I have the following training setup from tensorflow v1’s estimators:

config = tf.estimator.RunConfig(
    tf_random_seed=seed,
    log_step_count_steps=log_every_n_iter,
    save_checkpoints_steps=save_checkpoints_steps,
    session_config=gpu_config,
)

I followed some guides, like:

• Migrate from estimator to Keras API
• Migrate checkpoint saving
• Migrate canned estimators

And I tried to migrate some of the config options to keras, resulting in the following code:

tf.compat.v1.keras.utils.set_random_seed(seed)
model_checkpoint_callback = tf.compat.v1.keras.callbacks.ModelCheckpoint(
    filepath=model_dir,
    save_freq=save_checkpoints_steps,
    save_best_only=False,
)

Note: I still don’t know if it works as I still have to migrate the models and add the callback.

The questions are: Is set_random_seed the correct way of doing it? And what about log_every_n_iter? How can I replicate that behaviour.