TensorFlow2.0使用keras训练模型的实现

2023-07-05 09:25:04 74

1.一般的模型构造、训练、测试流程

#模型构造
inputs=keras.Input(shape=(784,),name='mnist_input')
h1=layers.Dense(64,activation='relu')(inputs)
h1=layers.Dense(64,activation='relu')(h1)
outputs=layers.Dense(10,activation='softmax')(h1)
model=keras.Model(inputs,outputs)
#keras.utils.plot_model(model,'net001.png',show_shapes=True)

model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])

#载入数据
(x_train,y_train),(x_test,y_test)=keras.datasets.mnist.load_data()
x_train=x_train.reshape(60000,784).astype('float32')/255
x_test=x_test.reshape(10000,784).astype('float32')/255

x_val=x_train[-10000:]
y_val=y_train[-10000:]

x_train=x_train[:-10000]
y_train=y_train[:-10000]

#训练模型
history=model.fit(x_train,y_train,batch_size=64,epochs=3,
validation_data=(x_val,y_val))
print('history:')
print(history.history)

result=model.evaluate(x_test,y_test,batch_size=128)
print('evaluate:')
print(result)
pred=model.predict(x_test[:2])
print('predict:')
print(pred)

2.自定义损失和指标

自定义指标只需继承Metric类，并重写一下函数

_init_(self)，初始化。

update_state(self，y_true，y_pred，sample_weight=None)，它使用目标y_true和模型预测y_pred来更新状态变量。

result(self)，它使用状态变量来计算最终结果。

reset_states(self)，重新初始化度量的状态。

#这是一个简单的示例，显示如何实现CatgoricalTruePositives指标，该指标计算正确分类为属于给定类的样本数量

classCatgoricalTruePostives(keras.metrics.Metric):
def__init__(self,name='binary_true_postives',**kwargs):
super(CatgoricalTruePostives,self).__init__(name=name,**kwargs)
self.true_postives=self.add_weight(name='tp',initializer='zeros')

defupdate_state(self,y_true,y_pred,sample_weight=None):
y_pred=tf.argmax(y_pred)
y_true=tf.equal(tf.cast(y_pred,tf.int32),tf.cast(y_true,tf.int32))

y_true=tf.cast(y_true,tf.float32)

ifsample_weightisnotNone:
sample_weight=tf.cast(sample_weight,tf.float32)
y_true=tf.multiply(sample_weight,y_true)

returnself.true_postives.assign_add(tf.reduce_sum(y_true))

defresult(self):
returntf.identity(self.true_postives)

defreset_states(self):
self.true_postives.assign(0.)


model.compile(optimizer=keras.optimizers.RMSprop(1e-3),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[CatgoricalTruePostives()])

model.fit(x_train,y_train,
batch_size=64,epochs=3)

#以定义网络层的方式添加网络loss
classActivityRegularizationLayer(layers.Layer):
defcall(self,inputs):
self.add_loss(tf.reduce_sum(inputs)*0.1)
returninputs

inputs=keras.Input(shape=(784,),name='mnist_input')
h1=layers.Dense(64,activation='relu')(inputs)
h1=ActivityRegularizationLayer()(h1)
h1=layers.Dense(64,activation='relu')(h1)
outputs=layers.Dense(10,activation='softmax')(h1)
model=keras.Model(inputs,outputs)
#keras.utils.plot_model(model,'net001.png',show_shapes=True)

model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])
model.fit(x_train,y_train,batch_size=32,epochs=1)

#也可以以定义网络层的方式添加要统计的metric
classMetricLoggingLayer(layers.Layer):
defcall(self,inputs):
self.add_metric(keras.backend.std(inputs),
name='std_of_activation',
aggregation='mean')

returninputs

inputs=keras.Input(shape=(784,),name='mnist_input')
h1=layers.Dense(64,activation='relu')(inputs)
h1=MetricLoggingLayer()(h1)
h1=layers.Dense(64,activation='relu')(h1)
outputs=layers.Dense(10,activation='softmax')(h1)
model=keras.Model(inputs,outputs)
#keras.utils.plot_model(model,'net001.png',show_shapes=True)

model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])
model.fit(x_train,y_train,batch_size=32,epochs=1)

#也可以直接在model上面加
#也可以以定义网络层的方式添加要统计的metric
classMetricLoggingLayer(layers.Layer):
defcall(self,inputs):
self.add_metric(keras.backend.std(inputs),
name='std_of_activation',
aggregation='mean')

returninputs

inputs=keras.Input(shape=(784,),name='mnist_input')
h1=layers.Dense(64,activation='relu')(inputs)
h2=layers.Dense(64,activation='relu')(h1)
outputs=layers.Dense(10,activation='softmax')(h2)
model=keras.Model(inputs,outputs)

model.add_metric(keras.backend.std(inputs),
name='std_of_activation',
aggregation='mean')
model.add_loss(tf.reduce_sum(h1)*0.1)

