如何在 Keras 的每个时期后禁用打印报告？

Question

After each epoch I have printout like below:

Train on 102 samples, validate on 26 samples
Epoch 1/1
Epoch 00000: val_acc did not improve
102/102 [==============================] - 3s - loss: 0.4934 - acc: 0.8997 - val_loss: 0.4984 - val_acc: 0.9231

I am not using built-in epochs, so I would like to disable these printouts and print something myself.

How to do that?

I am using tensorflow backend if it matters.

Answer 1

将verbose=0设置为模型的拟合方法。

Answer 2

First of all, you should set verbose=0 to the fit method for having a silent environment then we need to have a callback for controlling it. I always use the following code for showing every 10 epochs with its loss.

import tensorflow as tf

class Callback(tf.keras.callbacks.Callback):
    SHOW_NUMBER = 10
    counter = 0
    epoch = 0

    def on_epoch_begin(self, epoch, logs=None):
        self.epoch = epoch

    def on_train_batch_end(self, batch, logs=None):
        if self.counter == self.SHOW_NUMBER or self.epoch == 1:
            print('Epoch: ' + str(self.epoch) + ' loss: ' + str(logs['loss']))
            if self.epoch > 1:
                self.counter = 0
        self.counter += 1

model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, use_multiprocessing=True, callbacks=[Callback()], verbose=0)

Notice that if you increase SHOW_NUMBER , the processing will be finished sooner, so if your epochs are big, you should set it more.

Answer 3

In case anyone doesn't want to implement the progress bar from scratch:

import tensorflow as tf

class SelectiveProgbarLogger(tf.keras.callbacks.ProgbarLogger):
    def __init__(self, verbose, epoch_interval, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.default_verbose = verbose
        self.epoch_interval = epoch_interval
    
    def on_epoch_begin(self, epoch, *args, **kwargs):
        self.verbose = (
            0 
                if epoch % self.epoch_interval != 0 
                else self.default_verbose
        )
        super().on_epoch_begin(epoch, *args, **kwargs)

model.fit(
    x=..., y=..., 
    batch_size=..., epochs=..., 
    # Set the verbosity level here! This is important.
    callbacks=[SelectiveProgbarLogger(verbose = 1)], 
    # This disables the default `ProgbarLogger`
    verbose=0
)