1. 說明
本篇文章是CNN的另外一個(gè)例子，cifar10，是彩色的十分類數(shù)據(jù)集。
可以搭建卷積神經(jīng)網(wǎng)絡(luò)來訓(xùn)練模型。
2. cifar10實(shí)戰(zhàn)
2.1 導(dǎo)入相關(guān)庫
以下第三方庫是python專門用于深度學(xué)習(xí)的庫
# 導(dǎo)入tensorflow
import tensorflow as tf
# 導(dǎo)入keras
from tensorflow import keras
from keras.datasets import cifar10
# 引入繪制acc和loss曲線的庫
import matplotlib.pyplot as plt
# 引入ANN的必要的類
from keras.layers import Dense, Conv2D, MaxPool2D, Flatten
from keras.models import Sequential
from keras import optimizers, losses
from keras.preprocessing.image import ImageDataGenerator
2.2 加載數(shù)據(jù)
把cifar10數(shù)據(jù)集進(jìn)行加載
x_train是fashion_mnist訓(xùn)練集圖片,大小的32323的,y_train是對(duì)應(yīng)的標(biāo)簽是數(shù)字。
x_test是fashion_mnist測(cè)試集圖片,大小的32323的,y_test是對(duì)應(yīng)的標(biāo)簽是數(shù)字。
因?yàn)閥的形狀是(None,1)的,如果直接進(jìn)行獨(dú)熱編碼會(huì)變成(None,1,10),會(huì)和神經(jīng)網(wǎng)絡(luò)的輸出形狀(None,10)不匹配,因此需要對(duì)獨(dú)熱編碼之前的y的形狀進(jìn)行降維處理變成(None,)。
"1.加載數(shù)據(jù)"
"""
x_train是fashion_mnist訓(xùn)練集圖片,大小的32*32*3的,y_train是對(duì)應(yīng)的標(biāo)簽是數(shù)字
x_test是fashion_mnist測(cè)試集圖片,大小的32*32*3的,y_test是對(duì)應(yīng)的標(biāo)簽是數(shù)字
"""
(x_train, y_train), (x_test, y_test) = cifar10.load_data() # 加載cifar10數(shù)據(jù)集
print('mnist_data:', x_train.shape, y_train.shape, x_test.shape, y_test.shape) # 打印訓(xùn)練數(shù)據(jù)和測(cè)試數(shù)據(jù)的形狀
"""
因?yàn)閥的形狀是(None,1)的,如果直接進(jìn)行獨(dú)熱編碼會(huì)變成(None,1,10),會(huì)和神經(jīng)網(wǎng)絡(luò)的輸出形狀(None,10)不匹配,
因此需要對(duì)獨(dú)熱編碼之前的y的形狀進(jìn)行降維處理變成(None,)
"""
y_train = tf.squeeze(y_train, axis=1)
y_test = tf.squeeze(y_test, axis=1)
2.3 數(shù)據(jù)預(yù)處理
(1) 將輸入的圖片進(jìn)行歸一化，從0-255變換到0-1；
(2) 將標(biāo)簽y進(jìn)行獨(dú)熱編碼,因?yàn)樯窠?jīng)網(wǎng)絡(luò)的輸出是10個(gè)概率值,而y是1個(gè)數(shù)， 計(jì)算loss時(shí)無法對(duì)應(yīng)計(jì)算,因此將y進(jìn)行獨(dú)立編碼成為10個(gè)數(shù)的行向量，然后進(jìn)行l(wèi)oss的計(jì)算 獨(dú)熱編碼:例如數(shù)值1的10分類的獨(dú)熱編碼是[0 1 0 0 0 0 0 0 0 0,即1的位置為1，其余位置為0。
def preprocess(x, y): # 數(shù)據(jù)預(yù)處理函數(shù)
x = tf.cast(x, dtype=tf.float32) / 255. # 將輸入的圖片進(jìn)行歸一化，從0-255變換到0-1
y = tf.cast(y, dtype=tf.int32) # 將輸入圖片的標(biāo)簽轉(zhuǎn)換為int32類型
y = tf.one_hot(y, depth=10)
"""
# 將標(biāo)簽y進(jìn)行獨(dú)熱編碼,因?yàn)樯窠?jīng)網(wǎng)絡(luò)的輸出是10個(gè)概率值,而y是1個(gè)數(shù)，
計(jì)算loss時(shí)無法對(duì)應(yīng)計(jì)算,因此將y進(jìn)行獨(dú)立編碼成為10個(gè)數(shù)的行向量，然后進(jìn)行l(wèi)oss的計(jì)算
獨(dú)熱編碼:例如數(shù)值1的10分類的獨(dú)熱編碼是[0 1 0 0 0 0 0 0 0 0,即1的位置為1，其余位置為0
"""
return x, y
2.4 數(shù)據(jù)處理
數(shù)據(jù)加載進(jìn)入內(nèi)存后，需要轉(zhuǎn)換成 Dataset 對(duì)象，才能利用 TensorFlow 提供的各種便捷功能。
通過 Dataset.from_tensor_slices 可以將訓(xùn)練部分的數(shù)據(jù)圖片 x 和標(biāo)簽 y 都轉(zhuǎn)換成Dataset 對(duì)象
batchsz = 128 # 每次輸入給神經(jīng)網(wǎng)絡(luò)的圖片數(shù)
"""
