Skip to content

Fix solving linear and square normalizer #35

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

Open
mahaitongdae wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Fix solving linear and square normalizer #35

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
93e328b
fix solve square normalizer
Mar 26, 2026
fcb3a84
fix solve linear and square normalizer
Mar 26, 2026
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
9 changes: 7 additions & 2 deletions flowrl/agent/online/dpmd.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,11 @@ def solve_normalizer_exp(q: jnp.ndarray, temp: float):

def solve_normalizer_linear(q: jnp.ndarray, temp: float, negative: float=0.0):
num_particles = q.shape[-1]

# convert from sum to 1 to sum to num_particles
temp = temp * num_particles
negative = negative / num_particles

q_sorted = jnp.sort(q, axis=-1)[..., ::-1]
q_cumsum = jnp.cumsum(q_sorted, axis=-1)

Expand All @@ -40,7 +45,7 @@ def solve_normalizer_linear(q: jnp.ndarray, temp: float, negative: float=0.0):

def solve_normalizer_square(q: jnp.ndarray, temp: float):
num_particles = q.shape[-1]
target_sum = temp ** 2
target_sum = temp ** 2 * num_particles

q_sorted = jnp.sort(q, axis=-1)[..., ::-1]
q_cumsum = jnp.cumsum(q_sorted, axis=-1)
Expand Down Expand Up @@ -159,6 +164,7 @@ def critic_loss_fn(critic_params: Param, dropout_rng: PRNGKey) -> Tuple[jnp.ndar
if reweight == "exp":
nu = solve_normalizer_exp(q_batch, temp())
weights = jnp.exp((q_batch - nu) / temp())
weights = weights * num_particles
elif reweight == "linear":
nu = solve_normalizer_linear(q_batch, temp(), negative=negative_bound)
weights = jnp.maximum((q_batch - nu) / temp(), negative_bound)
Expand All @@ -170,7 +176,6 @@ def critic_loss_fn(critic_params: Param, dropout_rng: PRNGKey) -> Tuple[jnp.ndar
ent_weights = jnp.maximum(weights, 1e-6)
ent_weights = ent_weights / ent_weights.sum(axis=-1, keepdims=True)
entropy = - jnp.sum(ent_weights * jnp.log(ent_weights+1e-6), axis=-1)
weights = weights * num_particles

_, at, t, eps = actor.add_noise(add_noise_rng, action_batch)

Expand Down