Merge pull request #573 from lenskit/fix/fallback-bias

Improve pipeline logging and builder API

Author: mdekstrand

lenskit/lenskit/basic/bias.py

@@ -286,7 +286,7 @@ def __init__(
 
     @property
     def is_trained(self) -> bool:
-        return hasattr(self, "bias_")
+        return hasattr(self, "model_")
 
     def train(self, data: Dataset):
         """

lenskit/lenskit/logging/__init__.py

@@ -2,6 +2,7 @@
 Logging, progress, and resource records.
 """
 
+import os
 from typing import Any
 
 import structlog
@@ -21,6 +22,7 @@
 ]
 
 get_logger = structlog.stdlib.get_logger
+_trace_debug = os.environ.get("LK_TRACE", "no").lower() == "debug"
 
 
 def trace(logger: structlog.stdlib.BoundLogger, *args: Any, **kwargs: Any):
@@ -32,3 +34,5 @@ def trace(logger: structlog.stdlib.BoundLogger, *args: Any, **kwargs: Any):
     meth = getattr(logger, "trace", None)
     if meth is not None:
         meth(*args, **kwargs)
+    elif _trace_debug:
+        logger.debug(*args, **kwargs)

lenskit/lenskit/pipeline/__init__.py

@@ -11,12 +11,12 @@
 # pyright: strict
 from __future__ import annotations
 
-import logging
 import typing
 import warnings
 from types import FunctionType, UnionType
 from uuid import NAMESPACE_URL, uuid4, uuid5
 
+import structlog
 from typing_extensions import Any, Literal, Self, TypeAlias, TypeVar, cast, overload
 
 from lenskit.data import Dataset
@@ -51,7 +51,7 @@
     "topn_pipeline",
 ]
 
-_log = logging.getLogger(__name__)
+_log = structlog.stdlib.get_logger(__name__)
 
 # common type var for quick use
 T = TypeVar("T")
@@ -707,6 +707,7 @@ def run_all(self, *nodes: str | Node[Any], **kwargs: object) -> PipelineState:
 
         runner = PipelineRunner(self, kwargs)
         node_list = [self.node(n) for n in nodes]
+        _log.debug("running pipeline", name=self.name, nodes=[n.name for n in node_list])
         if not node_list:
             node_list = self.nodes
 

lenskit/lenskit/pipeline/common.py

@@ -58,7 +58,7 @@ def candidate_selector(self, sel: Component):
         self._selector = sel
 
     def predicts_ratings(
-        self, transform: Component | None = None, *, fallback: Component | None = None
+        self, *, transform: Component | None = None, fallback: Component | None = None
     ):
         """
         Specify that this pipeline will predict ratings, optionally providing a

lenskit/lenskit/pipeline/components.py

@@ -10,6 +10,7 @@
 from __future__ import annotations
 
 import inspect
+import json
 from abc import abstractmethod
 from importlib import import_module
 from types import FunctionType
@@ -214,6 +215,10 @@ def __call__(self, **kwargs: Any) -> COut:
         """
         ...
 
+    def __repr__(self) -> str:
+        params = json.dumps(self.get_config(), indent=2)
+        return f"<{self.__class__.__name__} {params}>"
+
 
 def instantiate_component(
     comp: str | type | FunctionType, config: dict[str, Any] | None

lenskit/lenskit/pipeline/runner.py

@@ -59,7 +59,7 @@ def run(self, node: Node[Any], *, required: bool = True) -> Any:
         elif status == "failed":  # pragma: nocover
             raise RuntimeError(f"{node} previously failed")
 
-        trace(self.log, "processing node %s", node)
+        trace(self.log, "processing node", node=node.name)
         self.status[node.name] = "in-progress"
         try:
             self._run_node(node, required)
@@ -96,6 +96,7 @@ def _inject_input(self, name: str, types: set[type] | None, required: bool) -> N
         if val is not None and types and not is_compatible_data(val, *types):
             raise TypeError(f"invalid data for input {name} (expected {types}, got {type(val)})")
 
+        trace(self.log, "injecting input", name=name, value=val)
         self.state[name] = val
 
     def _run_component(
@@ -107,7 +108,7 @@ def _run_component(
         required: bool,
     ) -> None:
         in_data = {}
-        log = self.log.bind(component=name)
+        log = self.log.bind(node=name)
         trace(log, "processing inputs")
         for iname, itype in inputs.items():
             # look up the input wiring for this parameter input
@@ -158,7 +159,7 @@ def _run_component(
 
             in_data[iname] = ival
 
-        trace(log, "running component")
+        trace(log, "running component", component=comp)
         self.state[name] = comp(**in_data)
 
 

lenskit/tests/basic/test_composite.py

@@ -4,6 +4,7 @@
 # Licensed under the MIT license, see LICENSE.md for details.
 # SPDX-License-Identifier: MIT
 
+import logging
 import pickle
 from typing import Any
 
@@ -20,9 +21,13 @@
 from lenskit.data import Dataset
 from lenskit.data.items import ItemList
 from lenskit.data.types import ID
+from lenskit.operations import predict, score
 from lenskit.pipeline import Pipeline
+from lenskit.pipeline.common import RecPipelineBuilder
 from lenskit.util.test import ml_ds, ml_ratings  # noqa: F401
 
+_log = logging.getLogger(__name__)
+
 
 def test_fallback_fill_missing(ml_ds: Dataset):
     pipe = Pipeline()
@@ -51,3 +56,29 @@ def test_fallback_fill_missing(ml_ds: Dataset):
 
     assert scores[:2] == approx(known(2, ItemList(item_ids=items[:2])).scores())
     assert scores[2:] == approx(bias(2, ItemList(item_ids=items[2:])).scores())
+
+
+def test_fallback_double_bias(rng: np.random.Generator, ml_ds: Dataset):
+    builder = RecPipelineBuilder()
+    builder.scorer(BiasScorer(damping=50))
+    builder.predicts_ratings(fallback=BiasScorer(damping=0))
+    pipe = builder.build("double-bias")
+
+    _log.info("pipeline configuration: %s", pipe.get_config().model_dump_json(indent=2))
+
+    pipe.train(ml_ds)
+
+    for user in rng.choice(ml_ds.users.ids(), 100):
+        items = rng.choice(ml_ds.items.ids(), 500)
+        scores = score(pipe, user, items)
+        scores = scores.scores()
+        assert scores is not None
+        assert not np.any(np.isnan(scores))
+
+        preds = predict(pipe, user, items)
+
+        preds = preds.scores()
+        assert preds is not None
+        assert not np.any(np.isnan(preds))
+
+        assert scores == approx(preds)