学习率动态修改

#    step
　　　　　lr_scheduler = tf.keras.optimizers.schedules.ExponentialDecay(
            initial_learning_rate=lr_rate,
            decay_steps=10,
            decay_rate=0.96)
　　　　　opt = tf.keras.optimizers.Adamax(lr=lr_scheduler, beta_1=0.9, beta_2=0.999, epsilon=1e-09)

#   epoch修改学习率
        def scheduler(epoch):
            if epoch < 5:
                return lr_rate
            else:
                lr = tf.maximum(lr_rate * tf.math.exp(0.1 * (5 - epoch)),1e-4)
            return lr.numpy()
        reduce_lr = tf.keras.callbacks.LearningRateScheduler(scheduler)
        reduce_lr2 = tf.keras.callbacks.ReduceLROnPlateau( monitor='val_dice_coef',
                                                                   factor=0.5,
                                                                   patience=3,
                                                                   verbose=0,
                                                                   mode='max',
                                                                   min_delta=1e-4,
                                                                   cooldown=0,
                                                                   min_lr=1e-4,)

        s_model.fit(train_db, epochs=epochs, validation_data=test_db, callbacks=[reduce_lr2,early_stoping, history2])

搜索

复制