Faster call ID lookups; move progress bar (#16)

ldp/alg/optimizer/opt.py

@@ -4,6 +4,8 @@
 from abc import ABC, abstractmethod
 from collections.abc import Iterable
 
+from tqdm import tqdm
+
 from ldp.data_structures import Trajectory
 
 logger = logging.getLogger(__name__)
@@ -21,9 +23,18 @@ def __init_subclass__(cls) -> None:
         _OPTIMIZER_REGISTRY[cls.__name__] = cls
         return super().__init_subclass__()
 
-    def aggregate(self, trajectories: Iterable[Trajectory]) -> None:
+    def aggregate(
+        self, trajectories: Iterable[Trajectory], show_pbar: bool = False
+    ) -> None:
         """Aggregate trajectories to construct training samples."""
-        for trajectory in trajectories:
+        trajectories_with_pbar = tqdm(
+            trajectories,
+            desc="Aggregating trajectories",
+            ncols=0,
+            mininterval=1,
+            disable=not show_pbar,
+        )
+        for trajectory in trajectories_with_pbar:
             self.aggregate_trajectory(trajectory)
 
     @abstractmethod
@@ -41,9 +52,11 @@ class ChainedOptimizer(Optimizer):
     def __init__(self, *optimizers: Optimizer):
         self.optimizers = optimizers
 
-    def aggregate(self, trajectories: Iterable[Trajectory]) -> None:
+    def aggregate(
+        self, trajectories: Iterable[Trajectory], show_pbar: bool = False
+    ) -> None:
         for optimizer in self.optimizers:
-            optimizer.aggregate(trajectories)
+            optimizer.aggregate(trajectories, show_pbar=show_pbar)
 
     async def update(self) -> None:
         for optimizer in self.optimizers:

ldp/alg/runners.py

@@ -300,14 +300,10 @@ async def train(self) -> None:
                 i_batch_start : i_batch_start + self.config.batch_size
             ]
 
-            batch_with_pbar = tqdm(
-                batch,
-                desc="Aggregating trajectories",
-                ncols=0,
-                # Only show the progress bar if we are doing full-batch optimization
-                disable=len(self.train_trajectories) > self.config.batch_size,
+            # Only show the progress bar if we are doing full-batch optimization
+            self.optimizer.aggregate(
+                batch, show_pbar=len(self.train_trajectories) <= self.config.batch_size
             )
-            self.optimizer.aggregate(batch_with_pbar)
 
             if (training_step + 1) % self.config.update_every == 0:
                 await self.optimizer.update()

ldp/graph/ops.py

@@ -338,22 +338,14 @@ def get_or_create(cls, op_name: str) -> OpCtx:
 
     def get(self, call_id: CallID, key: str, default: Any = NOT_FOUND) -> Any:
         """Get an attribute with an optional default, emulating dict.get."""
-        value = self.data[call_id.run_id].get((call_id.fwd_id, key), default)
+        value = self.data.get(call_id.run_id, {}).get((call_id.fwd_id, key), default)
         if value is NOT_FOUND:
             raise KeyError(f"call_id={call_id}, key='{key}' not found in context")
         return value
 
     def update(self, call_id: CallID, key: str, value: Any):
         self.data[call_id.run_id][(call_id.fwd_id, key)] = value
 
-    @property
-    def call_ids(self) -> set[CallID]:
-        return {
-            CallID(run_id, fwd_id)
-            for run_id, calls in self.data.items()
-            for fwd_id, _ in calls
-        }
-
     def get_input_grads(self, call_id: CallID) -> GradInType:
         # TODO: this function name is confusing. Let's deprecate it. We only use it
         # in tests as far as I can tell.
@@ -463,12 +455,13 @@ def backward(
         """
 
     def get_call_ids(self, run_ids: Collection[UUID] | None = None) -> set[CallID]:
-        call_ids = self.ctx.call_ids
-        return (
-            call_ids
-            if run_ids is None
-            else {c for c in call_ids if c.run_id in run_ids}
-        )
+        ctx = self.ctx
+        if run_ids is None:
+            run_ids = ctx.data.keys()
+
+        # de-duplicate before constructing CallIDs
+        ids = {(run_id, fwd_id) for run_id in run_ids for fwd_id, _ in ctx.data[run_id]}
+        return set(itertools.starmap(CallID, ids))
 
     # This compute_graph() decoration will do nothing if we are already inside a compute graph.
     # We add it here in case we are calling a bare op(), in which case we want a graph