Fix debug env, dont run utilizaiton if verbose=False, fixed slowdown …

…over time

config.yaml

@@ -15,11 +15,13 @@ debug:
     device: cpu
     compile: False
     compile_mode: default
-    num_envs: 2
+    num_envs: 1
     num_workers: 1
-    env_batch_size: 16
+    env_batch_size: 4
     env_pool: True
     zero_copy: False
+    batch_size: 128
+    minibatch_size: 128
     batch_rows: 4
     bptt_horizon: 2
     total_timesteps: 100_000_000
@@ -63,6 +65,7 @@ env:
   auto_pokeflute: True
   infinite_money: True
   use_global_map: False
+  save_state: False
 
 
 train:
@@ -73,7 +76,7 @@ train:
   compile_mode: "reduce-overhead"
   float32_matmul_precision: "high"
   total_timesteps: 100_000_000_000
-  batch_size: 65536 
+  batch_size: 131072 # 65536 
   minibatch_size: 2048
   learning_rate: 2.0e-4
   anneal_lr: False

pokemonred_puffer/cleanrl_puffer.py

@@ -133,7 +133,6 @@ class CleanPuffeRL:
     wandb_client: wandb.wandb_sdk.wandb_run.Run | None = None
     profile: Profile = field(default_factory=lambda: Profile())
     losses: Losses = field(default_factory=lambda: Losses())
-    utilization: Utilization = field(default_factory=lambda: Utilization())
     global_step: int = 0
     epoch: int = 0
     stats: dict = field(default_factory=lambda: {})
@@ -155,6 +154,8 @@ def __post_init__(self):
                 clear=True,
             )
 
+            self.utilization = Utilization()
+
         self.vecenv.async_reset(self.config.seed)
         obs_shape = self.vecenv.single_observation_space.shape
         obs_dtype = self.vecenv.single_observation_space.dtype
@@ -199,6 +200,14 @@ def __post_init__(self):
 
     @pufferlib.utils.profile
     def evaluate(self):
+        # Clear all self.infos except for the state
+        for k in list(self.infos.keys()):
+            if k != "state":
+                del self.infos[k]
+            elif len(self.infos["state"]) > 0:
+                # just in case
+                self.infos["state"] = self.infos["state"][-1]
+
         # now for a tricky bit:
         # if we have swarm_frequency, we will take the top swarm_keep_pct envs and evenly distribute
         # their states to the bottom 90%.
@@ -208,6 +217,7 @@ def evaluate(self):
             and hasattr(self.config, "swarm_keep_pct")
             and self.epoch % self.config.swarm_frequency == 0
             and "reward/event" in self.infos
+            and "state" in self.infos
         ):
             # collect the top swarm_keep_pct % of envs
             largest = [
@@ -275,11 +285,17 @@ def evaluate(self):
                 actions = actions.cpu().numpy()
                 mask = torch.as_tensor(mask)  # * policy.mask)
                 o = o if self.config.cpu_offload else o_device
+                if self.config.num_workers == 1:
+                    actions = np.expand_dims(actions, 0)
+                    logprob = logprob.unsqueeze(0)
                 self.experience.store(o, value, actions, logprob, r, d, env_id, mask)
 
                 for i in info:
                     for k, v in pufferlib.utils.unroll_nested_dict(i):
-                        self.infos[k].append(v)
+                        if k == "state":
+                            self.infos[k] = [v]
+                        else:
+                            self.infos[k].append(v)
 
             with self.profile.env:
                 self.vecenv.send(actions)
@@ -441,16 +457,17 @@ def train(self):
 
             done_training = self.global_step >= self.config.total_timesteps
             if self.profile.update(self) or done_training:
-                print_dashboard(
-                    self.config.env,
-                    self.utilization,
-                    self.global_step,
-                    self.epoch,
-                    self.profile,
-                    self.losses,
-                    self.stats,
-                    self.msg,
-                )
+                if self.config.verbose:
+                    print_dashboard(
+                        self.config.env,
+                        self.utilization,
+                        self.global_step,
+                        self.epoch,
+                        self.profile,
+                        self.losses,
+                        self.stats,
+                        self.msg,
+                    )
 
                 if (
                     self.wandb_client is not None
@@ -475,7 +492,8 @@ def train(self):
 
     def close(self):
         self.vecenv.close()
-        self.utilization.stop()
+        if self.config.verbose:
+            self.utilization.stop()
 
         if self.wandb_client is not None:
             artifact_name = f"{self.exp_name}_model"

pokemonred_puffer/environment.py

@@ -102,6 +102,7 @@ def __init__(self, env_config: pufferlib.namespace):
         self.auto_pokeflute = env_config.auto_pokeflute
         self.infinite_money = env_config.infinite_money
         self.use_global_map = env_config.use_global_map
+        self.save_state = env_config.save_state
         self.action_space = ACTION_SPACE
 
         # Obs space-related. TODO: avoid hardcoding?
@@ -552,7 +553,7 @@ def step(self, action):
         self.pokecenters[self.read_m("wLastBlackoutMap")] = 1
         info = {}
 
-        if self.get_events_sum() > self.max_event_rew:
+        if self.save_state and self.get_events_sum() > self.max_event_rew:
             state = io.BytesIO()
             self.pyboy.save_state(state)
             state.seek(0)