本文我們將學習使用Keras一步一步搭建一個卷積神經網絡。具體來說，我們將使用卷積神經網絡對手寫數字(MNIST數據集)進行識別，并達到99%以上的正確率。

@為什么選擇Keras呢？

主要是因為簡單方便。更多細節請看：https://keras.io/

@什么卷積神經網絡？

簡單地說，卷積神經網絡(CNNs)是一種多層神經網絡，它可以有效地減少全連接神經網絡參數量太大的問題。

下面就直接進入主題吧！

import keras 
keras.__version__

‘2.1.5’

from keras.models import Sequential
# 序貫模型
model = Sequential() 
from keras.layers import Dense 
import numpy as np 
import tensorflow as tf

配置keras模型

# units 矩陣運算輸出的特征維度，input_dim 輸入數據特征維度
model.add(Dense(units=64, activation='relu', input_dim=100))
model.add(Dense(units=10, activation='softmax')) 
model.add(Dense(units = 1024,activation='tanh')) 
model.output_shape

(None, 10)

model.output_shape

(None, 1024)

在完成了模型的構建后, 可以使用 .compile() 來配置學習過程

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

訓練

不需要寫for循環

model.fit(x_train, y_train, epochs=5, batch_size=32)

一批批交給模型，需要自己寫for循環

model.train_on_batch(x_batch, y_batch)

模型評估

loss_and_metrics = model.evaluate(x_test, y_test, batch_size=128)

模型預測

classes = model.predict(x_test, batch_size=128)

---

---

---

實例 mnist 手寫數字進行識別

（外網下載數據可能很慢或者timeouts）

導包、定義變量

import keras
# 數據集
from keras.datasets import mnist
# 序貫模型
from keras.models import Sequential
# Dense：矩陣運算
# Dropout：防止過擬合
# Flatten：reshape(None,-1)
from keras.layers import Dense, Dropout, Flatten
# Conv2D：卷積運算
# MaxPooling2D：池化
from keras.layers import Conv2D, MaxPooling2D# 后端，后臺
# 默認Tensorflow
from keras import backend as Kbatch_size = 128
num_classes = 10
epochs = 12# input image dimensions
img_rows, img_cols = 28, 28

數據操作，轉換

import tensorflow as tf 
# the data, split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()# 四維的NHWC---->卷積運算需要
x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
input_shape = (img_rows, img_cols, 1)# 類型轉換
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')# 歸一化 0 ~1
x_train /= 255
x_test /= 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')print(y_train.shape)
# convert class vectors to binary class matrices
# one-hot
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

模型構建

# 聲明序貫模型
model = Sequential()
# 第一層卷積
model.add(Conv2D(32, kernel_size=(3, 3),activation='relu',input_shape=input_shape))
# 第二層卷積
model.add(Conv2D(64, (3, 3), activation='relu'))
# 池化層
model.add(MaxPooling2D(pool_size=(2, 2)))
# dropout層
model.add(Dropout(0.25))
# reshape
model.add(Flatten())
# 全連接層，矩陣運算
model.add(Dense(1024, activation='relu'))
# dropout層
model.add(Dropout(0.5))
# 輸出層
model.add(Dense(num_classes, activation='softmax'))

編譯，最優化

model.compile(loss=keras.losses.categorical_crossentropy,optimizer=keras.optimizers.Adadelta(),metrics=['accuracy'])

訓練

x_train.shape

(60000, 28, 28, 1)

model.fit(x_train, y_train,batch_size=batch_size,epochs=epochs,verbose=1,validation_data=(x_test, y_test))

Train on 60000 samples, validate on 10000 samples
Epoch 1/12
26752/60000 [============>…] - ETA: 1:38 - loss: 0.3388 - acc: 0.8965

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

參考：
https://elitedatascience.com/keras-tutorial-deep-learning-in-python
http://adventuresinmachinelearning.com/keras-tutorial-cnn-11-lines/
https://keras.io/zh/#30-keras