tf.contrib.training.RandomStrategy

Returns a random PS task for op placement.

tf.contrib.training.RandomStrategy(
    num_ps_tasks, seed=0
)

This may perform better than the default round-robin placement if you have a large number of variables. Depending on your architecture and number of parameter servers, round-robin can lead to situations where all of one type of variable is placed on a single PS task, which may lead to contention issues.

This strategy uses a hash function on the name of each op for deterministic placement.

Methods

__call__

View source

__call__(
    op
)

Chooses a ps task index for the given Operation.