CODEWITHSUNDEEP
pythonforktensorflow
MBZ
MBZ

Reputation: 27592

forking a python process after loading TensorFlow

tf.Session() is not fork safe which means that the behavior of the system after forking a process while TensorFlow is loaded into the memory is unknown.

is there any work around for sharing multiple devices (on a single machine), between multiple processes?

Upvotes: 5

Views: 2152

Answers (1)

mrry
mrry

Reputation: 126184

The standard way to share a TensorFlow runtime between multiple processes is to use the distributed TensorFlow support, which also works on a single machine.

In one process, you can start a server by running the following code:

import tensorflow as tf
server = tf.train.Server.create_local_server()
print server.target  # for other processes to connect
server.join()

This process will own all of the devices on the machine, by default.

In the other processes, you can create tf.Session objects that connect to the server:

sess = tf.Session("grpc://localhost:...")  # Use value of `server.target`.

These sessions can be used just like in-process sessions.

Upvotes: 6

Related Questions