數(shù)據(jù)加載進(jìn)入內(nèi)存后，需要轉(zhuǎn)換成 Dataset 對(duì)象，才能利用 TensorFlow 提供的各種便捷功能。
通過 Dataset.from_tensor_slices 可以將訓(xùn)練部分的數(shù)據(jù)圖片 x 和標(biāo)簽 y 都轉(zhuǎn)換成Dataset 對(duì)象
"""
db = tf.data.Dataset.from_tensor_slices((x_train, y_train)) # 構(gòu)建訓(xùn)練集對(duì)象
db = db.map(preprocess).shuffle(60000).batch(batchsz) # 將數(shù)據(jù)進(jìn)行預(yù)處理，隨機(jī)打散和批量處理
ds_val = tf.data.Dataset.from_tensor_slices((x_test, y_test)) # 構(gòu)建測(cè)試集對(duì)象
ds_val = ds_val.map(preprocess).batch(batchsz) # 將數(shù)據(jù)進(jìn)行預(yù)處理，隨機(jī)打散和批量處理
2.5 構(gòu)建網(wǎng)絡(luò)模型
構(gòu)建了兩層卷積層，兩層池化層，然后是展平層(將二維特征圖拉直輸入給全連接層)，然后是三層全連接層。
"3.構(gòu)建網(wǎng)絡(luò)模型"
model = Sequential([Conv2D(filters=6, kernel_size=(5, 5), activation='relu'),
MaxPool2D(pool_size=(2, 2), strides=2),
Conv2D(filters=16, kernel_size=(5, 5), activation='relu'),
MaxPool2D(pool_size=(2, 2), strides=2),
Flatten(),
Dense(120, activation='relu'),
Dense(84, activation='relu'),
Dense(10,activation='softmax')])
model.build(input_shape=(None, 32, 32, 3)) # 模型的輸入大小
model.summary() # 打印網(wǎng)絡(luò)結(jié)構(gòu)
2.6 模型編譯
模型的優(yōu)化器是Adam,另外一種優(yōu)化方法，學(xué)習(xí)率是0.01,
損失函數(shù)是losses.CategoricalCrossentropy,多分類交叉熵，
性能指標(biāo)是正確率accuracy。
"4.模型編譯"
model.compile(optimizer='Adam',
loss=losses.CategoricalCrossentropy(from_logits=False),
metrics=['accuracy']
)
"""
模型的優(yōu)化器是Adam
損失函數(shù)是losses.CategoricalCrossentropy,
性能指標(biāo)是正確率accuracy
"""
2.7 模型訓(xùn)練
模型訓(xùn)練的次數(shù)是20,每1次循環(huán)進(jìn)行測(cè)試
"5.模型訓(xùn)練"
history = model.fit(db, epochs=20, validation_data=ds_val, validation_freq=1)
"""
模型訓(xùn)練的次數(shù)是20,每1次循環(huán)進(jìn)行測(cè)試
"""
2.8 模型保存
以.h5文件格式保存模型
"6.模型保存"
model.save('cnn_cifar10.h5') # 以.h5文件格式保存模型
2.9 模型評(píng)價(jià)
得到測(cè)試集的正確率
"7.模型評(píng)價(jià)"
model.evaluate(ds_val) # 得到測(cè)試集的正確率
2.10 模型測(cè)試
對(duì)模型進(jìn)行測(cè)試
"8.模型測(cè)試"
sample = next(iter(ds_val)) # 取一個(gè)batchsz的測(cè)試集數(shù)據(jù)
x = sample[0] # 測(cè)試集數(shù)據(jù)
y = sample[1] # 測(cè)試集的標(biāo)簽
pred = model.predict(x) # 將一個(gè)batchsz的測(cè)試集數(shù)據(jù)輸入神經(jīng)網(wǎng)絡(luò)的結(jié)果
pred = tf.argmax(pred, axis=1) # 每個(gè)預(yù)測(cè)的結(jié)果的概率最大值的下標(biāo),也就是預(yù)測(cè)的數(shù)字
y = tf.argmax(y, axis=1) # 每個(gè)標(biāo)簽的最大值對(duì)應(yīng)的下標(biāo),也就是標(biāo)簽對(duì)應(yīng)的數(shù)字
print(pred) # 打印預(yù)測(cè)結(jié)果
print(y) # 打印標(biāo)簽類別
2.11 模型訓(xùn)練結(jié)果的可視化
對(duì)模型的訓(xùn)練結(jié)果進(jìn)行可視化
"9.模型訓(xùn)練時(shí)的可視化"
# 顯示訓(xùn)練集和驗(yàn)證集的acc和loss曲線
acc = history.history['accuracy'] # 獲取模型訓(xùn)練中的accuracy
val_acc = history.history['val_accuracy'] # 獲取模型訓(xùn)練中的val_accuracy
loss = history.history['loss'] # 獲取模型訓(xùn)練中的loss
