Evaluator num_rollouts_per_env

ldp/alg/runners.py

@@ -28,6 +28,7 @@ async def _run_eval_loop(
     max_rollout_steps: int | None,
     callbacks: Sequence[Callback],
     shuffle: bool = False,
+    num_rollouts_per_env: int = 1,
 ) -> None:
     await asyncio.gather(*[callback.before_eval_loop() for callback in callbacks])
 
@@ -48,16 +49,20 @@ async def _run_eval_loop(
         # We use pbar.n as a counter for the number of training steps
         while pbar.n < num_iterations:
             for batch in dataset.iter_batches(batch_size, shuffle=shuffle):
-                trajectories = await rollout_manager.sample_trajectories(
-                    environments=batch, max_steps=max_rollout_steps
-                )
+                all_trajectories: list[Trajectory] = []
+
+                for _ in range(num_rollouts_per_env):
+                    trajectories = await rollout_manager.sample_trajectories(
+                        environments=batch, max_steps=max_rollout_steps
+                    )
+                    all_trajectories.extend(trajectories)
 
-                # Close the environment after we have sampled from it,
-                # in case it needs to tear down resources.
-                await _close_envs(batch, rollout_manager.catch_env_failures)
+                    # Close the environment after we have sampled from it,
+                    # in case it needs to tear down resources.
+                    await _close_envs(batch, rollout_manager.catch_env_failures)
 
                 await asyncio.gather(*[
-                    callback.after_eval_step(trajectories) for callback in callbacks
+                    callback.after_eval_step(all_trajectories) for callback in callbacks
                 ])
                 pbar.update()
 
@@ -79,6 +84,7 @@ class EvaluatorConfig(BaseModel):
             "If 0, will not run the eval loop. "
         ),
     )
+    num_rollouts_per_env: int = 1
     max_rollout_steps: int | None = None
     catch_agent_failures: bool = True
     catch_env_failures: bool = True
@@ -131,6 +137,7 @@ async def run(self, **kwargs) -> None:
             "num_iterations": self.config.num_eval_iterations,
             "max_rollout_steps": self.config.max_rollout_steps,
             "shuffle": self.config.shuffle,
+            "num_rollouts_per_env": self.config.num_rollouts_per_env,
         }
         eval_kwargs |= kwargs
         await _run_eval_loop(