只要是程式設計師都知道，學習一門新的語言或者框架的時候，第一個自己敲入且執行的程式，都一個HelloWorld的程式、筆者也不例外，當筆者把Keras在Unbuntu系統安裝好之後，早已掩蓋不住激動的心，然後想在Keras上面執行一個HelloWorld的例子。後面筆者參考了莫愁同學的Keras的學習視訊，上面提到了一個用Keras的API去反推出一個線性函式的例子，感覺非常的有趣，所有就模擬敲入了一把，現在讓我給大家來逐一解釋一下，解釋不到位的地方，還請海涵，畢竟也是樓主第一次接觸Keras框架。所有的程式碼全部黏貼出來，檔名為helloworld.py,如下，具體解釋請看程式碼中的註釋。

# Import package
import numpy as np
np.random.seed(1337)
from keras.models import Sequential
from keras.layers import Dense
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt

# 1.Build the trainning data
X=np.linspace(-1,1,200)
np.random.shuffle(X)
np.random.normal()
Y=0.5*X+2+np.random.normal(0,0.05,(200,))
plt.scatter(X,Y)
plt.show()
X_train,Y_train=X[:160],Y[:160]
X_test,Y_test=X[160:],Y[160:]
print(X)
print("*******************************************")
print(Y)

# 2.Build a neural network from the 1st layer to the last layer
model=Sequential()
model.add(Dense(output_dim=1,input_dim=1))

#3. Choose loss function and optimzing method
model.compile(loss='mse',optimizer='sgd')

#4. Trainning
print("Training......")
for step in range(1400):
        cost=model.train_on_batch(X_train,Y_train)
        if step % 100 ==0:
                print('train cost',cost)
#5.Test
print("\n Testing...........")
cost=model.evaluate(X_test,Y_test,batch_size=40)
print("Test cost:",cost)
W,b=model.layers[0].get_weights()
print('Weight=',W,"\nbiases=",b)

#6.Plotting the prediction
Y_pred=model.predict(X_test)
plt.scatter(X_test,Y_test)
plt.plot(X_test,Y_pred)
plt.show()
 