val_loss = history.history['val_loss'] # 獲取模型訓(xùn)練中的val_loss
# 繪值acc曲線
plt.figure(1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.title('Training and Validation Accuracy')
plt.legend()
# 繪制loss曲線
plt.figure(2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.title('Training and Validation Loss')
plt.legend()
plt.show() # 將結(jié)果顯示出來
3. cifar10的CNN模型可視化結(jié)果圖
Epoch 1/20
391/391 [==============================] - 20s 44ms/step - loss: 1.7699 - accuracy: 0.3471 - val_loss: 1.5087 - val_accuracy: 0.4547
Epoch 2/20
391/391 [==============================] - 20s 49ms/step - loss: 1.4580 - accuracy: 0.4737 - val_loss: 1.4031 - val_accuracy: 0.4884
Epoch 3/20
391/391 [==============================] - 21s 51ms/step - loss: 1.3519 - accuracy: 0.5168 - val_loss: 1.3040 - val_accuracy: 0.5345
Epoch 4/20
391/391 [==============================] - 21s 51ms/step - loss: 1.2646 - accuracy: 0.5506 - val_loss: 1.2333 - val_accuracy: 0.5646
Epoch 5/20
391/391 [==============================] - 19s 47ms/step - loss: 1.2030 - accuracy: 0.5753 - val_loss: 1.2309 - val_accuracy: 0.5673
Epoch 6/20
391/391 [==============================] - 19s 46ms/step - loss: 1.1519 - accuracy: 0.5941 - val_loss: 1.1947 - val_accuracy: 0.5716
Epoch 7/20
391/391 [==============================] - 18s 44ms/step - loss: 1.1104 - accuracy: 0.6088 - val_loss: 1.1496 - val_accuracy: 0.5987
Epoch 8/20
391/391 [==============================] - 19s 46ms/step - loss: 1.0726 - accuracy: 0.6235 - val_loss: 1.1330 - val_accuracy: 0.6031
Epoch 9/20
391/391 [==============================] - 19s 47ms/step - loss: 1.0393 - accuracy: 0.6327 - val_loss: 1.1079 - val_accuracy: 0.6119
Epoch 10/20
391/391 [==============================] - 21s 51ms/step - loss: 1.0149 - accuracy: 0.6440 - val_loss: 1.1023 - val_accuracy: 0.6160
Epoch 11/20
391/391 [==============================] - 19s 47ms/step - loss: 0.9798 - accuracy: 0.6550 - val_loss: 1.0828 - val_accuracy: 0.6265
Epoch 12/20
391/391 [==============================] - 19s 47ms/step - loss: 0.9594 - accuracy: 0.6621 - val_loss: 1.0978 - val_accuracy: 0.6191
Epoch 13/20
391/391 [==============================] - 18s 44ms/step - loss: 0.9325 - accuracy: 0.6709 - val_loss: 1.0803 - val_accuracy: 0.6264
Epoch 14/20
391/391 [==============================] - 20s 49ms/step - loss: 0.9106 - accuracy: 0.6801 - val_loss: 1.0792 - val_accuracy: 0.6212
Epoch 15/20
391/391 [==============================] - 23s 54ms/step - loss: 0.8928 - accuracy: 0.6873 - val_loss: 1.0586 - val_accuracy: 0.6382
Epoch 16/20
391/391 [==============================] - 20s 50ms/step - loss: 0.8695 - accuracy: 0.6931 - val_loss: 1.0825 - val_accuracy: 0.6303
