用示例学习 Keras
Posted zhuo木鸟
tags:
篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了用示例学习 Keras相关的知识,希望对你有一定的参考价值。
文章目录
但 CNN 的隐藏层大于或等于两层时,最好用 RELU 作为激活函数—— by zhuo 木鸟
网格寻优调参(包括网络层数、节点个数、编译方式等)
以神经网络+鸢尾花数据集为例:
from sklearn.datasets import load_iris
import numpy as np
from sklearn.metrics import make_scorer,f1_score,accuracy_score
from sklearn.linear_model import LogisticRegression
from keras.models import Sequential,model_from_json,model_from_yaml
from keras.layers import Dense
from keras.utils import to_categorical # one-hot 咱不用
from keras.callbacks import ModelCheckpoint
from keras.wrappers.scikit_learn import KerasClassifier
# import matplotlib.pyplot as plt
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import KFold
from sklearn.preprocessing import LabelEncoder #导入LabelEncoder模块
le = LabelEncoder() #实例一个LabelEncoder对象
X,y = load_iris(return_X_y=True) #导入数据集
y = le.fit_transform(y) #转换数据
span = list(set(y)) #将span赋值为test_df的取值范围(转换后)
le.inverse_transform(span) #数据拟转换,可以查看各个数字的含义
seed = 7
np.random.seed(seed)
def create_ann(units_list=[10],optimizer='rmsprop',init='glorot_uniform'):
ann = Sequential()
units=units_list[0]
ann.add(Dense(units=units,activation='relu',
input_shape=(4,),kernel_initializer=init))
for units in units_list[1:]:
ann.add(Dense(units=units,activation='relu',
kernel_initializer=init))
ann.add(Dense(units=3,activation='sigmoid',kernel_initializer=init))
ann.compile(loss='categorical_crossentropy',optimizer=optimizer)
return ann
# ann = create_ann()
# ann.fit(X,y,batch_size=5,epochs=100,verbose=0)
model = KerasClassifier(build_fn=create_ann,epochs=100,batch_size=5,
verbose=0) #可以看到,把 fit 的参数也带进去了。
# 但如果后文有用 grid 将其网格寻优,那么上面的设置其实可以不用的,只需要:
# model = KerasClassifier(build_fn=create_ann,verbose=0)
grid =
grid['units_list']=[[10],[4,6],[3,4,3]]
grid['optimizer']=['rmsprop','adam']
grid['init']=['glorot_uniform','normal']
grid['epochs']=[50,25]
grid['batch_size']=[5,3]
kfold = KFold(n_splits=5,shuffle=True,random_state=seed)
scorer = make_scorer(f1_score,average='macro') #用不了不知道为什么.......
# acc_scorer = make_scorer(accuracy_score)
grid_search = GridSearchCV(estimator=model,param_grid=grid,scoring=scorer,cv=kfold)
results = grid_search.fit(X,y)
print('Best:%f using %s'%(results.best_score_,results.best_params_))
means = results.cv_results_['mean_test_score']
stds = results.cv_results_['std_test_score']
params = results.cv_results_['params']
for mean,std,param in zip(means,stds,params):
print('%f(+-%f) with: %r'%(mean,std,param))
模型筛选——交叉验证
假设有两个模型,一个是上面的神经网络,一个是逻辑回归。如何筛选最好的模型呢?
