Stop Training in Keras when Accuracy is already 1.0

Question

How will I stop Keras Training when the accuracy already reached 1.0? I tried monitoring loss value, but I haven't tried stopping the training when the accuracy is already 1.

I tried the code below with no luck:

stopping_criterions =[
    EarlyStopping(monitor='loss', min_delta=0, patience = 1000),
    EarlyStopping(monitor='acc', base_line=1.0, patience =0)

]

model.summary()
model.compile(Adam(), loss='binary_crossentropy', metrics=['accuracy']) 
model.fit(scaled_train_samples, train_labels, batch_size=1000, epochs=1000000, callbacks=[stopping_criterions], shuffle = True, verbose=2)

UPDATE:

The training immediately stops at first epoch, even if the accuracy is still not 1.0.

Please help.

Answer 1

Update: tested in keras 2.4.3 (Dec.2020)

I don't know why EarlyStopping does not work in this case. Instead, I defined a custom callback that stops training when acc (or val_acc) reaches a specified baseline:

from keras.callbacks import Callback

class TerminateOnBaseline(Callback):
    """Callback that terminates training when either acc or val_acc reaches a specified baseline
    """
    def __init__(self, monitor='accuracy', baseline=0.9):
        super(TerminateOnBaseline, self).__init__()
        self.monitor = monitor
        self.baseline = baseline

    def on_epoch_end(self, epoch, logs=None):
        logs = logs or {}
        acc = logs.get(self.monitor)
        if acc is not None:
            if acc >= self.baseline:
                print('Epoch %d: Reached baseline, terminating training' % (epoch))
                self.model.stop_training = True

You can use it like this:

callbacks = [TerminateOnBaseline(monitor='accuracy', baseline=0.8)]
callbacks = [TerminateOnBaseline(monitor='val_accuracy', baseline=0.95)]

---

Note: This solution does not work.

If you want to stop training when the training (or validation) accuracy exactly reaches 100%, then use EarlyStopping callback and set the baseline argument to 1.0 and patience to zero:

EarlyStopping(monitor='acc', baseline=1.0, patience=0)  # use 'val_acc' instead to monitor validation accuarcy