Small fixes for resampling (sorry, last change)

baselines/ippo/config/ippo_ff_puffer.yaml

@@ -16,8 +16,8 @@ environment: # Overrides default environment configs (see pygpudrive/env/config.
   remove_non_vehicles: true # If false, all agents are included (vehicles, pedestrians, cyclists)
   use_lidar_obs: false # NOTE: Setting this to true currently turns of the other observation types
   reward_type: "weighted_combination"
-  collision_weight: -0.03
-  off_road_weight: -0.03
+  collision_weight: -0.035
+  off_road_weight: -0.035
   goal_achieved_weight: 1.0
   dynamics_model: "classic"
   collision_behavior: "ignore" # Options: "remove", "stop"
@@ -48,12 +48,13 @@ train:
   # # # Data sampling # # #
   resample_scenes: false
   resample_criterion: "global_step"
-  resample_interval: 3_000_000
+  resample_interval: 5_000_000
+  resample_limit: 10000 # Resample until the limit is reached; set to a large number to continue resampling indefinitely
   resample_mode: "random" # Options: random
 
   # # # PPO # # #
   torch_deterministic: false
-  total_timesteps: 500_000_000
+  total_timesteps: 1_000_000_000
   batch_size: 131_072
   minibatch_size: 16_384
   learning_rate: 3e-4

integrations/rl/puffer/ppo.py

@@ -32,12 +32,6 @@
 from integrations.rl.puffer.c_gae import compute_gae
 from integrations.rl.puffer.logging import print_dashboard, abbreviate
 
-global_steps_list = []
-perc_collisions_list = []
-perc_offroad_list = []
-perc_goal_achieved_list = []
-time_list = []
-
 
 def create(config, vecenv, policy, optimizer=None, wandb=None):
     seed_everything(config.seed, config.torch_deterministic)
@@ -110,11 +104,11 @@ def evaluate(data):
     # Resample data logic
     if data.config.resample_scenes:
         if (
-            (data.config.resample_limit is None or data.resample_counter < data.config.resample_limit)
+            data.resample_counter < int(data.config.resample_limit)
             and data.config.resample_criterion == "global_step"
             and data.resample_buffer >= data.config.resample_interval
         ):
-            print(f"Resampling scenarios: {data.resample_counter + 1}" + 
+            print(f"Resampling scenarios: {data.resample_counter + 1:,}" + 
                 (f" / {data.config.resample_limit}" if data.config.resample_limit is not None else ""))
 
             # Sample new batch of scenarios
@@ -124,17 +118,8 @@ def evaluate(data):
             if data.config.resample_limit is not None:  # Increment counter only if there is a limit
                 data.resample_counter += 1
 
-    data.vecenv.rendering_in_progress = {
-        env_idx: False for env_idx in range(data.config.render_k_scenarios)
-    }
-    data.vecenv.was_rendered_in_rollout = {
-        env_idx: False for env_idx in range(data.config.render_k_scenarios)
-    }
     data.vecenv.wandb_obj = data.wandb
-    data.vecenv.frames = {
-        env_idx: []
-        for env_idx in range(data.config.render_k_scenarios)
-    }
+    data.vecenv.clear_render_storage()
 
     config, profile, experience = data.config, data.profile, data.experience
 

integrations/rl/puffer/puffer_env.py

@@ -392,7 +392,6 @@ def render(self):
                     self.episode_lengths[render_env_idx, :][0] == 0
                     and not self.was_rendered_in_rollout[render_env_idx]
                 ):
-                    print(f"Render at iter: { self.iters}")
                     self.rendering_in_progress[render_env_idx] = True
 
         # Continue rendering if in progress
@@ -491,4 +490,5 @@ def clear_render_storage(self):
         for env_idx in range(self.train_config.render_k_scenarios):
             self.frames[env_idx] = []
             self.rendering_in_progress[env_idx] = False
-            self.was_rendered_in_rollout[env_idx] = True
+            self.was_rendered_in_rollout[env_idx] = False
+