#keras.utils.plot_model(model,'net001.png',show_shapes=True)

model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])
model.fit(x_train,y_train,batch_size=32,epochs=1)

处理使用validation_data传入测试数据，还可以使用validation_split划分验证数据

ps:validation_split只能在用numpy数据训练的情况下使用

model.fit(x_train,y_train,batch_size=32,epochs=1,validation_split=0.2)

3.使用tf.data构造数据

defget_compiled_model():
inputs=keras.Input(shape=(784,),name='mnist_input')
h1=layers.Dense(64,activation='relu')(inputs)
h2=layers.Dense(64,activation='relu')(h1)
outputs=layers.Dense(10,activation='softmax')(h2)
model=keras.Model(inputs,outputs)
model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])
returnmodel
model=get_compiled_model()
train_dataset=tf.data.Dataset.from_tensor_slices((x_train,y_train))
train_dataset=train_dataset.shuffle(buffer_size=1024).batch(64)

val_dataset=tf.data.Dataset.from_tensor_slices((x_val,y_val))
val_dataset=val_dataset.batch(64)

#model.fit(train_dataset,epochs=3)
#steps_per_epoch每个epoch只训练几步
#validation_steps每次验证，验证几步
model.fit(train_dataset,epochs=3,steps_per_epoch=100,
validation_data=val_dataset,validation_steps=3)

4.样本权重和类权重

“样本权重”数组是一个数字数组，用于指定批处理中每个样本在计算总损失时应具有多少权重。它通常用于不平衡的分类问题（这个想法是为了给予很少见的类更多的权重）。当使用的权重是1和0时，该数组可以用作损失函数的掩码（完全丢弃某些样本对总损失的贡献）。

“类权重”dict是同一概念的更具体的实例：它将类索引映射到应该用于属于该类的样本的样本权重。例如，如果类“0”比数据中的类“1”少两倍，则可以使用class_weight={0：1.，1：0.5}。

#增加第5类的权重
importnumpyasnp
#样本权重
model=get_compiled_model()
class_weight={i:1.0foriinrange(10)}
class_weight[5]=2.0
print(class_weight)
model.fit(x_train,y_train,
class_weight=class_weight,
batch_size=64,
epochs=4)
#类权重
model=get_compiled_model()
sample_weight=np.ones(shape=(len(y_train),))
sample_weight[y_train==5]=2.0
model.fit(x_train,y_train,
sample_weight=sample_weight,
batch_size=64,
epochs=4)

#tf.data数据
model=get_compiled_model()

sample_weight=np.ones(shape=(len(y_train),))
sample_weight[y_train==5]=2.0

train_dataset=tf.data.Dataset.from_tensor_slices((x_train,y_train,
sample_weight))
train_dataset=train_dataset.shuffle(buffer_size=1024).batch(64)

val_dataset=tf.data.Dataset.from_tensor_slices((x_val,y_val))
val_dataset=val_dataset.batch(64)

model.fit(train_dataset,epochs=3,)

5.多输入多输出模型

image_input=keras.Input(shape=(32,32,3),name='img_input')
timeseries_input=keras.Input(shape=(None,10),name='ts_input')

x1=layers.Conv2D(3,3)(image_input)
x1=layers.GlobalMaxPooling2D()(x1)

x2=layers.Conv1D(3,3)(timeseries_input)
x2=layers.GlobalMaxPooling1D()(x2)

x=layers.concatenate([x1,x2])

score_output=layers.Dense(1,name='score_output')(x)
class_output=layers.Dense(5,activation='softmax',name='class_output')(x)

model=keras.Model(inputs=[image_input,timeseries_input],
outputs=[score_output,class_output])
keras.utils.plot_model(model,'multi_input_output_model.png'
,show_shapes=True)

#可以为模型指定不同的loss和metrics
model.compile(
optimizer=keras.optimizers.RMSprop(1e-3),
loss=[keras.losses.MeanSquaredError(),
keras.losses.CategoricalCrossentropy()])

#还可以指定loss的权重
model.compile(
optimizer=keras.optimizers.RMSprop(1e-3),
loss={'score_output':keras.losses.MeanSquaredError(),
'class_output':keras.losses.CategoricalCrossentropy()},
metrics={'score_output':[keras.metrics.MeanAbsolutePercentageError(),
keras.metrics.MeanAbsoluteError()],
'class_output':[keras.metrics.CategoricalAccuracy()]},
loss_weight={'score_output':2.,'class_output':1.})

