feedback fixes

ldp/alg/callbacks.py

@@ -15,7 +15,7 @@
 
 from ldp.agent import Agent
 from ldp.data_structures import Trajectory, Transition
-from ldp.graph.ops import OpResult
+from ldp.graph.ops import OpCtx, OpResult
 
 try:
     import wandb
@@ -375,3 +375,11 @@ def _eval_log(metrics: list[dict[str, list[float]]]) -> None:
             f"eval/{key}_mean": sum(vals) / len(vals)
             for key, vals in flattened_metrics.items()
         })
+
+
+class ClearContextCallback(Callback):
+    async def after_eval_step(self, trajectories: Sequence[Trajectory]) -> None:
+        OpCtx.clear_data()
+
+    async def after_train_step(self, trajectories: Sequence[Trajectory]) -> None:
+        OpCtx.clear_data()

ldp/alg/runners.py

@@ -11,12 +11,12 @@
 from tqdm import tqdm, trange
 
 from ldp.agent import Agent
-from ldp.alg.callbacks import Callback
+from ldp.alg.callbacks import Callback, ClearContextCallback
 from ldp.alg.optimizer import Optimizer
 from ldp.alg.rollout import RolloutManager
 from ldp.data_structures import Trajectory
 from ldp.graph.op_utils import eval_mode, train_mode
-from ldp.graph.ops import OpCtx, OpResult
+from ldp.graph.ops import OpResult
 
 
 async def _run_eval_loop(
@@ -96,11 +96,15 @@ def __init__(
         agent: Agent,
         dataset: TaskDataset,
         callbacks: Sequence[Callback] | None = None,
+        clear_ctx_at_each_iter: bool = True,
     ):
         self.config = config
         self.agent = agent
         self.dataset = dataset
         self.callbacks = callbacks or []
+        if clear_ctx_at_each_iter:
+            clear_cb = ClearContextCallback()
+            self.callbacks = [*self.callbacks, clear_cb] if callbacks else [clear_cb]
         self.rollout_manager = self.config.make_rollout_manager(agent, self.callbacks)
 
     @eval_mode()
@@ -159,6 +163,7 @@ def __init__(
         train_dataset: TaskDataset,
         eval_dataset: TaskDataset | None = None,
         callbacks: Sequence[Callback] | None = None,
+        clear_ctx_at_each_iter: bool = True,
     ):
         if config.eval_every is not None and eval_dataset is None:
             raise ValueError("Must specify eval_dataset if eval_every is set")
@@ -169,6 +174,9 @@ def __init__(
         self.eval_dataset = eval_dataset
         self.optimizer = optimizer
         self.callbacks = callbacks or []
+        if clear_ctx_at_each_iter:
+            clear_cb = ClearContextCallback()
+            self.callbacks = [*self.callbacks, clear_cb] if callbacks else [clear_cb]
         self.rollout_manager = self.config.make_rollout_manager(
             agent=agent, callbacks=self.callbacks
         )
@@ -197,9 +205,6 @@ async def train(self) -> None:
                 if pbar.n == self.config.num_train_iterations:
                     break
 
-                # Clear all op contexts
-                OpCtx.clear_registry()
-
         pbar.close()
 
         await self._eval_loop()
@@ -279,13 +284,17 @@ def __init__(
         optimizer: Optimizer,
         train_trajectories: list[Trajectory],
         callbacks: Sequence[Callback] | None = None,
+        clear_ctx_at_each_iter: bool = True,
     ):
         self.config = config
         self.agent = agent
         self.optimizer = optimizer
         # copy so we can shuffle
         self.train_trajectories = train_trajectories.copy()
         self.callbacks = callbacks or []
+        if clear_ctx_at_each_iter:
+            clear_cb = ClearContextCallback()
+            self.callbacks = [*self.callbacks, clear_cb] if callbacks else [clear_cb]
 
     async def train(self) -> None:
         random.shuffle(self.train_trajectories)
@@ -317,6 +326,3 @@ async def train(self) -> None:
             await asyncio.gather(*[
                 callback.after_train_step(batch) for callback in self.callbacks
             ])
-
-            # Clear all op contexts
-            OpCtx.clear_registry()

ldp/graph/ops.py

@@ -337,7 +337,7 @@ def get_or_create(cls, op_name: str) -> OpCtx:
             return cls(op_name=op_name)  # Create
 
     @classmethod
-    def clear_registry(cls, op_names: Collection[str] | None = None) -> None:
+    def clear_data(cls, op_names: Collection[str] | None = None) -> None:
         """Clear the context registry. If op_names is provided, only clear those contexts."""
         if op_names is None:
             op_names = cls._CTX_REGISTRY.keys()

tests/test_ops.py

@@ -347,13 +347,13 @@ async def test_clear_contexts():
         # Test global clear
         await op2(await op1(1))
         assert len(op1.ctx.data) == len(op2.ctx.data) == 1
-        OpCtx.clear_registry()
+        OpCtx.clear_data()
         assert not op1.ctx.data
         assert not op2.ctx.data
 
         # Test global clear by op name
         await op2(await op1(1))
-        OpCtx.clear_registry(op_names=["op1"])
+        OpCtx.clear_data(op_names=["op1"])
         assert not op1.ctx.data
         assert len(op2.ctx.data) == 1