dodox/itpal_jax
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fixes for contextual KL

Browse Source
This commit is contained in:
Dominik Moritz Roth 2024-12-21 18:53:11 +01:00
parent de2b9a10d6
commit 3e2b988a2f
1 changed files with 20 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
itpal_jax/kl_projection.py
View File

@ -88,8 +88,18 @@ class KLProjection(BaseProjection):
def get_trust_region_loss(self, policy_params: Dict[str, jnp.ndarray], def get_trust_region_loss(self, policy_params: Dict[str, jnp.ndarray],
proj_policy_params: Dict[str, jnp.ndarray]) -> jnp.ndarray: proj_policy_params: Dict[str, jnp.ndarray]) -> jnp.ndarray:
mean, scale_or_tril = policy_params["loc"], policy_params["scale"] """Compute trust region loss between original and projected parameters."""
proj_mean, proj_scale_or_tril = proj_policy_params["loc"], proj_policy_params["scale"] # Get the right scale parameter based on full_cov
mean = policy_params["loc"]
proj_mean = proj_policy_params["loc"]
if self.full_cov:
scale_or_tril = policy_params["scale_tril"]
proj_scale_or_tril = proj_policy_params["scale_tril"]
else:
scale_or_tril = policy_params["scale"]
proj_scale_or_tril = proj_policy_params["scale"]
kl = sum(self._gaussian_kl((mean, scale_or_tril), (proj_mean, proj_scale_or_tril))) kl = sum(self._gaussian_kl((mean, scale_or_tril), (proj_mean, proj_scale_or_tril)))
return jnp.mean(kl) * self.trust_region_coeff return jnp.mean(kl) * self.trust_region_coeff
@ -151,14 +161,12 @@ class KLProjection(BaseProjection):
old_cov = old_scale_or_tril ** 2 old_cov = old_scale_or_tril ** 2
mask = cov_part > self.cov_bound mask = cov_part > self.cov_bound
proj_scale_or_tril = jnp.zeros_like(scale_or_tril)
proj_scale_or_tril = jnp.where(~mask, scale_or_tril, proj_scale_or_tril)
if mask.any(): if mask.any():
if self.full_cov: if self.full_cov:
proj_cov = project_full_covariance(cov, scale_or_tril, old_scale_or_tril, self.cov_bound) proj_cov = project_full_covariance(cov, scale_or_tril, old_scale_or_tril, self.cov_bound)
is_invalid = jnp.isnan(proj_cov.mean(axis=(-2, -1))) & mask is_invalid = jnp.isnan(proj_cov.mean(axis=(-2, -1)))
proj_scale_or_tril = jnp.where(is_invalid, old_scale_or_tril, proj_scale_or_tril) proj_scale_or_tril = jnp.where(is_invalid[..., None, None], old_scale_or_tril, scale_or_tril)
mask = mask & ~is_invalid mask = mask & ~is_invalid
chol = jnp.linalg.cholesky(proj_cov) chol = jnp.linalg.cholesky(proj_cov)
proj_scale_or_tril = jnp.where(mask[..., None, None], chol, proj_scale_or_tril) proj_scale_or_tril = jnp.where(mask[..., None, None], chol, proj_scale_or_tril)
@ -166,10 +174,12 @@ class KLProjection(BaseProjection):
proj_cov = project_diag_covariance(cov, old_cov, self.cov_bound) proj_cov = project_diag_covariance(cov, old_cov, self.cov_bound)
is_invalid = (jnp.isnan(proj_cov.mean(axis=-1)) | is_invalid = (jnp.isnan(proj_cov.mean(axis=-1)) |
jnp.isinf(proj_cov.mean(axis=-1)) | jnp.isinf(proj_cov.mean(axis=-1)) |
(proj_cov.min(axis=-1) < 0)) & mask (proj_cov.min(axis=-1) < 0))
proj_scale_or_tril = jnp.where(is_invalid, old_scale_or_tril, proj_scale_or_tril) proj_scale_or_tril = jnp.where(is_invalid[..., None], old_scale_or_tril, scale_or_tril)
mask = mask & ~is_invalid mask = mask & ~is_invalid
proj_scale_or_tril = jnp.where(mask[..., None], jnp.sqrt(proj_cov), proj_scale_or_tril) proj_scale_or_tril = jnp.where(mask[..., None], jnp.sqrt(proj_cov), scale_or_tril)
else:
proj_scale_or_tril = scale_or_tril
return proj_scale_or_tril return proj_scale_or_tril