Epoch 17/20
391/391 [==============================] - 22s 54ms/step - loss: 0.8524 - accuracy: 0.6993 - val_loss: 1.0917 - val_accuracy: 0.6296
Epoch 18/20
391/391 [==============================] - 19s 46ms/step - loss: 0.8314 - accuracy: 0.7074 - val_loss: 1.0753 - val_accuracy: 0.6341
Epoch 19/20
391/391 [==============================] - 19s 48ms/step - loss: 0.8117 - accuracy: 0.7136 - val_loss: 1.0701 - val_accuracy: 0.6376
Epoch 20/20
391/391 [==============================] - 20s 50ms/step - loss: 0.7967 - accuracy: 0.7200 - val_loss: 1.0715 - val_accuracy: 0.6376
79/79 [==============================] - 1s 17ms/step - loss: 1.0715 - accuracy: 0.6376
4/4 [==============================] - 0s 6ms/step
從以上結(jié)果可知，模型的準(zhǔn)確率達(dá)到了63%，太低了，原因是網(wǎng)絡(luò)結(jié)構(gòu)有點(diǎn)簡單，因此進(jìn)行改變網(wǎng)絡(luò)結(jié)構(gòu)，同時(shí)為了降低過擬合的影響，加入數(shù)據(jù)增強(qiáng)部分。文章來源：http://www.zghlxwxcb.cn/news/detail-614102.html
4. 完整代碼
# 導(dǎo)入tensorflow
import tensorflow as tf
# 導(dǎo)入keras
from tensorflow import keras
from keras.datasets import cifar10
# 引入繪制acc和loss曲線的庫
import matplotlib.pyplot as plt
# 引入ANN的必要的類
from keras.layers import Dense, Conv2D, MaxPool2D, Flatten
from keras.models import Sequential
from keras import optimizers, losses
from keras.preprocessing.image import ImageDataGenerator
"1.加載數(shù)據(jù)"
"""
x_train是fashion_mnist訓(xùn)練集圖片,大小的32*32*3的,y_train是對(duì)應(yīng)的標(biāo)簽是數(shù)字
x_test是fashion_mnist測(cè)試集圖片,大小的32*32*3的,y_test是對(duì)應(yīng)的標(biāo)簽是數(shù)字
"""
(x_train, y_train), (x_test, y_test) = cifar10.load_data() # 加載cifar10數(shù)據(jù)集
print('mnist_data:', x_train.shape, y_train.shape, x_test.shape, y_test.shape) # 打印訓(xùn)練數(shù)據(jù)和測(cè)試數(shù)據(jù)的形狀
"""
因?yàn)閥的形狀是(None,1)的,如果直接進(jìn)行獨(dú)熱編碼會(huì)變成(None,1,10),會(huì)和神經(jīng)網(wǎng)絡(luò)的輸出形狀(None,10)不匹配,
因此需要對(duì)獨(dú)熱編碼之前的y的形狀進(jìn)行降維處理變成(None,)
"""
y_train = tf.squeeze(y_train, axis=1)
y_test = tf.squeeze(y_test, axis=1)
"2.數(shù)據(jù)預(yù)處理"
def preprocess(x, y): # 數(shù)據(jù)預(yù)處理函數(shù)
x = tf.cast(x, dtype=tf.float32) / 255. # 將輸入的圖片進(jìn)行歸一化，從0-255變換到0-1
y = tf.cast(y, dtype=tf.int32) # 將輸入圖片的標(biāo)簽轉(zhuǎn)換為int32類型
y = tf.one_hot(y, depth=10)
"""
# 將標(biāo)簽y進(jìn)行獨(dú)熱編碼,因?yàn)樯窠?jīng)網(wǎng)絡(luò)的輸出是10個(gè)概率值,而y是1個(gè)數(shù)，
計(jì)算loss時(shí)無法對(duì)應(yīng)計(jì)算,因此將y進(jìn)行獨(dú)立編碼成為10個(gè)數(shù)的行向量，然后進(jìn)行l(wèi)oss的計(jì)算
獨(dú)熱編碼:例如數(shù)值1的10分類的獨(dú)熱編碼是[0 1 0 0 0 0 0 0 0 0,即1的位置為1，其余位置為0
"""
return x, y
batchsz = 128 # 每次輸入給神經(jīng)網(wǎng)絡(luò)的圖片數(shù)
"""
