Removed old TODOs

metastable_baselines/distributions/distributions.py

@@ -200,7 +200,6 @@ class UniversalGaussianDistribution(SB3_Distribution):
 
         assert std_init >= 0.0, "std can not be initialized to a negative value."
 
-        # TODO: Implement SDE
         self.latent_sde_dim = latent_sde_dim
 
         mean_actions = nn.Linear(latent_dim, self.action_dim)
@@ -348,7 +347,6 @@ class UniversalGaussianDistribution(SB3_Distribution):
 
     def get_noise(self, latent_sde: th.Tensor) -> th.Tensor:
         latent_sde = latent_sde if self.learn_features else latent_sde.detach()
-        # # TODO: Good idea?
         latent_sde = th.nn.functional.normalize(latent_sde, dim=-1)
         # Default case: only one exploration matrix
         if len(latent_sde) == 1 or len(latent_sde) != len(self.exploration_matrices):
@@ -579,7 +577,6 @@ class CholNet(nn.Module):
                                                                         dim2=-1)).diag_embed() + chol.triu(1)
 
     def string(self):
-        # TODO
         return '<CholNet />'
 
 

metastable_baselines/ppo/policies.py

@@ -79,8 +79,6 @@ class ActorCriticPolicy(BasePolicy):
         excluding the learning rate, to pass to the optimizer
     """
 
-    # TODO: Allow passing of dist_kwargs into dist
-
     def __init__(
         self,
         observation_space: gym.spaces.Space,

test.py

@@ -15,18 +15,21 @@ import columbus
 
 from metastable_baselines.distributions import Strength, ParametrizationType, EnforcePositiveType, ProbSquashingType
 
+import torch as th
+
 root_path = '.'
 
 
 def main(env_name='ColumbusCandyland_Aux10-v0', timesteps=1_000_000, showRes=True, saveModel=True, n_eval_episodes=0):
     env = gym.make(env_name)
     use_sde = False
+    # th.autograd.set_detect_anomaly(True)
     ppo = PPO(
         MlpPolicyPPO,
         env,
-        projection=BaseProjectionLayer(), # KLProjectionLayer(trust_region_coeff=0.01),
-        policy_kwargs={'dist_kwargs': {'neural_strength': Strength.NONE, 'cov_strength': Strength.DIAG, 'parameterization_type':
-                       ParametrizationType.NONE, 'enforce_positive_type': EnforcePositiveType.ABS, 'prob_squashing_type': ProbSquashingType.NONE}},
+        projection=BaseProjectionLayer(),  # KLProjectionLayer(trust_region_coeff=0.01),
+        policy_kwargs={'dist_kwargs': {'neural_strength': Strength.NONE, 'cov_strength': Strength.FULL, 'parameterization_type':
+                       ParametrizationType.CHOL, 'enforce_positive_type': EnforcePositiveType.ABS, 'prob_squashing_type': ProbSquashingType.NONE}},
         verbose=0,
         tensorboard_log=root_path+"/logs_tb/" +
         env_name+"/ppo"+(['', '_sde'][use_sde])+"/",
@@ -37,7 +40,7 @@ def main(env_name='ColumbusCandyland_Aux10-v0', timesteps=1_000_000, showRes=Tru
         ent_coef=0.1,  # 0.1
         vf_coef=0.5,
         use_sde=use_sde,  # False
-        clip_range=0.2 # 1  # 0.2,
+        clip_range=None  # 1  # 0.2,
     )
     # trl_frob = PPO(
     #    MlpPolicy,