数据分析师养成之路--keras学习中metrics和callbacks笔记
complie中的metrics参数
model.compile(...metrics=['accuracy'])如上,
metrics中有两个arguments,y_true,和y_pred
它所接收的评估指标为,损失函数,(mse等)和准确率(accuracy等),它们都作用于训练集和验证集(eg: loss:..acc..mae .. val_loss:..val_acc..val_mae..),当然该性能的评估结果不会真正用于训练,只用来显示(作为参考)
(每个batch,输出:loss acc mae (后两个就是设置metrics得到),每个epoch(一轮),输出:loss,acc mae … val_loss…val_acc… val_mae)
metrics中的损失函数和准确率函数可由自己改写:
如下:
def top_3_accuracy(y_true,y_pred):
# 返回的函数即为metrics 中支持的函数
return metrics.top_k_categorical_accuracy(y_true,y_pred,k=3)
model.compile(loss='categorical_crossentropy',
optimizer=sgd,
metrics=[metrics.mae,top_3_accuracy])import keras.backend as K
# 对metrics中的accuracy函数进行重写
def mean_pred(y_true, y_pred):
return K.mean(y_pred)
model.compile(optimizer='rmsprop',
loss='binary_crossentropy',
metrics=['accuracy', mean_pred])fit中的callbacks参数:
from keras.callbacks import ModelCheckpoint
model = Sequential()
model.add(Dense(10, input_dim=784, kernel_initializer='uniform'))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
'''
saves the model weights after each epoch if the validation loss decreased
'''
checkpointer = ModelCheckpoint(filepath='/tmp/weights.hdf5', verbose=1, save_best_only=True)
model.fit(x_train, y_train, batch_size=128, epochs=20, verbose=0, validation_data=(X_test, Y_test), callbacks=[checkpointer])
class LossHistory(keras.callbacks.Callback):
def on_train_begin(self, logs={}):
self.losses = []
def on_batch_end(self, batch, logs={}):
self.losses.append(logs.get('loss'))
model = Sequential()
model.add(Dense(10, input_dim=784, kernel_initializer='uniform'))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy', optimizer='rmsprop')
history = LossHistory()
model.fit(x_train, y_train, batch_size=128, epochs=20, verbose=0, callbacks=[history])
print(history.losses)
# outputs
'''
[0.66047596406559383, 0.3547245744908703, ..., 0.25953155204159617, 0.25901699725311789]
'''就我而言,callbacks很方便作为 验证集数据评估的参考
如下,为fit_generator的callbacks参数写一个RocAuc接口
class RocAuc(keras.callbacks.Callback):
def __init__(self,validation_generate,interval=1):
self.interval=interval
self.validation_generate=validation_generate
def on_epoch_end(self,epoch, logs={}):
# 每次epoch,读取一批生成的数据
x_val,y_val=next(self.validation_generate)
#print(y_val)
if epoch % self.interval == 0:
try:
y_pred=self.model.predict(x_val,verbose=0)
score=roc_auc_score(y_val,y_pred)
print('\n ROC_AUC - epoch:%d - score:%.6f \n' % (epoch + 1, score*100))
except:
print('\n epoch:%d only one class!!\n' % (epoch + 1))
def validata(iter_test):
for index in iter_test:
yield index
vli=validata(iter_test)
rocauc=RocAuc(validation_generate=vli)
model.fit_generator(...callbacks=[rocauc])当然也可以把earlyStopping加入其中,不过需要传入roc_auc_val 参数,
class RocAuc(keras.callbacks.Callback):
def __init__(self,validation_generate,interval=1):
self.interval=interval
self.validation_generate=validation_generate
def on_train_begin(self,logs={}):
#添加roc_auc_val属性
self.roc_auc_score=[]
if not ('roc_auc_val' in self.params['metrics']):
self.params['metrics'].append('roc_auc_val')
def on_epoch_end(self,epoch, logs={}):
logs['roc_auc_val']=float('-inf')
x_val,y_val=next(self.validation_generate)
#print(y_val)
if epoch % self.interval == 0:
try:
y_pred=self.model.predict(x_val,verbose=0)
score=roc_auc_score(y_val,y_pred)
logs['roc_auc_val']=score*100
print('\n ROC_AUC - epoch:%d - score:%.6f \n' % (epoch + 1, score*100))
except:
logs['roc_auc_val']='cant compute!!'
print('\n epoch:%d only one class!!\n' % (epoch + 1))
self.roc_auc_score.append(logs.get('roc_auc_val'))
def validata(iter_test):
for index in iter_test:
yield index
vli=validata(iter_test)
model.fit_generator(...callbacks=[rocauc, EarlyStopping(monitor='roc_auc_val',patience=30, verbose=2,mode='max')])
print(rocauc.roc_auc_score)当然也可以添加plot,显示batch,epoch的loss,acc输出趋势.
( 比如,我们看我们的模型是否收敛时,(如果源数据太大,可以抽取一部分(.sample(xxx)),看其模型的Loss,val_loss,是否逐渐下降(主要看loss)acc,val_acc是否逐渐增长,若loss已经平衡了/acc已经达到最大,不变了,那么可以考虑改变learningRate)
class LossHistory(keras.callbacks.Callback):
def __init__(self):
self.loss={'count':[],'batch':[],'epoch':[]}
self.acc={'count':[],'batch':[],'epoch':[]}
self.val_loss={'count':[],'batch':[],'epoch':[]}
self.val_acc={'count':[],'batch':[],'epoch':[]}
def on_batch_end(self,batch,logs=[]):
self.loss['batch'].append(logs.get('loss'))
self.acc['batch'].append(logs.get('acc'))
self.loss['count'].append(batch)
self.acc['count'].append(batch)
def on_epoch_end(self,epoch,logs=[]):
self.val_loss['epoch'].append(logs.get('loss'))
self.val_acc['epoch'].append(logs.get('acc'))
self.val_loss['epoch'].append(logs.get('val_loss'))
self.val_acc['epoch'].append(logs.get('val_acc'))
self.val_loss['count'].append(epoch)
self.val_acc['count'].append(epoch)
def loss_plot(self,loss_type):
iters=range(len(loss_type)-1)
plt.figure()
if loss_type=='batch':
plt.plot(iters,self.acc[loss_type],'g',label='train_acc')
plt.plot(iters,self.loss[loss_type],'b',label='train_loss')
if loss_type=='epoch':
plt.plot(iters,self.val_acc[loss_type],'r',label='val_acc')
plt.plot(iters,self.val_loss[loss_type],'r',label='val_loss')
plt.grid(True)
plt.xlabel(loss_type)
plt.ylabel('acc-loss')
plt.legend(loc='upper right')
plt.show()