數(shù)據(jù)加載進(jìn)入內(nèi)存后，需要轉(zhuǎn)換成 Dataset 對(duì)象，才能利用 TensorFlow 提供的各種便捷功能。
通過 Dataset.from_tensor_slices 可以將訓(xùn)練部分的數(shù)據(jù)圖片 x 和標(biāo)簽 y 都轉(zhuǎn)換成Dataset 對(duì)象
"""
db = tf.data.Dataset.from_tensor_slices((x_train, y_train)) # 構(gòu)建訓(xùn)練集對(duì)象
db = db.map(preprocess).shuffle(60000).batch(batchsz) # 將數(shù)據(jù)進(jìn)行預(yù)處理，隨機(jī)打散和批量處理
ds_val = tf.data.Dataset.from_tensor_slices((x_test, y_test)) # 構(gòu)建測(cè)試集對(duì)象
ds_val = ds_val.map(preprocess).batch(batchsz) # 將數(shù)據(jù)進(jìn)行預(yù)處理，隨機(jī)打散和批量處理
"3.構(gòu)建網(wǎng)絡(luò)模型"
model = Sequential([Conv2D(filters=6, kernel_size=(5, 5), activation='relu'),
MaxPool2D(pool_size=(2, 2), strides=2),
Conv2D(filters=16, kernel_size=(5, 5), activation='relu'),
MaxPool2D(pool_size=(2, 2), strides=2),
Flatten(),
Dense(120, activation='relu'),
Dense(84, activation='relu'),
Dense(10,activation='softmax')])
model.build(input_shape=(None, 32, 32, 3)) # 模型的輸入大小
model.summary() # 打印網(wǎng)絡(luò)結(jié)構(gòu)
"4.模型編譯"
model.compile(optimizer='Adam',
loss=losses.CategoricalCrossentropy(from_logits=False),
metrics=['accuracy']
)
"""
模型的優(yōu)化器是Adam
損失函數(shù)是losses.CategoricalCrossentropy,
性能指標(biāo)是正確率accuracy
"""
"5.模型訓(xùn)練"
history = model.fit(db, epochs=20, validation_data=ds_val, validation_freq=1)
"""
模型訓(xùn)練的次數(shù)是20,每1次循環(huán)進(jìn)行測(cè)試
"""
"6.模型保存"
model.save('cnn_cifar10.h5') # 以.h5文件格式保存模型
"7.模型評(píng)價(jià)"
model.evaluate(ds_val) # 得到測(cè)試集的正確率
"8.模型測(cè)試"
sample = next(iter(ds_val)) # 取一個(gè)batchsz的測(cè)試集數(shù)據(jù)
x = sample[0] # 測(cè)試集數(shù)據(jù)
y = sample[1] # 測(cè)試集的標(biāo)簽
pred = model.predict(x) # 將一個(gè)batchsz的測(cè)試集數(shù)據(jù)輸入神經(jīng)網(wǎng)絡(luò)的結(jié)果
pred = tf.argmax(pred, axis=1) # 每個(gè)預(yù)測(cè)的結(jié)果的概率最大值的下標(biāo),也就是預(yù)測(cè)的數(shù)字
y = tf.argmax(y, axis=1) # 每個(gè)標(biāo)簽的最大值對(duì)應(yīng)的下標(biāo),也就是標(biāo)簽對(duì)應(yīng)的數(shù)字
print(pred) # 打印預(yù)測(cè)結(jié)果
print(y) # 打印標(biāo)簽類別
"9.模型訓(xùn)練時(shí)的可視化"
# 顯示訓(xùn)練集和驗(yàn)證集的acc和loss曲線
acc = history.history['accuracy'] # 獲取模型訓(xùn)練中的accuracy
val_acc = history.history['val_accuracy'] # 獲取模型訓(xùn)練中的val_accuracy
loss = history.history['loss'] # 獲取模型訓(xùn)練中的loss
val_loss = history.history['val_loss'] # 獲取模型訓(xùn)練中的val_loss
# 繪值acc曲線
plt.figure(1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.title('Training and Validation Accuracy')
plt.legend()
# 繪制loss曲線
plt.figure(2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.title('Training and Validation Loss')
plt.legend()
plt.show() # 將結(jié)果顯示出來
5. 改進(jìn)后的代碼和結(jié)果
# 導(dǎo)入tensorflow
import tensorflow as tf
# 導(dǎo)入keras
from tensorflow import keras
