How to use Keras + tensorflow datasets and custom callbacks with validation datasets

[kmkolasinski]

Hello,
recently I've started to work with Keras together with tensorflow datasets.
The API I use is following:

• I create net which inputs are tf.Tensors generated by iterator
• Then I compile the model, by providing target tensors: target_tensors

model.compile(Adam(lr=lr), loss=losses, metrics=metrics, target_tensors=target_tensors)

• Then fit the model by calling model.fit(...)

However, it's not clear to me how could I add validation datasets into training loop, like it was done in previous API. I understand that datasets are quite new in Keras, but maybe there are some examples which could be useful.
Thanks in advance.

[zotroneneis]

I'm not sure whether I understand your question correctly. There are two easy ways in which you can include validation data in the training process:

1. Tell the model.fit() function to use part of your training data for validation, e.g.
   model.fit(x, y, validation_split=0.3)

2. Tell the model.fit() function which validation data to use
   model.fit(x, y, validation_data=(x_val, y_val))

[kmkolasinski]

Hi, thanks for answer, however my question was about how to add properly evaluation on validation data when working with tensorflow datasets API, which reads data directly from tfrecords. I mean this. Normally, I read data from e.g. *.tfrecords files and create model which inputs tensors take data from tf dataset. So basically for training data I run:

...
iterator = dataset.make_initializable_iterator()
training_filenames = ["/var/data/file1.tfrecord", "/var/data/file2.tfrecord"]
sess.run(iterator.initializer, feed_dict={filenames: training_filenames})

The workaround for my problem would be to write special callback, which evaluates model after each epoch on other dataset which reads from validation_filenames and then switch back to training_filenames. This works, but maybe there is a simpler one.