多种条件提前停止

Question

我正在为推荐系统（项目推荐）进行多类分类，目前正在使用sparse_categorical_crossentropy损失来训练我的网络。EarlyStopping因此，通过监控我的验证损失来执行是合理的，val_loss如下所示：tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=10)其按预期工作。然而，网络（推荐系统）的性能是通过 Average-Precision-at-10 来衡量的，并在训练期间作为指标进行跟踪，如average_precision_at_k10。因此，我还可以使用此指标执行提前停止：tf.keras.callbacks.EarlyStopping(monitor='average_precision_at_k10', patience=10)这也按预期工作。我的问题： 有时验证损失会增加，而 10 处的平均精度会提高，反之亦然。因此，当且仅当两者都恶化时，我需要监测两者，并尽早停止。我想做的事：tf.keras.callbacks.EarlyStopping(monitor=['val_loss', 'average_precision_at_k10'], patience=10)这显然不起作用。有什么想法可以做到这一点吗？

Answer 1

我成功创建了自己的自定义 EarlyStopping 回调，并认为我将其发布在这里，以防其他人想要实现类似的东西。

如果验证损失和10 时的 平均精度对于epoch 数没有改善，则执行早期停止。patience

class CustomEarlyStopping(keras.callbacks.Callback):

    def __init__(self, patience=0):

        super(CustomEarlyStopping, self).__init__()

        self.patience = patience

        self.best_weights = None

        

    def on_train_begin(self, logs=None):

        # The number of epoch it has waited when loss is no longer minimum.

        self.wait = 0

        # The epoch the training stops at.

        self.stopped_epoch = 0

        # Initialize the best as infinity.

        self.best_v_loss = np.Inf

        self.best_map10 = 0

    def on_epoch_end(self, epoch, logs=None): 

        v_loss=logs.get('val_loss')

        map10=logs.get('val_average_precision_at_k10')

        # If BOTH the validation loss AND map10 does not improve for 'patience' epochs, stop training early.

        if np.less(v_loss, self.best_v_loss) and np.greater(map10, self.best_map10):

            self.best_v_loss = v_loss

            self.best_map10 = map10

            self.wait = 0

            # Record the best weights if current results is better (less).

            self.best_weights = self.model.get_weights()

        else:

            self.wait += 1

            if self.wait >= self.patience:

                self.stopped_epoch = epoch

                self.model.stop_training = True

                print("Restoring model weights from the end of the best epoch.")

                self.model.set_weights(self.best_weights)

                

    def on_train_end(self, logs=None):

        if self.stopped_epoch > 0:

            print("Epoch %05d: early stopping" % (self.stopped_epoch + 1))

然后将其用作：

model.fit(

    x_train,

    y_train,

    batch_size=64,

    steps_per_epoch=5,

    epochs=30,

    verbose=0,

    callbacks=[CustomEarlyStopping(patience=10)],

)

Answer 2

您可以通过创建自定义回调来实现此目的。下面是一些代码，说明了您可以在自定义回调中执行哪些操作。我引用的文档显示了许多其他选项。

class LRA(keras.callbacks.Callback): # subclass the callback class

# create class variables as below. These can be accessed in your code outside the class definition as LRA.my_class_variable, LRA.best_weights

    my_class_variable=something  # a class variable

    best_weights=model.get_weights() # another  class variable

# define an initialization function with parameters you want to feed to the callback

    def __init__(self, param1, param2, etc):

        super(LRA, self).__init__()

        self.param1=param1

        self.param2=param2

        etc for all parameters

        # write any initialization code you need here

    def on_epoch_end(self, epoch, logs=None):  # method runs on the end of each epoch

        v_loss=logs.get('val_loss')  # example of getting log data at end of epoch the validation loss for this epoch

        acc=logs.get('accuracy') # another example of getting log data 

        LRA.best_weights=model.get_weights() # example of setting class variable value

        print(f'Hello epoch {epoch} has just ended') # print a message at the end of every epoch

        lr=float(tf.keras.backend.get_value(self.model.optimizer.lr)) # get the current learning rate

        if v_loss > self.param1:

           new_lr=lr * self.param2

           tf.keras.backend.set_value(model.optimizer.lr, new_lr) # set the learning rate in the optimizer