from keras.datasets import cifar10
# 引入繪制acc和loss曲線的庫
import matplotlib.pyplot as plt
# 引入ANN的必要的類
from keras.layers import Dense, Conv2D, MaxPool2D, Flatten, Dropout
from keras.models import Sequential
from keras import optimizers, losses
from keras.preprocessing.image import ImageDataGenerator
"1.加載數(shù)據(jù)"
"""
x_train是fashion_mnist訓(xùn)練集圖片,大小的32*32*3的,y_train是對(duì)應(yīng)的標(biāo)簽是數(shù)字
x_test是fashion_mnist測(cè)試集圖片,大小的32*32*3的,y_test是對(duì)應(yīng)的標(biāo)簽是數(shù)字
"""
(x_train, y_train), (x_test, y_test) = cifar10.load_data() # 加載cifar10數(shù)據(jù)集
print('mnist_data:', x_train.shape, y_train.shape, x_test.shape, y_test.shape) # 打印訓(xùn)練數(shù)據(jù)和測(cè)試數(shù)據(jù)的形狀
image_gen_train = ImageDataGenerator(
rescale=1. / 1., # 如為圖像，分母為255時(shí)，可歸至0～1
rotation_range=45, # 隨機(jī)45度旋轉(zhuǎn)
width_shift_range=.15, # 寬度偏移
height_shift_range=.15, # 高度偏移
horizontal_flip=False, # 水平翻轉(zhuǎn)
zoom_range=0.5 # 將圖像隨機(jī)縮放閾量50％
)
image_gen_train.fit(x_train)
"""
因?yàn)閥的形狀是(None,1)的,如果直接進(jìn)行獨(dú)熱編碼會(huì)變成(None,1,10),會(huì)和神經(jīng)網(wǎng)絡(luò)的輸出形狀(None,10)不匹配,
因此需要對(duì)獨(dú)熱編碼后的y的形狀進(jìn)行降維處理變成(None,10)
"""
y_train = tf.squeeze(y_train, axis=1)
y_test = tf.squeeze(y_test, axis=1)
"2.數(shù)據(jù)預(yù)處理"
def preprocess(x, y): # 數(shù)據(jù)預(yù)處理函數(shù)
x = tf.cast(x, dtype=tf.float32) / 255. # 將輸入的圖片進(jìn)行歸一化，從0-255變換到0-1
y = tf.cast(y, dtype=tf.int32) # 將輸入圖片的標(biāo)簽轉(zhuǎn)換為int32類型
y = tf.one_hot(y, depth=10)
"""
# 將標(biāo)簽y進(jìn)行獨(dú)熱編碼,因?yàn)樯窠?jīng)網(wǎng)絡(luò)的輸出是10個(gè)概率值,而y是1個(gè)數(shù)，
計(jì)算loss時(shí)無法對(duì)應(yīng)計(jì)算,因此將y進(jìn)行獨(dú)立編碼成為10個(gè)數(shù)的行向量，然后進(jìn)行l(wèi)oss的計(jì)算
獨(dú)熱編碼:例如數(shù)值1的10分類的獨(dú)熱編碼是[0 1 0 0 0 0 0 0 0 0,即1的位置為1，其余位置為0
"""
return x, y
batchsz = 128 # 每次輸入給神經(jīng)網(wǎng)絡(luò)的圖片數(shù)
"""
數(shù)據(jù)加載進(jìn)入內(nèi)存后，需要轉(zhuǎn)換成 Dataset 對(duì)象，才能利用 TensorFlow 提供的各種便捷功能。
通過 Dataset.from_tensor_slices 可以將訓(xùn)練部分的數(shù)據(jù)圖片 x 和標(biāo)簽 y 都轉(zhuǎn)換成Dataset 對(duì)象
"""
db = tf.data.Dataset.from_tensor_slices((x_train, y_train)) # 構(gòu)建訓(xùn)練集對(duì)象
db = db.map(preprocess).shuffle(50000).batch(batchsz) # 將數(shù)據(jù)進(jìn)行預(yù)處理，隨機(jī)打散和批量處理
ds_val = tf.data.Dataset.from_tensor_slices((x_test, y_test)) # 構(gòu)建測(cè)試集對(duì)象
ds_val = ds_val.map(preprocess).batch(batchsz) # 將數(shù)據(jù)進(jìn)行預(yù)處理，隨機(jī)打散和批量處理
"3.構(gòu)建網(wǎng)絡(luò)模型"
model = Sequential([Conv2D(32, kernel_size=(3, 3), padding='same', activation='relu'),
Dropout(0.25),
MaxPool2D(pool_size=(2, 2), strides=2),
Conv2D(filters=64, kernel_size=(3, 3),padding='same', activation='relu'),
Dropout(0.25),
MaxPool2D(pool_size=(2, 2), strides=2),
