Skip to content
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

[RLlib] Add metrics to buffers. #49822

Merged
merged 8 commits into from
Jan 20, 2025
Merged

[RLlib] Add metrics to buffers. #49822

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
f4e56ee
Defined metrics for replay buffers and added metrics logging to 'Epis…
simonsays1980 Jan 8, 2025
22d406f
WIP. Added counting metrics to 'EpisodeReplayBuffer' and a logging me…
simonsays1980 Jan 9, 2025
5eea74e
LINTER
simonsays1980 Jan 9, 2025
78bcb4f
Added metrics for adding and sampling in 'EpisodeReplayBuffer'. Furth…
simonsays1980 Jan 13, 2025
d440c93
Added further metrics to the 'EpisodeReplayBuffer' and added all metr…
simonsays1980 Jan 14, 2025
12a251b
Merged master and resolved conflicts.
simonsays1980 Jan 14, 2025
9f5ab29
Merge branch 'master' into offpolicy-add-metrics-to-buffers
simonsays1980 Jan 15, 2025
a4df011
Fixed a small bugs in 'EpisodeReplayBuffer._sample_batch'.
simonsays1980 Jan 16, 2025
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
7 changes: 7 additions & 0 deletions rllib/algorithms/algorithm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4002,6 +4002,13 @@ def _create_local_replay_buffer_if_necessary(
):
return

# Add parameters, if necessary.
if config["replay_buffer_config"]["type"] == "EpisodeReplayBuffer":
# TODO (simon): If all episode buffers have metrics, check for sublassing.
config["replay_buffer_config"][
"metrics_num_episodes_for_smoothing"
] = self.config.metrics_num_episodes_for_smoothing

return from_config(ReplayBuffer, config["replay_buffer_config"])

@OldAPIStack
Expand Down
6 changes: 6 additions & 0 deletions rllib/algorithms/dqn/dqn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,6 +51,7 @@
NUM_ENV_STEPS_SAMPLED_LIFETIME,
NUM_TARGET_UPDATES,
REPLAY_BUFFER_ADD_DATA_TIMER,
REPLAY_BUFFER_RESULTS,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cool!

REPLAY_BUFFER_SAMPLE_TIMER,
REPLAY_BUFFER_UPDATE_PRIOS_TIMER,
SAMPLE_TIMER,
Expand Down Expand Up @@ -660,6 +661,11 @@ def _training_step_new_api_stack(self):
sample_episodes=True,
)

replay_buffer_results = self.local_replay_buffer.get_metrics()
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nice. Unified API names get_metrics. Analogous to EnvRunners.

self.metrics.merge_and_log_n_dicts(
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, I wonder why log_dict doesn't work here. It should be the better choice here b/c we don't have more than one buffer.

self.metrics.log_dict(
    replay_buffer_results,
    key=REPLAY_BUFFER_RESULTS,
)

Maybe b/c in replay_buffer_results there are already Stats objects with their individual settings? ...

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I need to check it.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So, basically, the lifetime metrics are somehow wrongly accumulated and grow exponentially. They probably need to be reduced before given to the log_dict method.

[replay_buffer_results], key=REPLAY_BUFFER_RESULTS
)

# Perform an update on the buffer-sampled train batch.
with self.metrics.log_time((TIMERS, LEARNER_UPDATE_TIMER)):
learner_results = self.learner_group.update_from_episodes(
Expand Down
10 changes: 7 additions & 3 deletions rllib/tuned_examples/dqn/cartpole_dqn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,11 +18,15 @@
lr=0.0005 * (args.num_learners or 1) ** 0.5,
train_batch_size_per_learner=32,
replay_buffer_config={
"type": "PrioritizedEpisodeReplayBuffer",
"type": "EpisodeReplayBuffer",
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is this just for testing?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, I wanted to check with you, if we proceed like this and then all buffers get the metrics. Then I can test with any of them.

"capacity": 50000,
"alpha": 0.6,
"beta": 0.4,
},
# replay_buffer_config={
# "type": "PrioritizedEpisodeReplayBuffer",
# "capacity": 50000,
# "alpha": 0.6,
# "beta": 0.4,
# },
n_step=(2, 5),
double_q=True,
dueling=True,
Expand Down
28 changes: 28 additions & 0 deletions rllib/utils/metrics/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,34 @@
NUM_MODULE_STEPS_SAMPLED = "num_module_steps_sampled"
NUM_MODULE_STEPS_SAMPLED_LIFETIME = "num_module_steps_sampled_lifetime"

# Counters for adding and evicting in replay buffers.
# TODO (Simon): Check, if we should prefix with 'REPLAY_BUFFER'.
NUM_ENV_STEPS = "num_env_steps"
NUM_ENV_STEPS_ADDED = "num_env_steps_added"
NUM_ENV_STEPS_ADDED_LIFETIME = "num_env_steps_added_lifetime"
NUM_ENV_STEPS_EVICTED = "num_env_steps_evicted"
NUM_ENV_STEPS_EVICTED_LIFETIME = "num_env_steps_evicted_lifetime"
NUM_AGENT_EPISODES = "num_agent_episodes"
NUM_AGENT_EPISODES_ADDED = "num_agent_episodes_added"
NUM_AGENT_EPISODES_ADDED_LIFETIME = "num_agent_episodes_added_lifetime"
NUM_AGENT_EPISODES_EVICTED = "num_agent_episodes_evicted"
NUM_AGENT_EPISODES_EVICTED_LIFETIME = "num_agent_episodes_evicted_lifetime"
NUM_AGENT_EPISODES_PER_SAMPLE = ("num_agent_episodes_per_sample",)
NUM_AGENT_STEPS = "num_agent_steps"
NUM_AGENT_STEPS_ADDED = "num_agent_steps_added"
NUM_AGENT_STEPS_ADDED_LIFETIME = "num_agent_steps_added_lifetime"
NUM_AGENT_STEPS_EVICTED = "num_agent_steps_evicted"
NUM_AGENT_STEPS_EVICTED_LIFETIME = "num_agent_steps_evicted_lifetime"
NUM_EPISODES = "num_episodes"
NUM_EPISODES_ADDED = "num_episodes_added"
NUM_EPISODES_ADDED_LIFETIME = "num_episodes_added_lifetime"
NUM_EPISODES_EVICTED = "num_episodes_evicted"
NUM_EPISODES_EVICTED_LIFETIME = "num_episodes_evicted_lifetime"
NUM_EPISODES_PER_SAMPLE = "num_episodes_per_sample"
NUM_AGENT_STEPS_EVICTED = "num_agent_steps_evicted"
# If some requirements (like length) are not matched we resample.
NUM_MISSED_SAMPLES = "num_missed_samples"

EPISODE_DURATION_SEC_MEAN = "episode_duration_sec_mean"
EPISODE_LEN_MEAN = "episode_len_mean"
EPISODE_LEN_MAX = "episode_len_max"
Expand Down
Loading