问题描述
为了训练我的神经网络模型，我使用 Keras 的 EarlyStopping 回调来最小化训练时间（通过 talos.utils.early_stopper 包装器）：
history = model.fit(x=X_train,y=y_train,validation_data=(X_val,y_val),batch_size=params["batch"],epochs=params["epoch"],callbacks=[talos.utils.early_stopper(epochs=params["epoch"],mode='strict',min_delta=0.001)],verbose=1)
但是，我注意到它的工作相当不一致：
附件 A
Epoch 1/42
160/160 [==============================] - 19s 73ms/step - loss: 116.8279 - accuracy: 0.3124 - val_loss: 0.5561 - val_accuracy: 0.3708
Epoch 2/42
160/160 [==============================] - 6s 36ms/step - loss: 0.5676 - accuracy: 0.3440 - val_loss: 0.5564 - val_accuracy: 0.3708
Epoch 3/42
160/160 [==============================] - 6s 35ms/step - loss: 0.5720 - accuracy: 0.3337 - val_loss: 0.5573 - val_accuracy: 0.3708
-> TRAIN STOPS
图表 B
Epoch 14/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5664 - accuracy: 0.3501 - val_loss: 0.5581 - val_accuracy: 0.3708
Epoch 15/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5684 - accuracy: 0.3414 - val_loss: 0.5575 - val_accuracy: 0.3708
Epoch 16/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5699 - accuracy: 0.3348 - val_loss: 0.5570 - val_accuracy: 0.3708
Epoch 17/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5686 - accuracy: 0.3415 - val_loss: 0.5567 - val_accuracy: 0.3708
Epoch 18/42
160/160 [==============================] - 6s 38ms/step - loss: 0.5666 - accuracy: 0.3457 - val_loss: 0.5566 - val_accuracy: 0.3708
Epoch 19/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5694 - accuracy: 0.3367 - val_loss: 0.5563 - val_accuracy: 0.3708
Epoch 20/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5671 - accuracy: 0.3418 - val_loss: 0.5562 - val_accuracy: 0.3708
Epoch 21/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5654 - accuracy: 0.3472 - val_loss: 0.5561 - val_accuracy: 0.3708
Epoch 22/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5673 - accuracy: 0.3416 - val_loss: 0.5560 - val_accuracy: 0.3708
Epoch 23/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5637 - accuracy: 0.3542 - val_loss: 0.5560 - val_accuracy: 0.3708
Epoch 24/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5670 - accuracy: 0.3417 - val_loss: 0.5560 - val_accuracy: 0.3708
Epoch 25/42
160/160 [==============================] - 6s 38ms/step - loss: 0.5652 - accuracy: 0.3495 - val_loss: 0.5560 - val_accuracy: 0.3708
Epoch 26/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5663 - accuracy: 0.3454 - val_loss: 0.5560 - val_accuracy: 0.3708
Epoch 27/42
160/160 [==============================] - 6s 38ms/step - loss: 0.5679 - accuracy: 0.3384 - val_loss: 0.5560 - val_accuracy: 0.3708
Epoch 28/42
160/160 [==============================] - 6s 38ms/step - loss: 0.5639 - accuracy: 0.3505 - val_loss: 0.5560 - val_accuracy: 0.3708
Epoch 29/42
160/160 [==============================] - 7s 42ms/step - loss: 0.5636 - accuracy: 0.3515 - val_loss: 0.5559 - val_accuracy: 0.3708
Epoch 30/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5680 - accuracy: 0.3399 - val_loss: 0.5559 - val_accuracy: 0.3708
Epoch 31/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5696 - accuracy: 0.3338 - val_loss: 0.5559 - val_accuracy: 0.3708
Epoch 32/42
160/160 [==============================] - 6s 39ms/step - loss: 0.5705 - accuracy: 0.3321 - val_loss: 0.5559 - val_accuracy: 0.3708
Epoch 33/42
160/160 [==============================] - 6s 40ms/step - loss: 0.5724 - accuracy: 0.3273 - val_loss: 0.5559 - val_accuracy: 0.3708
-> TRAIN STOPS
为什么它在 Exhibit B 上没有提前停止，即使很明显 val_loss 上面的 min_delta 没有改进？我看过 talos source ，在我看来 early_stopper 只是回调的包装器，一切看起来都很好。我还注意到，这只会在每次 val_loss 围绕 0.5559 - 0.5560 浮动时才会发生。
FWIW 我正在 Colab 上运行它，使用 TPU。
谢谢！
解决方法
出于某种原因，将 monitor 从 val_loss 更改为 val_accuracy (EarlyStopping(monitor="val_accuracy",min_delta=0.01,patience=2,verbose=1,mode='auto') 似乎提供了更一致的回调。