Conv2D(filters=128, kernel_size=(3, 3), padding='same', activation='relu'),
Dropout(0.25),
MaxPool2D(pool_size=(2, 2), strides=2),
Flatten(),
Dense(1024, activation='relu'),
Dense(10,activation='softmax')])
model.build(input_shape=(None, 32, 32, 3)) # 模型的輸入大小
model.summary() # 打印網(wǎng)絡(luò)結(jié)構(gòu)
"4.模型編譯"
model.compile(optimizer='Adam',
loss=losses.CategoricalCrossentropy(from_logits=False),
metrics=['accuracy']
)
"""
模型的優(yōu)化器是Adam
損失函數(shù)是losses.CategoricalCrossentropy,
性能指標(biāo)是正確率accuracy
"""
"5.模型訓(xùn)練"
history = model.fit(db, epochs=20, validation_data=ds_val, validation_freq=1)
"""
模型訓(xùn)練的次數(shù)是10,每1次循環(huán)進(jìn)行測(cè)試
"""
"6.模型保存"
model.save('cnn_cifar10_4.h5') # 以.h5文件格式保存模型
"7.模型評(píng)價(jià)"
model.evaluate(ds_val) # 得到測(cè)試集的正確率
"8.模型測(cè)試"
sample = next(iter(ds_val)) # 取一個(gè)batchsz的測(cè)試集數(shù)據(jù)
x = sample[0] # 測(cè)試集數(shù)據(jù)
y = sample[1] # 測(cè)試集的標(biāo)簽
pred = model.predict(x) # 將一個(gè)batchsz的測(cè)試集數(shù)據(jù)輸入神經(jīng)網(wǎng)絡(luò)的結(jié)果
pred = tf.argmax(pred, axis=1) # 每個(gè)預(yù)測(cè)的結(jié)果的概率最大值的下標(biāo),也就是預(yù)測(cè)的數(shù)字
y = tf.argmax(y, axis=1) # 每個(gè)標(biāo)簽的最大值對(duì)應(yīng)的下標(biāo),也就是標(biāo)簽對(duì)應(yīng)的數(shù)字
print(pred) # 打印預(yù)測(cè)結(jié)果
print(y) # 打印標(biāo)簽類別
"9.模型訓(xùn)練時(shí)的可視化"
# 顯示訓(xùn)練集和驗(yàn)證集的acc和loss曲線
acc = history.history['accuracy'] # 獲取模型訓(xùn)練中的accuracy
val_acc = history.history['val_accuracy'] # 獲取模型訓(xùn)練中的val_accuracy
loss = history.history['loss'] # 獲取模型訓(xùn)練中的loss
val_loss = history.history['val_loss'] # 獲取模型訓(xùn)練中的val_loss
# 繪值acc曲線
plt.figure(1)
plt.plot(acc, label='Training Accuracy')
plt.plot(val_acc, label='Validation Accuracy')
plt.title('Training and Validation Accuracy')
plt.legend()
# 繪制loss曲線
plt.figure(2)
plt.plot(loss, label='Training Loss')
plt.plot(val_loss, label='Validation Loss')
plt.title('Training and Validation Loss')
plt.legend()
plt.show() # 將結(jié)果顯示出來
Epoch 1/20
391/391 [==============================] - 109s 272ms/step - loss: 1.4598 - accuracy: 0.4688 - val_loss: 1.2981 - val_accuracy: 0.5713
Epoch 2/20
391/391 [==============================] - 103s 262ms/step - loss: 1.0536 - accuracy: 0.6297 - val_loss: 1.0809 - val_accuracy: 0.6476
Epoch 3/20
391/391 [==============================] - 104s 264ms/step - loss: 0.8955 - accuracy: 0.6840 - val_loss: 0.9347 - val_accuracy: 0.7052
Epoch 4/20
391/391 [==============================] - 103s 262ms/step - loss: 0.7785 - accuracy: 0.7258 - val_loss: 0.9195 - val_accuracy: 0.6880
Epoch 5/20
391/391 [==============================] - 104s 264ms/step - loss: 0.6896 - accuracy: 0.7575 - val_loss: 0.8039 - val_accuracy: 0.7330
