Skip to content

[bug-fix] Fix memory leak when using LSTMs #5048

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Mar 8, 2021
Merged

[bug-fix] Fix memory leak when using LSTMs #5048

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f66dbe6
Detach memory before storing
Mar 6, 2021
b137059
Add test
Mar 6, 2021
54f45ee
Evaluate with no_grad
Mar 8, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 14 additions & 11 deletions ml-agents/mlagents/trainers/optimizer/torch_optimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -166,18 +166,21 @@ def get_trajectory_value_estimates(

# If we're using LSTM, we want to get all the intermediate memories.
all_next_memories: Optional[AgentBufferField] = None
if self.policy.use_recurrent:
(
value_estimates,
all_next_memories,
next_memory,
) = self._evaluate_by_sequence(current_obs, memory)
else:
value_estimates, next_memory = self.critic.critic_pass(
current_obs, memory, sequence_length=batch.num_experiences
)

# Store the memory for the next trajectory
# To prevent memory leak and improve performance, evaluate with no_grad.
with torch.no_grad():
if self.policy.use_recurrent:
(
value_estimates,
all_next_memories,
next_memory,
) = self._evaluate_by_sequence(current_obs, memory)
else:
value_estimates, next_memory = self.critic.critic_pass(
current_obs, memory, sequence_length=batch.num_experiences
)

# Store the memory for the next trajectory. This should NOT have a gradient.
self.critic_memory_dict[agent_id] = next_memory

next_value_estimate, _ = self.critic.critic_pass(
Expand Down
5 changes: 5 additions & 0 deletions ml-agents/mlagents/trainers/tests/torch/test_ppo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -207,6 +207,11 @@ def test_ppo_get_value_estimates(dummy_config, rnn, visual, discrete):
run_out, final_value_out, all_memories = optimizer.get_trajectory_value_estimates(
trajectory.to_agentbuffer(), trajectory.next_obs, done=False
)
if rnn:
# Check that memories don't have a Torch gradient
for mem in optimizer.critic_memory_dict.values():
assert not mem.requires_grad

for key, val in run_out.items():
assert type(key) is str
assert len(val) == 15
Expand Down