ScopedGraphManager

Applicability

Description

Unloads the variable initialization graph in one go and releases the memory held by constant nodes in the graph.

Prototype

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

class ScopedGraphManager(object):
    def __enter__(self):
        self.start()
        return self

    def __exit__(self, exe_type, exe_val, exc_tb):
        self.stop()

    def start(self):
        tf_adapter.EnableControl()

    def stop(self):
        tf_adapter.Clear()

Parameters

None

Returns

None

Restrictions

• The ScopedGraphManager class needs to be called using the with statement. Once the scope is exceeded, the variable initialization graph executed in the scope is automatically unloaded, and the occupied constant memory is released.
• Only a variable initialization graph can be placed in the scope of ScopedGraphManager. If a non-variable initialization graph is placed in this scope, it may interfere with the normal operation of training functions.
• This API is applicable only to the scenario where variables are initialized at a time in the main process. It does not support scenarios where variables are initialized in multiple threads, sessions, or times.

Example

1
2
3
4
5
6

import tensorflow as tf
from npu_bridge.npu_init import *

with scoped_graph_manager.ScopedGraphManager():
    # Initialize variables.
    sess.run(tf.global_variables_initializer())