python - 当验证损失满足某些标准时提前停止

Question

我正在 Keras 中训练神经网络模型。我想监控验证损失并在达到特定条件时停止训练。

我知道我可以使用 EarlyStopping 在给定的 patience 轮次训练没有改善时停止训练。

我想要一些不同的东西。我想在 n 轮后 val_loss 超过 x 值时停止训练。

为了清楚起见，假设 0.5 中的 x 和 n 是 50。我只想在 epoch 数大于 50 且 val_loss 高于 0.5 时停止模型训练。

我如何在 Keras 中执行此操作？

Answer 1

您可以通过继承 Keras EarlyStopping 回调并使用您自己的逻辑覆盖它来定义您自己的回调:

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

说明它应该工作的小例子:

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)