TensorFlow的培训实用程序。
train的Python项目详细描述
TensorFlow的培训实用程序。
安装
pip install tensorflow
然后运行:
pip install train
开始
fromtrainimportModel,GradientDescentimporttensorflowastf# Define the network architecture - layers, number of units, activations etc.defnetwork(inputs):hidden=tf.layers.Dense(units=64,activation=tf.nn.relu)(inputs)outputs=tf.layers.Dense(units=10)(hidden)returnoutputs# Configure the learning process - loss, optimizer, evaluation metrics etc.model=Model(network,loss='sparse_softmax_cross_entropy',optimizer=GradientDescent(0.001),metrics=['accuracy'])# Train the model using training datamodel.train(x_train,y_train,epochs=30,batch_size=128)# Evaluate the model performance on test or validation dataloss_and_metrics=model.evaluate(x_test,y_test)# Use the model to make predictions for new datapredictions=model.predict(x)# or call the model directlypredictions=model(x)
更多配置选项可用:
model=Model(network,loss='sparse_softmax_cross_entropy',optimizer=GradientDescent(0.001),metrics=['accuracy'],model_dir='/tmp/my_model')
您还可以使用自定义的损失和度量函数:
defcustom_loss(labels,outputs):passdefcustom_metric(labels,outputs):passmodel=Model(network,loss=custom_loss,optimizer=GradientDescent(0.001),metrics=['accuracy',custom_metric])
模型函数
要获得更多控制,可以使用Estimator
类在函数内部配置模型:
fromtrainimportEstimator,PREDICTimporttensorflowastfdefmodel(features,labels,mode):# Define the network architecturehidden=tf.layers.Dense(units=64,activation=tf.nn.relu)(features)outputs=tf.layers.Dense(units=10)(hidden)predictions=tf.argmax(outputs,axis=1)# In prediction mode, simply return predictions without configuring learning processifmode==PREDICT:returnpredictions# Configure the learning process for training and evaluation modesloss=tf.losses.sparse_softmax_cross_entropy(labels,outputs)optimizer=tf.train.GradientDescentOptimizer(0.001)accuracy=tf.metrics.accuracy(labels,predictions)returndict(loss=loss,optimizer=optimizer,metrics={'accuracy':accuracy})# Create the model using model functionmodel=Estimator(model)# Train the modelmodel.train(x_train,y_train,epochs=30,batch_size=128)
mode
参数指定模型是用于训练、评估还是预测
培训模式
对于像Dropout
这样的层,可以使用training mode变量:
fromtrainimporttrainingx=tf.layers.Dropout(rate=0.4)(x,training=training())
Model
和Estimator
类自动管理训练模式变量。要手动更改,请使用:
# Enter training modetraining(True,session=session)# Exit training modetraining(False,session=session)