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

Jump to bottom

[RLlib] Issue 21297: Ignore PPO KL-loss term completely if kl-coeff == 0.0 to avoid NaN values due to some discrete action probs==0.0 #21456

Merged
merged 3 commits into from
Jan 10, 2022
Merged

[RLlib] Issue 21297: Ignore PPO KL-loss term completely if kl-coeff == 0.0 to avoid NaN values due to some discrete action probs==0.0 #21456

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
72f978a
wip.
sven1977 Jan 7, 2022
966e9ab
merge
sven1977 Jan 10, 2022
d2fd71b
fix
sven1977 Jan 10, 2022
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
14 changes: 11 additions & 3 deletions rllib/agents/ppo/ppo_tf_policy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,8 +80,13 @@ def reduce_mean_valid(t):
logp_ratio = tf.exp(
curr_action_dist.logp(train_batch[SampleBatch.ACTIONS]) -
train_batch[SampleBatch.ACTION_LOGP])
action_kl = prev_action_dist.kl(curr_action_dist)
mean_kl_loss = reduce_mean_valid(action_kl)

# Only calculate kl loss if necessary (kl-coeff > 0.0).
if policy.config["kl_coeff"] > 0.0:
action_kl = prev_action_dist.kl(curr_action_dist)
mean_kl_loss = reduce_mean_valid(action_kl)
else:
mean_kl_loss = 0.0

curr_entropy = curr_action_dist.entropy()
mean_entropy = reduce_mean_valid(curr_entropy)
Expand Down Expand Up @@ -110,9 +115,12 @@ def reduce_mean_valid(t):
vf_loss = mean_vf_loss = tf.constant(0.0)

total_loss = reduce_mean_valid(-surrogate_loss +
policy.kl_coeff * action_kl +
policy.config["vf_loss_coeff"] * vf_loss -
policy.entropy_coeff * curr_entropy)
# Add mean_kl_loss (already processed through `reduce_mean_valid`),
# if necessary.
if policy.config["kl_coeff"] > 0.0:
Copy link
Member

Choose a reason for hiding this comment

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

seems this if statement is not necessary, 0.0 case, mean_kl_loss is zero. same below.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Normally, yes, but if the calculated kl is Inf (due to some actions having almost 0.0 probability -> log(a) = -inf), then multiplying it with 0.0 will result in NaN. :)

Copy link
Member

Choose a reason for hiding this comment

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

sorry, wasn't clear enough.
I mean if policy.config["kl_coeff"] <= 0.0, mean_kl_loss is hardcoded to be 0.0 now.
so it seems like we don't really need to have this particular if statement here.

the above one is critical, for sure.

total_loss += policy.kl_coeff * mean_kl_loss

# Store stats in policy for stats_fn.
policy._total_loss = total_loss
Expand Down
15 changes: 12 additions & 3 deletions rllib/agents/ppo/ppo_torch_policy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,8 +106,13 @@ def reduce_mean_valid(t):
logp_ratio = torch.exp(
curr_action_dist.logp(train_batch[SampleBatch.ACTIONS]) -
train_batch[SampleBatch.ACTION_LOGP])
action_kl = prev_action_dist.kl(curr_action_dist)
mean_kl_loss = reduce_mean_valid(action_kl)

# Only calculate kl loss if necessary (kl-coeff > 0.0).
if self.config["kl_coeff"] > 0.0:
action_kl = prev_action_dist.kl(curr_action_dist)
mean_kl_loss = reduce_mean_valid(action_kl)
else:
mean_kl_loss = torch.tensor(0.0, device=logp_ratio.device)

curr_entropy = curr_action_dist.entropy()
mean_entropy = reduce_mean_valid(curr_entropy)
Expand Down Expand Up @@ -137,10 +142,14 @@ def reduce_mean_valid(t):
vf_loss = mean_vf_loss = 0.0

total_loss = reduce_mean_valid(-surrogate_loss +
self.kl_coeff * action_kl +
self.config["vf_loss_coeff"] * vf_loss -
self.entropy_coeff * curr_entropy)

# Add mean_kl_loss (already processed through `reduce_mean_valid`),
# if necessary.
if self.config["kl_coeff"] > 0.0:
total_loss += self.kl_coeff * mean_kl_loss

# Store values for stats function in model (tower), such that for
# multi-GPU, we do not override them during the parallel loss phase.
model.tower_stats["total_loss"] = total_loss
Expand Down