#可以把不需要传播的loss置0
model.compile(
optimizer=keras.optimizers.RMSprop(1e-3),
loss=[None,keras.losses.CategoricalCrossentropy()])

#Ordictlossversion
model.compile(
optimizer=keras.optimizers.RMSprop(1e-3),
loss={'class_output':keras.losses.CategoricalCrossentropy()})

6.使用回调

Keras中的回调是在训练期间（在epoch开始时，batch结束时，epoch结束时等）在不同点调用的对象，可用于实现以下行为：

在培训期间的不同时间点进行验证（超出内置的每个时期验证）
定期检查模型或超过某个精度阈值
在训练似乎平稳时改变模型的学习率
在训练似乎平稳时对顶层进行微调
在培训结束或超出某个性能阈值时发送电子邮件或即时消息通知等等。

可使用的内置回调有

ModelCheckpoint：定期保存模型。
EarlyStopping：当训练不再改进验证指标时停止培训。
TensorBoard：定期编写可在TensorBoard中显示的模型日志（更多细节见“可视化”）。
CSVLogger：将丢失和指标数据流式传输到CSV文件。
等等

6.1回调使用

model=get_compiled_model()

callbacks=[
keras.callbacks.EarlyStopping(
#Stoptrainingwhen`val_loss`isnolongerimproving
monitor='val_loss',
#"nolongerimproving"beingdefinedas"nobetterthan1e-2less"
min_delta=1e-2,
#"nolongerimproving"beingfurtherdefinedas"foratleast2epochs"
patience=2,
verbose=1)
]
model.fit(x_train,y_train,
epochs=20,
batch_size=64,
callbacks=callbacks,
validation_split=0.2)

#checkpoint模型回调
model=get_compiled_model()
check_callback=keras.callbacks.ModelCheckpoint(
filepath='mymodel_{epoch}.h5',
save_best_only=True,
monitor='val_loss',
verbose=1
)

model.fit(x_train,y_train,
epochs=3,
batch_size=64,
callbacks=[check_callback],
validation_split=0.2)

#动态调整学习率
initial_learning_rate=0.1
lr_schedule=keras.optimizers.schedules.ExponentialDecay(
initial_learning_rate,
decay_steps=10000,
decay_rate=0.96,
staircase=True
)
optimizer=keras.optimizers.RMSprop(learning_rate=lr_schedule)

#使用tensorboard
tensorboard_cbk=keras.callbacks.TensorBoard(log_dir='./full_path_to_your_logs')
model.fit(x_train,y_train,
epochs=5,
batch_size=64,
callbacks=[tensorboard_cbk],
validation_split=0.2)

6.2创建自己的回调方法

classLossHistory(keras.callbacks.Callback):
defon_train_begin(self,logs):
self.losses=[]
defon_epoch_end(self,batch,logs):
self.losses.append(logs.get('loss'))
print('\nloss:',self.losses[-1])

model=get_compiled_model()

callbacks=[
LossHistory()
]
model.fit(x_train,y_train,
epochs=3,
batch_size=64,
callbacks=callbacks,
validation_split=0.2)

7.自己构造训练和验证循环

#Getthemodel.
inputs=keras.Input(shape=(784,),name='digits')
x=layers.Dense(64,activation='relu',name='dense_1')(inputs)
x=layers.Dense(64,activation='relu',name='dense_2')(x)
outputs=layers.Dense(10,activation='softmax',name='predictions')(x)
model=keras.Model(inputs=inputs,outputs=outputs)

#Instantiateanoptimizer.
optimizer=keras.optimizers.SGD(learning_rate=1e-3)
#Instantiatealossfunction.
loss_fn=keras.losses.SparseCategoricalCrossentropy()

#Preparethetrainingdataset.
batch_size=64
train_dataset=tf.data.Dataset.from_tensor_slices((x_train,y_train))
train_dataset=train_dataset.shuffle(buffer_size=1024).batch(batch_size)

#自己构造循环
forepochinrange(3):
print('epoch:',epoch)
forstep,(x_batch_train,y_batch_train)inenumerate(train_dataset):
#开一个gradienttape,计算梯度
withtf.GradientTape()astape:
logits=model(x_batch_train)

loss_value=loss_fn(y_batch_train,logits)
grads=tape.gradient(loss_value,model.trainable_variables)
optimizer.apply_gradients(zip(grads,model.trainable_variables))

ifstep%200==0:
print('Trainingloss(foronebatch)atstep%s:%s'%(step,float(loss_value)))
print('Seensofar:%ssamples'%((step+1)*64))

