Python 验证丢失满足特定标准时提前停止

我正在Keras中训练一个神经网络模型。我希望监控验证丢失，并在达到特定条件时停止培训

我知道，在给定的

耐心

轮数下，如果训练没有改善，我可以使用

提前停止

停止训练

我想做些不同的事情。当

val_loss

在

n

轮后高于某个值时，我想停止训练，比如

x

为了清楚起见，我们假设

0.5中的
x
，
n
是
50
。仅当
epoch
数字大于
50
且
valu loss
大于
0.5
时，我才想停止模型的训练

如何在Keras中执行此操作？
您可以通过继承Keras
早期停止
回调并使用自己的逻辑覆盖它来定义自己的回调：

from keras.callbacks import EarlyStopping # use as base class

class MyCallBack(EarlyStopping):
    def __init__(self, threshold, min_epochs, **kwargs):
        super(MyCallBack, self).__init__(**kwargs)
        self.threshold = threshold # threshold for validation loss
        self.min_epochs = min_epochs # min number of epochs to run

    def on_epoch_end(self, epoch, logs=None):
        current = logs.get(self.monitor)
        if current is None:
            warnings.warn(
                'Early stopping conditioned on metric `%s` '
                'which is not available. Available metrics are: %s' %
                (self.monitor, ','.join(list(logs.keys()))), RuntimeWarning
            )
            return

        # implement your own logic here
        if (epoch >= self.min_epochs) & (current >= self.threshold):
            self.stopped_epoch = epoch
            self.model.stop_training = True

举例说明它应该起作用：

from keras.layers import Input, Dense
from keras.models import Model
import numpy as np

# Generate some random data
features = np.random.rand(100, 5)
labels = np.random.rand(100, 1)

validation_feat = np.random.rand(100, 5)
validation_labels = np.random.rand(100, 1)

# Define a simple model
input_layer = Input((5, ))
dense_layer = Dense(10)(input_layer)
output_layer = Dense(1)(dense_layer)
model = Model(inputs=input_layer, outputs=output_layer)
model.compile(loss='mse', optimizer='sgd')

# Fit with custom callback
callbacks = [MyCallBack(threshold=0.001, min_epochs=10, verbose=1)] 
model.fit(features, labels, validation_data=(validation_feat, validation_labels), callbacks=callbacks, epochs=100)   

不错。我试着自己实现这个。我犯了一些错误。我会参考这个