无法记录使用Keras构建但使用Tensorflow的Estimator训练的模型的损失

Question

I created a neural model with Keras and want to train it with the Tensorflow Estimator API for Keras models.

Further I want to simply log the loss value with a LoggingTensorHook. But I get the following error:

ValueError: Passed Tensor("loss/mul:0", shape=(), dtype=float32) should have graph attribute that is equal to current graph <tensorflow.python.framework.ops.Graph object at 0x11fca7358>.

I have a rough idea of what is going on. (Estimator create a new graph or something), but can't solve this issue by myself.

---

Now some code:

model = create_keras_model(shapes)

adam = tf.keras.optimizers.Adam(lr=conf.lr)
model.compile(loss='categorical_crossentropy',
              optimizer=adam, metrics=['acc'])

estimator = tf.keras.estimator.model_to_estimator(
    keras_model=model, model_dir=model_dir,config=run_config)


### I want to log this tensor
loss_to_log = model.total_loss
log_hook = tf.train.LoggingTensorHook(
    {'loss': loss_to_log}, every_n_iter=10, at_end=True)

train_spec = tf.estimator.TrainSpec(
    input_fn=...,
    max_steps=..., hooks=[log_hook])

tf.estimator.train_and_evaluate(estimator, train_spec, ...)

Answer 1

Try add codes:

import tensorflow as tf
tf.logging.set_verbosity(tf.logging.INFO)

and modify your codes

{'loss': loss_to_log}

to

{'loss': loss_to_log.name}