问题:

With Tensorflow it is possible to monitor quantities during training, using tf.summary.

Is it possible to do the same using Keras ? Could you include an example by modifying the code at https://github.com/fchollet/keras/blob/master/examples/variational_autoencoder.py and monitoring the KL loss (defined at line 53)

Thank you in advance !

回答1:

Have you tried the TensorBoard callback? [1]

tensorboard = keras.callbacks.TensorBoard(log_dir='./logs',
                 histogram_freq=1, 
                 write_graph=True, 
                 write_images=False)
vae.fit(x_train,
        shuffle=True,
        epochs=epochs,
        batch_size=batch_size,
        validation_data=(x_test, x_test),
        callbacks=[tensorboard])

Then run:

tensorboard --logdir=./logs

You could write a modified version of the callback to handle the specific items you are interested in.

[1] https://keras.io/callbacks/#tensorboard

回答2:

Actually a workaround consists in adding the quantities to monitor as metrics when compiling the model.

For instance, I wanted to monitor the KL divergence (in the context of variational auto encoders), so I wrote this:

def kl_loss(y_true, y_pred):
    kl_loss = - 0.5 * K.sum(1 + K.log(z_var_0+1e-8) - K.square(z_mean_0) - z_var_0, axis=-1)
    return kl_loss

vae.compile(optimizer='rmsprop', loss=vae_loss, metrics=['accuracy', kl_loss])

And it does what I need