为什么Keras Early Stopping功能不停止训练，虽然监测到的值在增加？

Question

我正在尝试为回归问题训练神经网络，并且我实现了 Keras 早期停止功能以避免过度拟合。

现在，当我监视“val_loss”时，提前停止功能几乎直接停止程序，结果是一个无用的神经网络，但是当我监视“val_mse”时，训练继续进行而不会停止，尽管我可以看到“val_mse”在整个过程中不断增加训练，我设置耐心= 0。

我似乎误解了早期停止回调，因为我认为它确实会监控值并在值再次开始增加时立即停止训练。

np.random.seed(7)

#Define Input
tf_features_64 = np.load("IN_2.npy")
tf_labels_64 = np.load("OUT_2.npy")
tf_features_32 = tf_features_64.astype(np.float32)
tf_labels_32 = tf_labels_64.astype(np.float32)

X = tf_features_32
Y = tf_labels_32[0:10680, 4:8]
#Define Callback
tbCallBack = TensorBoard(log_dir='./Graph{}', histogram_freq=0, write_graph=True, write_images=True) #TensorBoard Monitoring
esCallback = EarlyStopping(monitor='val_mse',
                           min_delta=0,
                           patience=0,
                           verbose=1,
                           mode='min')

#create Layers
visible = Input(shape=(33,))
x = Dropout(.1)(visible)
#x = Dense(63)(x)
#x = Dropout(.4)(x)
output = Dense(4)(x)  

Optimizer = optimizers.Adam(lr=0.001
                            #amsgrad = True)

model = Model(inputs=visible, outputs = output)
model.compile(optimizer=Optimizer,
              loss=['mse'],
              metrics=['mae', 'mse']
              )
model.fit(X, Y, epochs=8000, batch_size=20, shuffle=True, validation_split=0.35, callbacks=[tbCallBack, esCallback])

例如，我得到以下输出，我可以清楚地看到，val_mse 在各个时期内增加。

  20/6942 [..............................] - ETA: 0s - loss: 0.0022 - mean_absolute_error: 0.0373 - mean_squared_error: 0.0022
1620/6942 [======>.......................] - ETA: 0s - loss: 0.0011 - mean_absolute_error: 0.0251 - mean_squared_error: 0.0011
3260/6942 [=============>................] - ETA: 0s - loss: 0.0015 - mean_absolute_error: 0.0290 - mean_squared_error: 0.0015
4900/6942 [====================>.........] - ETA: 0s - loss: 0.0017 - mean_absolute_error: 0.0301 - mean_squared_error: 0.0017
6500/6942 [===========================>..] - ETA: 0s - loss: 0.0016 - mean_absolute_error: 0.0301 - mean_squared_error: 0.0016
6942/6942 [==============================] - 0s 37us/step - loss: 0.0016 - mean_absolute_error: 0.0294 - mean_squared_error: 0.0016 - val_loss: 0.0011 - val_mean_absolute_error: 0.0240 - **val_mean_squared_error: 0.0011**
**Epoch 334/8000**

  20/6942 [..............................] - ETA: 0s - loss: 0.0025 - mean_absolute_error: 0.0367 - mean_squared_error: 0.0025
1620/6942 [======>.......................] - ETA: 0s - loss: 0.0012 - mean_absolute_error: 0.0257 - mean_squared_error: 0.0012
3260/6942 [=============>................] - ETA: 0s - loss: 0.0014 - mean_absolute_error: 0.0274 - mean_squared_error: 0.0014
4860/6942 [====================>.........] - ETA: 0s - loss: 0.0014 - mean_absolute_error: 0.0268 - mean_squared_error: 0.0014
6400/6942 [==========================>...] - ETA: 0s - loss: 0.0012 - mean_absolute_error: 0.0254 - mean_squared_error: 0.0012
6942/6942 [==============================] - 0s 39us/step - loss: 0.0012 - mean_absolute_error: 0.0249 - mean_squared_error: 0.0012 - val_loss: 0.0032 - val_mean_absolute_error: 0.0393 - **val_mean_squared_error: 0.0032**
**Epoch 335/8000**

  20/6942 [..............................] - ETA: 0s - loss: 9.5175e-04 - mean_absolute_error: 0.0243 - mean_squared_error: 9.5175e-04
1620/6942 [======>.......................] - ETA: 0s - loss: 0.0017 - mean_absolute_error: 0.0312 - mean_squared_error: 0.0017        
3260/6942 [=============>................] - ETA: 0s - loss: 0.0013 - mean_absolute_error: 0.0271 - mean_squared_error: 0.0013
4860/6942 [====================>.........] - ETA: 0s - loss: 0.0014 - mean_absolute_error: 0.0277 - mean_squared_error: 0.0014
6460/6942 [==========================>...] - ETA: 0s - loss: 0.0013 - mean_absolute_error: 0.0266 - mean_squared_error: 0.0013
6942/6942 [==============================] - 0s 38us/step - loss: 0.0013 - mean_absolute_error: 0.0268 - mean_squared_error: 0.0013 - val_loss: 0.0046 - val_mean_absolute_error: 0.0491 - **val_mean_squared_error: 0.0046**
**Epoch 336/8000**

Answer 1

val_mse您的代码中没有调用指标，您的回调正在监控错误的指标。有，val_mean_squared_error但这与val_mse.

您应该将要监控的指标从 更改val_mse为val_mean_squared_error，它应该可以工作。