tfset 1.1

Set Tensor values during training in Tensorflow.

Change the hyper-parameters of your Tensorflow training session on the fly. The package allows you to schedule events that change the values of arbitrary Tensors with a simple command.

Requirements

• Python >= 3

• tensorflow >= 1.0

Set Up

Install the package with pip:

pip install tfset

Or clone and install from github:

git clone https://github.com/ondrejba/tfset.git
cd tfset
python setup.py install

Usage

tfset DEMO

Check MNIST_demo.ipynb for a demostration of the usage of tfset in a simple training script.

Server

Import tfset server.

import tfset.server as server

Create Tensors for your hyper-parameters.

learning_rate = tf.get_variable("learning_rate", initializer=tf.constant(0.1, dtype=tf.float32))
dropout_prob = tf.get_variable("dropout_prob", initializer=tf.constant(0.9, dtype=tf.float32))

Create and start a Session Server.

# "session" is a Tensorflow session
s, thread = server.run_server([learning_rate, dropout_prob], session)

Periodically check for events.

# "step" is the global step of your training procedure
s.check_events(step)

Stop the server.

s.shutdown()
thread.join(timeout=10)

Client

Get status.

tfset -s

Add an event (this event sets the learning rate to 0.01 at iteration 10000).

tfset -a -n learning_rate:0 -i 10000 --value 0.01

Remove an event (with index 0 in this case).

tfset -r -e 0

Events

tfset schedules hyper-parameter changes based on events. An event contains the following information:

• iteration: when to execute the event

• Tensor name: which Tensor to change

• value: value to set the Tensor to

The reason for the use of events is that you might want to schedule hyper-parameter changes in the future (e.g. lower learning rate to 10e-3 at 800k iteration). If two events targeting the same Tensor are scheduled at the same iteration, the one that was scheduled later is going to be executed.