        # write whatever code you need

Answer 3

我建议您创建自己的回调。下面我添加了一个监控准确性和损失的解决方案。您可以将 acc 替换为您自己的指标：

class CustomCallback(keras.callbacks.Callback):

    acc = {}

    loss = {}

    best_weights = None

    

    def __init__(self, patience=None):

        super(CustomCallback, self).__init__()

        self.patience = patience

    

    def on_epoch_end(self, epoch, logs=None):

        epoch += 1

        self.loss[epoch] = logs['loss']

        self.acc[epoch] = logs['accuracy']

    

        if self.patience and epoch > self.patience:

            # best weight if the current loss is less than epoch-patience loss. Simiarly for acc but when larger

            if self.loss[epoch] < self.loss[epoch-self.patience] and self.acc[epoch] > self.acc[epoch-self.patience]:

                self.best_weights = self.model.get_weights()

            else:

                # to stop training

                self.model.stop_training = True

                # Load the best weights

                self.model.set_weights(self.best_weights)

        else:

            # best weight are the current weights

            self.best_weights = self.model.get_weights()

请记住，如果您想控制受监控数量的最小变化（又名 min_delta），您必须将其集成到代码中。

Answer 4

此时，创建自定义循环并仅使用 if 语句会更简单。例如：

def main(epochs=50):

    for epoch in range(epochs):

        fit(epoch)

        if test_acc.result() > .8 and topk_acc.result() > .9:

            print(f'\nEarly stopping. Test acc is above 80% and TopK acc is above 90%.')

            break

if __name__ == '__main__':

    main(epochs=100)

这是使用此方法的简单自定义训练循环：

import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

import tensorflow_datasets as tfds

import tensorflow as tf

data, info = tfds.load('iris', split='train',

                       as_supervised=True,

                       shuffle_files=True,

                       with_info=True)

def preprocessing(inputs, targets):

    scaled = tf.divide(inputs, tf.reduce_max(inputs, axis=0))

    return scaled, targets

dataset = data.filter(lambda x, y: tf.less_equal(y, 2)).\

    map(preprocessing).\

    shuffle(info.splits['train'].num_examples)

train_dataset = dataset.take(120).batch(4)

test_dataset = dataset.skip(120).take(30).batch(4)

model = tf.keras.Sequential([

    tf.keras.layers.Dense(8, activation='relu'),

    tf.keras.layers.Dense(16, activation='relu'),

    tf.keras.layers.Dense(info.features['label'].num_classes, activation='softmax')

    ])

loss_object = tf.losses.SparseCategoricalCrossentropy(from_logits=True)

train_loss = tf.metrics.Mean()

test_loss = tf.metrics.Mean()

train_acc = tf.metrics.SparseCategoricalAccuracy()

test_acc = tf.metrics.SparseCategoricalAccuracy()

topk_acc = tf.metrics.SparseTopKCategoricalAccuracy(k=2)

opt = tf.keras.optimizers.Adam(learning_rate=1e-3)

@tf.function

def train_step(inputs, labels):

    with tf.GradientTape() as tape:

        logits = model(inputs)

        loss = loss_object(labels, logits)

    gradients = tape.gradient(loss, model.trainable_variables)

    opt.apply_gradients(zip(gradients, model.trainable_variables))

    train_loss(loss)

    train_acc(labels, logits)

@tf.function

def test_step(inputs, labels):

    logits = model(inputs)

    loss = loss_object(labels, logits)

    test_loss.update_state(loss)

    test_acc.update_state(labels, logits)

    topk_acc.update_state(labels, logits)

def fit(epoch):

    template = 'Epoch {:>2} Train Loss {:.3f} Test Loss {:.3f} ' \

               'Train Acc {:.2f} Test Acc {:.2f} Test TopK Acc {:.2f} '

    train_loss.reset_states()

    test_loss.reset_states()

    train_acc.reset_states()

    test_acc.reset_states()

    topk_acc.reset_states()

    for X_train, y_train in train_dataset:

        train_step(X_train, y_train)

    for X_test, y_test in test_dataset:

        test_step(X_test, y_test)

    print(template.format(

        epoch + 1,

        train_loss.result(),

        test_loss.result(),

        train_acc.result(),

        test_acc.result(),

        topk_acc.result()

    ))

def main(epochs=50):

    for epoch in range(epochs):

        fit(epoch)

        if test_acc.result() > .8 and topk_acc.result() > .9:

            print(f'\nEarly stopping. Test acc is above 80% and TopK acc is above 90%.')

            break

if __name__ == '__main__':

    main(epochs=100)