#训练并验证
#Getmodel
inputs=keras.Input(shape=(784,),name='digits')
x=layers.Dense(64,activation='relu',name='dense_1')(inputs)
x=layers.Dense(64,activation='relu',name='dense_2')(x)
outputs=layers.Dense(10,activation='softmax',name='predictions')(x)
model=keras.Model(inputs=inputs,outputs=outputs)

#Instantiateanoptimizertotrainthemodel.
optimizer=keras.optimizers.SGD(learning_rate=1e-3)
#Instantiatealossfunction.
loss_fn=keras.losses.SparseCategoricalCrossentropy()

#Preparethemetrics.
train_acc_metric=keras.metrics.SparseCategoricalAccuracy()
val_acc_metric=keras.metrics.SparseCategoricalAccuracy()

#Preparethetrainingdataset.
batch_size=64
train_dataset=tf.data.Dataset.from_tensor_slices((x_train,y_train))
train_dataset=train_dataset.shuffle(buffer_size=1024).batch(batch_size)

#Preparethevalidationdataset.
val_dataset=tf.data.Dataset.from_tensor_slices((x_val,y_val))
val_dataset=val_dataset.batch(64)


#Iterateoverepochs.
forepochinrange(3):
print('Startofepoch%d'%(epoch,))

#Iterateoverthebatchesofthedataset.
forstep,(x_batch_train,y_batch_train)inenumerate(train_dataset):
withtf.GradientTape()astape:
logits=model(x_batch_train)
loss_value=loss_fn(y_batch_train,logits)
grads=tape.gradient(loss_value,model.trainable_variables)
optimizer.apply_gradients(zip(grads,model.trainable_variables))

#Updatetrainingmetric.
train_acc_metric(y_batch_train,logits)

#Logevery200batches.
ifstep%200==0:
print('Trainingloss(foronebatch)atstep%s:%s'%(step,float(loss_value)))
print('Seensofar:%ssamples'%((step+1)*64))

#Displaymetricsattheendofeachepoch.
train_acc=train_acc_metric.result()
print('Trainingaccoverepoch:%s'%(float(train_acc),))
#Resettrainingmetricsattheendofeachepoch
train_acc_metric.reset_states()

#Runavalidationloopattheendofeachepoch.
forx_batch_val,y_batch_valinval_dataset:
val_logits=model(x_batch_val)
#Updatevalmetrics
val_acc_metric(y_batch_val,val_logits)
val_acc=val_acc_metric.result()
val_acc_metric.reset_states()
print('Validationacc:%s'%(float(val_acc),))

##添加自己构造的loss,每次只能看到最新一次训练增加的loss
classActivityRegularizationLayer(layers.Layer):

defcall(self,inputs):
self.add_loss(1e-2*tf.reduce_sum(inputs))
returninputs

inputs=keras.Input(shape=(784,),name='digits')
x=layers.Dense(64,activation='relu',name='dense_1')(inputs)
#Insertactivityregularizationasalayer
x=ActivityRegularizationLayer()(x)
x=layers.Dense(64,activation='relu',name='dense_2')(x)
outputs=layers.Dense(10,activation='softmax',name='predictions')(x)

model=keras.Model(inputs=inputs,outputs=outputs)
logits=model(x_train[:64])
print(model.losses)
logits=model(x_train[:64])
logits=model(x_train[64:128])
logits=model(x_train[128:192])
print(model.losses)

#将loss添加进求导中
optimizer=keras.optimizers.SGD(learning_rate=1e-3)

forepochinrange(3):
print('Startofepoch%d'%(epoch,))

forstep,(x_batch_train,y_batch_train)inenumerate(train_dataset):
withtf.GradientTape()astape:
logits=model(x_batch_train)
loss_value=loss_fn(y_batch_train,logits)

#Addextralossescreatedduringthisforwardpass:
loss_value+=sum(model.losses)

grads=tape.gradient(loss_value,model.trainable_variables)
optimizer.apply_gradients(zip(grads,model.trainable_variables))

#Logevery200batches.
ifstep%200==0:
print('Trainingloss(foronebatch)atstep%s:%s'%(step,float(loss_value)))
print('Seensofar:%ssamples'%((step+1)*64))

到此这篇关于TensorFlow2.0使用keras训练模型的实现的文章就介绍到这了,更多相关TensorFlow2.0keras训练模型内容请搜索毛票票以前的文章或继续浏览下面的相关文章希望大家以后多多支持毛票票！

声明：本文内容来源于网络，版权归原作者所有，内容由互联网用户自发贡献自行上传，本网站不拥有所有权，未作人工编辑处理，也不承担相关法律责任。如果您发现有涉嫌版权的内容，欢迎发送邮件至：czq8825#qq.com（发邮件时，请将#更换为@）进行举报，并提供相关证据，一经查实，本站将立刻删除涉嫌侵权内容。