I am following the tutorial on this link:
In the task.py file (https://github.com/GoogleCloudPlatform/tensorflow-lifetime-value/blob/0f8c16ea70a2e7da370965e23e9e2154978364fa/clv_mle/trainer/task.py)
of the used github Repo are the default hyperparameters given. They are using the parameter ‘train_size’ = 100’000, but the actual length of the trainingset is only 883.
It is used for the following:
checkpoint_steps = int((train_size/batch_size) * (num_epochs/num_eval))
train_steps = (train_size/batch_size) * num_epochs
which in turn are used like this:
lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
initial_learning_rate = 0.096505,
decay_steps = checkpoint_steps,
decay_rate = 0.7,
staircase=True)
…and this:
train_spec = tf.estimator.TrainSpec(
input_fn = read_train,
max_steps = train_steps)
I don’t get it. Why is the given parameter so huge?
(p.s.: I have some more issues, in case anyone is interested in helping out a lost student.)