今天要為cifar10訓練一個模型，它相較mnist來說，難度提高了不少，我們先從簡單的模型架構開始，看看我們有什麼能改善的。

from tensorflow.keras.datasets import cifar10
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
x_train = x_train / 255
x_test = x_test / 255

from tensorflow.keras.utils import to_categorical
y_train = to_categorical(y_train)
y_test = to_categorical(y_test)

from tensorflow.keras.models import Sequential
model = Sequential()

from tensorflow.keras.layers import Conv2D, Dense, Dropout, Flatten, MaxPool2D

model.add(
    Conv2D(
        filters = 64,
        input_shape = (32, 32, 3),
        kernel_size = (3, 3),
        strides = (1, 1),
        padding = 'same',
        activation = 'relu'
    )
)

model.add(
    MaxPool2D(
        pool_size = (2, 2)
    )
)

model.add(
    Conv2D(
        filters = 64,
        kernel_size = (3, 3),
        strides = (1, 1),
        padding = 'same',
        activation = 'relu'
    )
)

model.add(
    MaxPool2D(
        pool_size = (2, 2)
    )
)

model.add(
    Conv2D(
        filters = 64,
        kernel_size = (3, 3),
        strides = (1, 1),
        padding = 'same',
        activation = 'relu'
    )
)

model.add(
    MaxPool2D(
        pool_size = (2, 2)
    )
)

model.add(
    Flatten()
)

model.add(
    Dropout(
        rate = 0.2
    )
)

model.add(
    Dense(
        units = 10,
        activation = 'softmax'
    )
)

model.compile(
    optimizer = 'adam',
    loss = 'categorical_crossentropy',
    metrics = ['accuracy']
)

model.summary()

from tensorflow.keras.callbacks import ModelCheckpoint, CSVLogger, TerminateOnNaN, EarlyStopping

mcp = ModelCheckpoint(filepath='cifar10-{epoch:02d}.h5', monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto', save_freq='epoch')
log = CSVLogger(filename='cifar10.csv', separator=',', append=False)
ton = TerminateOnNaN()
esl = EarlyStopping(monitor='val_loss', patience=7, mode='auto', restore_best_weights=True)
esa = EarlyStopping(monitor='val_accuracy', patience=7, mode='auto', restore_best_weights=True)

from tensorflow.keras.preprocessing.image import ImageDataGenerator
datagen = ImageDataGenerator(
    width_shift_range = 0.1,
    height_shift_range = 0.1,
    shear_range = 0.1,
    rotation_range = 20,
    horizontal_flip = True
)

from sklearn.model_selection import train_test_split
import time
x_train, x_valid, y_train, y_valid = train_test_split(x_train, y_train, test_size=0.1, random_state=int(time.time()))

batch_size = 50
hist = model.fit(
    x = datagen.flow(x_train, y_train, batch_size=batch_size),
    steps_per_epoch = x_train.shape[0] // batch_size,
    epochs = 50,
    validation_data = (x_valid, y_valid),
    callbacks = [mcp, log, ton, esl, esa],
    verbose = 2
)

score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 32, 32, 64)        1792      
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 16, 16, 64)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 16, 16, 64)        36928     
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 8, 8, 64)          0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 8, 8, 64)          36928     
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 4, 4, 64)          0         
_________________________________________________________________
flatten (Flatten)            (None, 1024)              0         
_________________________________________________________________
dropout (Dropout)            (None, 1024)              0         
_________________________________________________________________
dense (Dense)                (None, 10)                10250     
=================================================================
Total params: 85,898
Trainable params: 85,898
Non-trainable params: 0
_________________________________________________________________

Test loss =  0.6636186135292054
Test accuracy =  0.7777

77.77%的正確率，看來我們還有很多進步的空間，比起mnist一開始的97%少了很多，這還是我們已經應用了前面所提到的技巧後的成績，我們畫些錯誤的圖出來看一下。

y_predict = model.predict(x_test)
import numpy as np
y_predict = np.argmax(y_predict, axis=1)
y_test = np.argmax(y_test, axis=1)

wrong = np.not_equal(y_predict, y_test)
label = np.arange(*y_test.shape)[wrong]

text = ['飛機', '汽車', '鳥' ,'貓', '鹿', '狗', '青蛙', '馬', '船', '卡車']

import matplotlib.pyplot as plt
from random import choice
plt.figure(figsize=(16,10),facecolor='w')
for i in range(5):
  for j in range(8):
    index = choice(label)
    plt.subplot(5, 8, i*8+j+1)
    plt.title("label: {}, predict: {}".format(text[y_test[index]], text[y_predict[index]]), fontproperties="Microsoft YaHei")
    plt.imshow(x_test[index])
    plt.axis('off')

plt.show()

大多都是把一種動物辨識成另一種動物，或是一種交通工具辨識成另一種交通工具，混淆兩者的情況沒有太多，模型還是有學到一些東西的，畫個混淆矩陣來看，因為要把標籤換成中文，所以改了一下程式。

import pandas as pd
import numpy as np

y_test = np.array(list(map(lambda x: text[x], y_test)))
y_predict = np.array(list(map(lambda x: text[x], y_predict)))
df = pd.DataFrame({'y_Actual': y_test, 'y_Predicted': y_predict})
pd.crosstab(df['y_Actual'], df['y_Predicted'], rownames=['Actual'], colnames=['Predicted'])

貓好像經常被誤認為狗，發生了206次，不過也不能怪模型了，這圖解析度也不高，有些圖片用人眼看也需要一些時間才能反應出來，這裡畫個訓練時的loss和accuracy給大家看吧，程式有所更改，因為我們將epoch調高了，也有可能發生EarlyStopping的情況，所以不能把數字寫死了，在本次訓練中，模型在第34次epoch停止，模型被回退到第27次時的狀態了，比起mnist的圖看起來，有種掙扎的感覺。

history = hist.history
epoch = len(history['loss'])
x = np.arange(epoch)
plt.figure(facecolor='w')
plt.plot(x, history['loss'], label='loss')
plt.plot(x, history['val_loss'], label='val_loss')
plt.plot(x, history['accuracy'], label='accuracy')
plt.plot(x, history['val_accuracy'], label='val_accuracy')

plt.xlim(0, epoch-1)
plt.xticks([i for i in range(0,epoch,epoch//5)],[str(i) for i in range(0,epoch,epoch//5)])
plt.xlabel('epoch')
plt.ylim(0,1)
plt.ylabel('acc-loss')
plt.legend()
plt.show()