gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2026-02-21 01:05:50 +01:00
Code Issues Projects Releases Wiki Activity

Avoid Memory Leak in Rollout worker

Browse Source 
ISSUE: When we restore checkpoints, we create new nodes in the
Tensorflow graph. This happens when we assign new value (op node) to
RefVariable in GlobalVariableSaver. With every restore the size of TF
graph increases as new nodes are created and old unused nodes are not
removed from the graph. This causes the memory leak in
restore_checkpoint codepath.

FIX: We reset the Tensorflow graph and recreate the Global, Online and
Target networks on every restore. This ensures that the old unused nodes
in TF graph is dropped.
This commit is contained in:
Gourav Roy
2018-12-25 20:50:34 -08:00
parent 02f2db1264
commit c694766fad
3 changed files with 17 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
rl_coach/agents/agent.py
View File

@@ -953,6 +953,12 @@ class Agent(AgentInterface):
self.input_filter.restore_state_from_checkpoint(checkpoint_dir, checkpoint_prefix)
self.pre_network_filter.restore_state_from_checkpoint(checkpoint_dir, checkpoint_prefix)
if self.ap.task_parameters.framework_type == Frameworks.tensorflow:
import tensorflow as tf
tf.reset_default_graph()
# Recreate all the networks of the agent
self.networks = self.create_networks()
# no output filters currently have an internal state to restore
# self.output_filter.restore_state_from_checkpoint(checkpoint_dir)