为了筛选模型,就要评价模型。评价模型的方法有交叉验证:
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score
ann = KerasClassifier(build_fn=create_ann,epochs=100,batch_size=5,
units_list=[4,6],init='normal',
verbose=0)
lg = LogisticRegression(penalty='none')
S_lg_i = cross_val_score(lg,X,y,scoring=scorer,cv=kfold) #计算出逻辑回归模型的Si
S_ann_i = cross_val_score(ann,X,y,scoring=scorer,cv=kfold) #计算出ANN模型的Si
print('逻辑回归: Baseline:%.2f (+-%.2f) f1_score:',%(S_lg_i.mean(),S_lg_i.std()))
print('神经网络: Baseline:%.2f (+-%.2f) f1_score:',%(S_ann_i.mean(),S_ann_i.std()))
正则化 Dropout 与最大范数约束
from keras.layers import Dropout
from keras.constraints import maxnorm
def create_ann(optimizer='rmsprop',init='glorot_uniform'):
ann = Sequential()
ann.add(Dense(units=10,activation='relu',
input_shape=(4,),kernel_initializer=init))
ann.add(Dropout(rate=0.2))
ann.add(Dense(units=6,activation='relu',
kernel_initializer=init))
ann.add(Dropout(rate=0.5)) #最好是 0.2~0.5 这个范围
ann.add(Dense(units=6,activation='relu',
kernel_initializer=init,
kernel_constrain=maxnorm(3)))
ann.add(Dense(units=3,activation='sigmoid',kernel_initializer=init))
ann.compile(loss='categorical_crossentropy',optimizer=optimizer)
return ann
学习率调整
线性衰减
衰减函数:
l
r
k
=
l
r
k
+
1
×
1
1
+
d
e
c
a
y
×
e
p
o
c
h
s
lr_k = lr_k+1\\times \\frac11+decay\\times epochs
lrk=lrk+1×1+decay×epochs1
from keras.optimizers import SGD
learningRate = 0.1 #大学习率
momentum = 0.9 #大动量值 0.9~0.99
decay_rate = 0.005
sgd = SGD(lr=learningRate,momentum=momentum,decay=decay_rate,nesterov=False)
model = KerasClassifier(build_fn=create_ann,epochs=100,batch_size=5,optimizer=sgd,
verbose=0)
指数衰减
l r = l r × d r o p R a t e f l o o r ( 1 + e p o c h s e p o c h D r o p s ) lr = lr \\times dropRate^floor(\\frac1+epochsepochDrops) lr=lr×dropRatefloor(epochDrops1+epochs)
from keras.callbacks import LearningRateScheduler
from math import pow,floor
def step_decay(epoch):
init_lrate = 0.1
dropRate = 0.5
epochDrops = 10
lrate = init_lrate*pow(drop,floor(1+epoch)/epochDrops)
return lrate
learningRate = 0.1 #大学习率
momentum = 0.9 #大动量值 0.9~0.99
decay_rate = 0
sgd = SGD(lr=learningRate,momentum=momentum,decay=decay_rate,nesterov=False)
lrate = LearningRateScheduler(step_decay)
model = KerasClassifier(build_fn=create_ann,epochs=100,batch_size=5,optimizer=sgd,
verbose=0,callbacks=[lrate])
结果可视化
from sklearn.datasets import load_iris
import numpy as np
from sklearn.linear_model import LogisticRegression
X,y = load_iris(return_X_y=True) #导入数据集
from keras.models import Sequential,model_from_json,model_from_yaml
from keras.layers import Dense
from keras.utils import to_categorical
from keras.callbacks import ModelCheckpoint
import matplotlib.pyplot as plt
y = to_categorical(y,num_classes=3)
seed = 7
np.random.seed(seed)
def create_model(optimizer='rmsprop',init='glorot_uniform'):
model = Sequential()
model.add(Dense(units=4,activation='relu',input_dim=4,kernel_initializer=init))
model.add(Dense(units=6,activation='relu',kernel_initializer=init))
model.add(Dense(units=3,activation='sigmoid',kernel_initializer=init))
model.compile(loss='categorical_crossentropy',optimizer=optimizer,metrics=['acc'])
return model
model = create_model()
filepath = 'weights-best.h5'
checkpoint = ModelCheckpoint(filepath=filepath,monitor='val_acc',
verbose=1,save_best_only=True,mode='max')
callback_list=[checkpoint]
history = model.fit(X,y,epochs=50,batch_size=5,verbose=0,callbacks=callback_list,validation_split=0.3)
print(history.history.keys())
font1 = 'family' : 'Times New Roman',
'weight' : 'normal',
'size' : 20,
plt.rcParams['font.sans-serif']=['SimHei']
plt.rcParams['axes.unicode_minus'] = False
plt.figure(figsize=(12,4))
plt.subplots_adjust(left=0.125, bottom=None, right=0.9, top=None,
wspace=0.3, hspace=None)
plt.subplot(1,2,1)
plt.plot(history.history['loss'],linewidth=3,label='Train')
