gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2026-02-01 13:25:45 +01:00
Code Issues Projects Releases Wiki Activity

Parallel agents fixes (#95)

Browse Source 
* Parallel agents related bug fixes: checkpoint restore, tensorboard integration.
Adding narrow networks support.
Reference code for unlimited number of checkpoints
This commit is contained in:
Itai Caspi
2018-05-24 14:24:19 +03:00
committed by GitHub
parent 6c0b59b4de
commit d302168c8c
10 changed files with 75 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
agents/agent.py
View File

@@ -550,9 +550,10 @@ class Agent(object):
if current_snapshot_period > model_snapshots_periods_passed:
model_snapshots_periods_passed = current_snapshot_period
self.save_model(model_snapshots_periods_passed)
to_pickle(self.running_observation_stats,
os.path.join(self.tp.save_model_dir,
"running_stats.p".format(model_snapshots_periods_passed)))
if self.running_observation_stats is not None:
to_pickle(self.running_observation_stats,
os.path.join(self.tp.save_model_dir,
"running_stats.p".format(model_snapshots_periods_passed)))
# play and record in replay buffer
if self.tp.agent.collect_new_data:

2
agents/clipped_ppo_agent.py
View File

@@ -69,7 +69,7 @@ class ClippedPPOAgent(ActorCriticAgent):
screen.warning("WARNING: The requested policy gradient rescaler is not available")
# standardize
advantages = (advantages - np.mean(advantages)) / np.std(advantages)
advantages = (advantages - np.mean(advantages)) / (np.std(advantages) + 1e-8)
for transition, advantage, value_target in zip(batch, advantages, value_targets):
transition.info['advantage'] = advantage