Epoch 6/20
391/391 [==============================] - 103s 262ms/step - loss: 0.6085 - accuracy: 0.7864 - val_loss: 0.8141 - val_accuracy: 0.7296
Epoch 7/20
391/391 [==============================] - 102s 261ms/step - loss: 0.5378 - accuracy: 0.8102 - val_loss: 0.7418 - val_accuracy: 0.7463
Epoch 8/20
391/391 [==============================] - 102s 259ms/step - loss: 0.4698 - accuracy: 0.8335 - val_loss: 0.6867 - val_accuracy: 0.7693
Epoch 9/20
391/391 [==============================] - 102s 259ms/step - loss: 0.4042 - accuracy: 0.8563 - val_loss: 0.6729 - val_accuracy: 0.7734
Epoch 10/20
391/391 [==============================] - 103s 261ms/step - loss: 0.3474 - accuracy: 0.8763 - val_loss: 0.6690 - val_accuracy: 0.7696
Epoch 11/20
391/391 [==============================] - 102s 261ms/step - loss: 0.2969 - accuracy: 0.8958 - val_loss: 0.6920 - val_accuracy: 0.7660
Epoch 12/20
391/391 [==============================] - 102s 260ms/step - loss: 0.2597 - accuracy: 0.9072 - val_loss: 0.7243 - val_accuracy: 0.7558
Epoch 13/20
391/391 [==============================] - 102s 259ms/step - loss: 0.2146 - accuracy: 0.9235 - val_loss: 0.7107 - val_accuracy: 0.7599
Epoch 14/20
391/391 [==============================] - 106s 269ms/step - loss: 0.1904 - accuracy: 0.9318 - val_loss: 0.6790 - val_accuracy: 0.7760
Epoch 15/20
391/391 [==============================] - 103s 262ms/step - loss: 0.1703 - accuracy: 0.9404 - val_loss: 0.7177 - val_accuracy: 0.7706
Epoch 16/20
391/391 [==============================] - 103s 262ms/step - loss: 0.1526 - accuracy: 0.9463 - val_loss: 0.7038 - val_accuracy: 0.7781
Epoch 17/20
391/391 [==============================] - 103s 262ms/step - loss: 0.1333 - accuracy: 0.9520 - val_loss: 0.7267 - val_accuracy: 0.7770
Epoch 18/20
391/391 [==============================] - 103s 263ms/step - loss: 0.1253 - accuracy: 0.9576 - val_loss: 0.7518 - val_accuracy: 0.7728
Epoch 19/20
391/391 [==============================] - 103s 263ms/step - loss: 0.1210 - accuracy: 0.9570 - val_loss: 0.7855 - val_accuracy: 0.7670
Epoch 20/20
391/391 [==============================] - 104s 265ms/step - loss: 0.1105 - accuracy: 0.9607 - val_loss: 0.8065 - val_accuracy: 0.7721
79/79 [==============================] - 4s 48ms/step - loss: 0.8065 - accuracy: 0.7721
4/4 [==============================] - 0s 15ms/step
結(jié)果有所改善，但是過擬合仍然存在。文章來源地址http://www.zghlxwxcb.cn/news/detail-614102.html
到了這里，關(guān)于python與深度學(xué)習(xí)(九):CNN和cifar10的文章就介紹完了。如果您還想了解更多內(nèi)容，請(qǐng)?jiān)谟疑辖撬阉鱐OY模板網(wǎng)以前的文章或繼續(xù)瀏覽下面的相關(guān)文章，希望大家以后多多支持TOY模板網(wǎng)！