執行後的結果如下：

[email protected]:~/keraslearn$ python3 helloworld.py
Using TensorFlow backend.
[-0.70854271  0.1758794  -0.30653266  0.74874372 -0.02512563  0.33668342
 -0.85929648  0.01507538 -0.13567839  0.72864322  0.24623116 -0.74874372
 -0.78894472  0.50753769  0.03517588  0.35678392 -0.55778894  0.2361809
 -0.25628141 -0.44723618  0.2160804  -0.43718593 -0.64824121  0.69849246
 -0.03517588 -0.45728643  0.86934673  0.73869347  0.53768844 -0.67839196
 -0.75879397  0.55778894  0.28643216 -0.05527638 -0.86934673  0.1959799
 -0.57788945 -0.9798995  -0.6080402  -0.63819095  0.84924623  0.41708543
  0.13567839  0.79899497 -0.47738693  0.46733668  0.59798995 -0.80904523
 -0.98994975 -0.36683417 -0.5678392  -0.00502513 -0.53768844 -0.37688442
 -0.65829146 -0.1959799   0.06532663  0.44723618 -0.01507538 -0.6281407
  0.02512563 -0.71859296 -0.14572864 -0.46733668  0.07537688  0.85929648
  0.76884422  0.40703518 -0.68844221  0.68844221 -0.29648241  0.66834171
 -0.95979899 -0.33668342  0.26633166 -0.82914573  1.         -0.5879397
 -0.69849246 -0.20603015  0.63819095 -0.88944724 -0.40703518 -0.32663317
  0.15577889 -0.41708543  0.10552764  0.20603015 -0.04522613  0.00502513
 -0.31658291  0.43718593  0.42713568  0.45728643 -0.59798995 -0.66834171
  0.83919598  0.75879397 -0.24623116  0.71859296 -0.92964824  0.39698492
  0.61809045 -0.84924623 -0.87939698 -0.96984925  0.87939698  0.6281407
  0.25628141  0.27638191  0.12562814  0.09547739 -0.89949749  0.80904523
 -0.16582915 -0.12562814  0.30653266  0.49748744  0.5879397  -0.51758794
 -0.10552764  0.54773869 -0.94974874  0.92964824  0.16582915 -0.83919598
 -0.35678392 -0.48743719  0.08542714 -0.61809045  0.18592965  0.57788945
  0.65829146  0.38693467  0.91959799 -0.26633166 -0.50753769 -1.
 -0.54773869  0.6080402  -0.49748744 -0.22613065  0.9798995   0.98994975
  0.5678392   0.32663317  0.64824121 -0.52763819  0.36683417  0.81909548
 -0.11557789  0.31658291 -0.2160804   0.95979899  0.77889447 -0.73869347
 -0.81909548 -0.79899497  0.78894472  0.88944724 -0.2361809   0.37688442
  0.70854271  0.22613065 -0.28643216 -0.38693467  0.90954774 -0.91959799
  0.48743719 -0.42713568 -0.08542714  0.11557789 -0.18592965  0.47738693
 -0.39698492 -0.34673367  0.04522613  0.05527638  0.93969849 -0.77889447
 -0.93969849 -0.06532663 -0.72864322  0.29648241  0.52763819 -0.76884422
  0.94974874  0.82914573  0.34673367 -0.90954774 -0.27638191 -0.15577889
 -0.1758794   0.14572864 -0.09547739  0.96984925  0.67839196 -0.07537688
  0.89949749  0.51758794]
*******************************************
[ 1.67999883  2.15732476  1.86215827  2.35166736  1.93572236  2.18016038
  1.61608065  2.03423062  1.99563479  2.37665292  2.18481192  1.68394294
  1.62204409  2.26767077  1.98888272  2.22171139  1.71104094  2.17892658
  1.91484027  1.82024419  2.16604711  1.81951961  1.54483872  2.34601517
  2.00887174  1.77313928  2.37743864  2.31814604  2.32716849  1.68305339
  1.60808704  2.24648868  2.26423579  1.96845047  1.66389213  2.1151824
  1.76544914  1.5408011   1.68818555  1.65445629  2.34232127  2.21861488
  2.07349814  2.44520484  1.68813908  2.23841182  2.28008276  1.61527365
  1.45239671  1.86837714  1.73063808  2.01510789  1.78710972  1.81579434
  1.7282564   1.88204563  2.12479282  2.27787801  2.04415445  1.64586279
  2.02443768  1.69792605  1.90274178  1.78137239  2.02497528  2.42132526
  2.42777448  2.16564271  1.68763444  2.31923599  1.81271447  2.34163208
  1.48248635  1.78163246  2.0686663   1.56576829  2.57615655  1.65904659
  1.57722512  1.81629077  2.26556137  1.61007302  1.88352331  1.71313059
  2.10104146  1.7038275   1.97249884  2.10127479  2.07381404  2.11472407
  1.83211308  2.26576731  2.23235442  2.28700014  1.72663184  1.71425067
  2.44958621  2.48620331  1.89267965  2.35438093  1.57974268  2.22254286
  2.23280715  1.50250543  1.58598131  1.40667751  2.50025819  2.37980426
  2.19965578  2.12256837  2.05612666  1.98159549  1.49287023  2.43805289
  1.87463195  2.04988238  2.12882858  2.2546237   2.34223092  1.65392499
  1.99025412  2.28291716  1.45807616  2.51044748  1.97988068  1.66059626
  1.7961923   1.75129171  2.0241939   1.69914728  2.07346622  2.33384759
  2.35610171  2.22847144  2.39450148  1.84129314  1.69452659  1.55065089
  1.66503047  2.28833885  1.65681898  1.9227888   2.46308462  2.55316759
  2.3420487   2.1586477   2.31608103  1.72595194  2.20947867  2.43358179
  1.92711345  2.09783848  1.90011119  2.3742013   2.39525115  1.61921788
  1.60036827  1.61487721  2.32314506  2.47010224  1.78412147  2.20253779
  2.29520865  2.13392445  1.88572837  1.80240843  2.37782058  1.53947432
  2.21682792  1.80241496  1.9832941   2.19691473  1.98446141  2.25189264
  1.79373247  1.91751681  2.04975192  1.95401037  2.47210723  1.64954777
  1.53229024  1.89811048  1.56746323  2.10295781  2.18903255  1.5645517
  2.47228584  2.29823263  2.17956294  1.51738366  1.86618515  1.92440005
  1.77639738  2.04854509  1.89432423  2.44443885  2.40954465  1.93669521
  2.4738211   2.26099266]
helloworld.py:25: UserWarning: Update your `Dense` call to the Keras 2 API: `Dense(units=1, input_dim=1)`
  model.add(Dense(output_dim=1,input_dim=1))
Training......
2017-05-30 13:52:34.569118: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-30 13:52:34.569158: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-30 13:52:34.569165: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-05-30 13:52:34.569169: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-30 13:52:34.569176: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
train cost 4.02812
train cost 0.0807406
train cost 0.00588549
train cost 0.00316872
train cost 0.00276465
train cost 0.00266697
train cost 0.00264223
train cost 0.00263594
train cost 0.00263434
train cost 0.00263394
train cost 0.00263383
train cost 0.00263381
train cost 0.0026338
train cost 0.0026338