plt.plot(history.history['val_loss'],linewidth=3,linestyle='dashed',label='Test')
plt.xlabel('Epoch',fontsize=20)
plt.ylabel('loss',fontsize=20)
plt.legend(prop=font1)
plt.subplot(1,2,2)
plt.plot(history.history['acc'],linewidth=3,label='Train')
plt.plot(history.history['val_acc'],linewidth=3,linestyle='dashed',label='Test')
plt.xlabel('Epoch',fontsize=20)
plt.ylabel('Acc',fontsize=20)
plt.legend(prop=font1)
两种学习方法(重要)
1、从 create_model 函数中,直接建立:ann = create_model()。然后,在用 ann.fit。当然,在 fit 里面,要设置好 fit 的参数,如 validation_data,validation_split,epochs,batch_size,verbose 等等。
2、用 KerasClassifier 封装,ann = KerasClassifier(build_fn = create_model, arg*) 。这里的参数,不仅可以设置 fit 的参数,同时还可以设置 build_fn 的参数。不过,build_fn 的参数主要是编译时的参数,编译时的参数有:metrics,loss,optimizer。然后,metrics 不可以用 scorer 替代,只能用 keras 内置的 acc、mse 填进去。当然,build_fn 的参数可能还有 units_list,用来调整网络拓扑的、或者 rate,maxnorm,用来调整正则化参数。
模型文件管理
保存为 pickle 文件
pickle 不仅仅可以用于保存模型(包括权重),而且可以用来保存数据:
from sklearn.externals import joblib
:
joblib.dump(要保存的变量名,r'D:\\xxx\\xxx\\data.pkl')
载入 pickle 文件:
data = joblib.load(r'D:\\xxx\\xxx\\data.pkl')
当然,模型也可以用同样的方式,保存为 pickle 文件:
joblib.dump(要保存的模型变量,r'D:\\xxx\\xxx\\model.pkl')
保存为 HDF5 文件
注意,保存为 HDF5、JSON、YAML 只有 keras 才可以,sklearn 不可以
from keras.models import Sequential
from keras.layers import Dense,Activation #导入神经层构造包
from keras.utils import to_categorical #导入one-hot编码法
from sklearn.datasets import load_boston
from keras.models import load_model
X,y = load_boston(return_X_y=True)
ANN = Sequential() #定义一个sequential类,以便构造神经网络
ANN.add(Dense(units=64,activation=’relu’,input_shape=(len(X[1,:]),)))
ANN.add(Dense(units=1,activation=’linear’)) #输出层,units即节点个数
ANN.compile(optimizer=’adam’,loss=’mse’)
#编译模型
ANN.fit(X,y,epochs=100,batch_size=50)
#对模型进行训练
ANN.save(‘model.h5’) #保存模型为model.h5文件
model = load_model(“model.h5”) #读取model.h5文件,并导入模型
保存为 JSON 与 YAML
from sklearn.datasets import load_iris
import numpy as np
from sklearn.linear_model import LogisticRegression
X,y = load_iris(return_X_y=True) #导入数据集
from keras.models import Sequential,model_from_json,model_from_yaml
from keras.layers import Dense
from keras.utils import to_categorical
y = to_categorical(y,num_classes=3)
seed = 7
np.random.seed(seed)
def create_model(optimizer='rmsprop',init='glorot_uniform'):
model = Sequential()
model.add(Dense(units=4,activation='relu',input_dim=4,kernel_initializer=init))
model.add(Dense(units=6,activation='relu',kernel_initializer=init))
model.add(Dense(units=3,activation='sigmoid',kernel_initializer=init))
model.compile(loss='categorical_crossentropy',optimizer=optimizer,metrics=['accuracy'])
return model
model = create_model()
model.fit(X,y,epochs=50,batch_size=5,verbose=0)
model_json = model.to_json()
model_yaml = model.to_yaml()
with open('model.json','w') as file:
file.write(model_json)
with open('model.yaml','w') as file:
file.write(model_yaml)
model.save_weights('model_weights.h5')
with open('model.json','r') as file:
model_json_load = file.read()
with open('model.yaml','r') as file:
model_yaml_load = file.read()
model_load1 = model_from_json(model_json_load)
model_load2 = model_from_yaml(model_yaml_load)
model_load1.load_weights('model_weights.h5')
model_load2.load_weights('model_weights.h5')
这里的保存模型,是将网络的拓扑结构保存为 JSON 或者 YAML。与保存为 HDF5 或者 Pickle 不同的是,它们保存的是模型的“形式”,而后者既保存了模型的拓扑、又同时保存了模型的参数。
当然,可以用 model.save_weights 的方式,将模型的参数单独保存为一个 HDF5 文件,然后再将网络的拓扑结构,保存为 JSON 或者 YAML 文件。
模型更新
一般的,为了保证模型的时效性,通常需要定期对模型进行更新。这个时间通常是 1~2 个月,或者 3-6 个月。更新的方法可以是全量更新、也可以是增量更新。
假设数据的增量为 X_increment,y_increment,那么增量更新可以直接是:
model.compile(loss='categorical_crossentropy',optimizer='rmsprop',metrics=['acc']
model.fit( X_increment,y_increment,epochs=10,batch_size=5,verbose=2) #verbose 控制显示
也就是说,增量更新可以让我们重新编译模型(当然也可以不)。当然,全量更新是将新增的数据,加入
以上是关于用示例学习 Keras的主要内容,如果未能解决你的问题,请参考以下文章
Tensorflow+kerasKeras 用Class类封装的模型如何调试call子函数的模型内部变量
基于Docker的TensorFlow机器学习框架搭建和实例源码解读