feature_normal)
    showPredict(test_feature, test_label, prediction)
    del model

Test Your Own Handwritten Numbers Image

為了讓訓練好的模型預測看看資料集以外的圖片，我用FireAlpaca「手寫」了10張28x28的數字圖片😆，並將圖片命名為「真實數字_圖片順序編碼.jpg」這樣的格式，例如「8_image2.jpg」代表這張圖片為我製作的第2張圖片，內容為數字8，這樣的命名規則是為了方便讀取圖片時能從檔名擷取其label。

import blob & opencv

from glob import glob
from cv2 import cv2 as cv

P.S. 在VS Code中若只寫「import cv2」的話會報錯...

data preprocessing

def get_test_process(files):
    test_image = []
    test_label = []
    for file in files:
        label = int(file[0:1])  # get label from file name
        image = cv.imread(file, cv.IMREAD_GRAYSCALE)  # read image as grayscale
        # retval, dst = cv.threshold(src, thresh, maxval, type[,dst])
        image = cv.threshold(image, 120, 255, cv.THRESH_BINARY_INV)[1]  # binary invert
        test_image.append(image)
        test_label.append(label)

    # list -> numpy.array
    test_image = np.array(test_image)
    test_label = np.array(test_label)

    # reshape(flatten) & normalization
    test_image_normal = test_image.reshape(len(test_image), 784).astype('float32') / 255
    # one-hot encoding
    test_label_onehot = np_utils.to_categorical(test_label)

    return (test_image, test_label), (test_image_normal, test_label_onehot)

Prediction

model = load_model("mdl_mlp_mnist.h5")
print("=== Load mdl_mlp_mnist.h5 ===")
files = glob('*.jpg')  # find all images (path)

# data preprocessing
(test_image, test_label), (test_image_normal, test_label_onehot) = get_test_process(files)

prediction = model.predict_classes(test_image_normal)
showPredict(test_image, test_label, prediction)
del model

Result

哎呀，其中一張數字8的圖片預測錯誤😂

和數據集的圖片比較起來，我的手寫圖片經過影像處理完筆跡變得超細，或許特徵相對不那麼明顯吧，把原圖多點幾個像素上去再預測一次就過了呢。

---

＊測試程式指定隨機亂數種子是為了再現性

＊下載MNIST數據集時若發生 ssl.SSLError: [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed，在Terminal 執行以下命令 :

/Applications/Python\ 3.6/Install\ Certificates.command

📄Keras中文說明文件