 Testing...........
40/40 [==============================] - 0s
Test cost: 0.00337669742294
Weight= [[ 0.5063296]]
biases= [ 2.00427961]

@ 匯入相關的Python和Keras的模組（module）

import numpy as np
np.random.seed(1337)
from keras.models import Sequential
from keras.layers import Dense
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt

需要特別注意的是，如果在Ubuntu下，如果我們用的遠端命令列方式，因為在遠端命令列的環境下顯示不了圖形介面，所以需要加入下面的兩行程式碼，否則會執行報錯。

import matplotlib
matplotlib.use('Agg')
 

但是在Ubuntu的圖形化介面下（比如，遠端桌面）不需要。

@ 隨機生成200個數字並模擬一個線性函式

隨機生成200個範圍在-1到1之間一個浮點數。並模擬一個線性函式的公式,0.5*X+2 並加上一些隨機的干擾，生成200個函式結果Y

然後從中抽選出160組資料作為訓練資料，40組作為測試訓練的結果的資料。

# 1.Build the trainning data
X=np.linspace(-1,1,200)
np.random.shuffle(X)
np.random.normal()
Y=0.5*X+2+np.random.normal(0,0.05,(200,))
plt.scatter(X,Y)
plt.show()
X_train,Y_train=X[:160],Y[:160]
X_test,Y_test=X[160:],Y[160:]
print(X)
print("*******************************************")
print(Y)

把下面的兩行註釋掉，並在Ubuntu的圖形介面模式下，顯示出來的圖形效果如下：

import matplotlib
matplotlib.use('Agg')

@ 用Keras的API建立一個神經網路模型

這個模型是總共有隻要一層，1個輸入和一個輸出,建立好神經網路後，選擇損失函式和優化器。從下面的程式碼來看，其實還是非常的簡潔的，如果用TensorFlow或者Theano實現的話，程式碼會多不少，但是用Keras實現非常的簡潔和簡單。

# 2.Build a neural network from the 1st layer to the last layer
model=Sequential()
model.add(Dense(output_dim=1,input_dim=1))

#3. Choose loss function and optimzing method
model.compile(loss='mse',optimizer='sgd')

損失函式只要有下面的幾種：

• mean_squared_error
• mean_absolute_error
• mean_absolute_percentage_error
• mean_squared_logarithmic_error
• squared_hinge
• hinge
• logcosh
• categorical_crossentropy
• sparse_categorical_crossentropy
• binary_crossentropy
• kullback_leibler_divergence
• poisson
• cosine_proximity

具體的意思，請參考：

https://keras.io/losses/

https://github.com/fchollet/keras/blob/master/keras/losses.py

我們當前選擇的是：sgd，其支援，隨機梯度下降法，支援動量引數，支援學習衰減率，支援Nesterov動量。

此外還有優化器：

• RMSprop
• Adagrad
• Adadelta
• Adam
• Adamax
• Nadam
• TFOptimizer

不同的優化器有有不同的適用場景，限於篇幅，請讀者自己補腦和查詢相關資料。

@ 分批次訓練

批次的次數不是越多越好，在當前的例子中，批次的訓練次數達到1300次左右基本上已經達到損失函式能夠達到的最好的結果了，在增加次數也增加了不了精度。具體請見
日誌的輸出。

#4. Trainning
print("Training......")
for step in range(1400):
        cost=model.train_on_batch(X_train,Y_train)
        if step % 100 ==0:
                print('train cost',cost)

@ 測試資料測試訓練結果。

在上面的部分，經過1400次的分批次訓練，神經網路已經完全模擬出了上面的線性函式的模型。這個時候代入剩下的40組測試資料進行測試。我們會發現0.5*X+2 這個線性函式完全被建立起來了。從輸出的weight和biases的值其實就是上面的0.5和2； weight和0.5越接近，說明效果越好；biases和2越接近說明效果越好。

#5.Test
print("\n Testing...........")
cost=model.evaluate(X_test,Y_test,batch_size=40)
print("Test cost:",cost)
W,b=model.layers[0].get_weights()
print('Weight=',W,"\nbiases=",b)

@ Keras模型結果 VS 原始測試資料結果

把通過神經網路得出的結果與原始的測試結果得出的結果進行比較，並顯示其對比影象。

#6.Plotting the prediction
Y_pred=model.predict(X_test)
plt.scatter(X_test,Y_test)
plt.plot(X_test,Y_pred)
plt.show()