TensorFlow - 当损失达到定义值时停止训练

Question

I used the first example here as an example of network.

How to stop the training when the loss reach a fixed value ?

So, for example, I would like to fix a maximum of 3000 epochs and the training will stop when the loss will be under 0.2.

I read this topic but it is not the solution I found.

I would want to stop the training when the loss reach a value, not when there is no improvement like with this function proposed in the precedent topic.

Here is the code:

import keras
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation
from keras.optimizers import SGD

# Generate dummy data
import numpy as np
x_train = np.random.random((1000, 20))
y_train = keras.utils.to_categorical(np.random.randint(10, size=(1000, 1)), num_classes=10)
x_test = np.random.random((100, 20))
y_test = keras.utils.to_categorical(np.random.randint(10, size=(100, 1)), num_classes=10)

model = Sequential()
# Dense(64) is a fully-connected layer with 64 hidden units.
# in the first layer, you must specify the expected input data shape:
# here, 20-dimensional vectors.
model.add(Dense(64, activation='relu', input_dim=20))
model.add(Dropout(0.5))
model.add(Dense(64, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(10, activation='softmax'))

sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy',
              optimizer=sgd,
              metrics=['accuracy'])

model.fit(x_train, y_train,
          epochs=3000,
          batch_size=128)
score = model.evaluate(x_test, y_test, batch_size=128)  

Answer 1

You can use some method like this if you would switch to TensorFlow 2.0:

class haltCallback(tf.keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs={}):
    if(logs.get('loss') <= 0.05):
        print("\n\n\nReached 0.05 loss value so cancelling training!\n\n\n")
        self.model.stop_training = True

You just need to create a callback like that and then add that callback to your model.fit so it becomes something like this:

model.fit(x_train, y_train,
      epochs=3000,
      batch_size=128,
      callbacks=['trainingStopCallback'])

This way, the fitting should stop whenever it reaches down below 0.05 (or whatever value you put on while defining it).

Also, since it's been a long time you asked this question but it still has no actual answer for using it with TensorFlow 2.0, I updated your code snippet to TensorFlow 2.0 so everyone can now easily find and use it with their new projects.

import tensorflow as tf

# Generate dummy data
import numpy as np


x_train = np.random.random((1000, 20))
y_train = tf.keras.utils.to_categorical(
    np.random.randint(10, size=(1000, 1)), num_classes=10)
x_test = np.random.random((100, 20))
y_test = tf.keras.utils.to_categorical(
    np.random.randint(10, size=(100, 1)), num_classes=10)

model = tf.keras.models.Sequential()
# Dense(64) is a fully-connected layer with 64 hidden units.
# in the first layer, you must specify the expected input data shape:
# here, 20-dimensional vectors.
model.add(tf.keras.layers.Dense(64, activation='relu', input_dim=20))
model.add(tf.keras.layers.Dropout(0.5))
model.add(tf.keras.layers.Dense(64, activation='relu'))
model.add(tf.keras.layers.Dropout(0.5))
model.add(tf.keras.layers.Dense(10, activation='softmax'))


class haltCallback(tf.keras.callbacks.Callback):
    def on_epoch_end(self, epoch, logs={}):
        if(logs.get('loss') <= 0.05):
            print("\n\n\nReached 0.05 loss value so cancelling training!\n\n\n")
            self.model.stop_training = True


trainingStopCallback = haltCallback()

sgd = tf.keras.optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy',
              optimizer=sgd,
              metrics=['accuracy', 'loss'])

model.fit(x_train, y_train,
          epochs=3000,
          batch_size=128,
          callbacks=['trainingStopCallback'])
score = model.evaluate(x_test, y_test, batch_size=128)

Answer 2

此处的文档： EarlyStopping

keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=0, verbose=0, mode='auto', baseline=None)