lefts 0.1.0__py3-none-any.whl

lefts/__init__.py

@@ -0,0 +1,11 @@
+from .interface import Model, leaf, lift, split, ensemble, tune, feed
+
+__all__ = [
+    "Model",
+    "leaf",
+    "lift",
+    "split",
+    "ensemble",
+    "tune",
+    "feed",
+]

lefts/helpers.py

@@ -0,0 +1,42 @@
+import copy
+import polars as pl
+from typing import Any, Callable
+
+
+def tabular_model(
+    estimator,
+    features: list[str],
+    target: str,
+) -> Callable[..., Any]:
+    """
+    Wraps a sklearn-compatible estimator into a Lefts model factory.
+
+    The returned factory can be passed directly to `leaf(model_constructor=...)`.
+
+    Parameters
+    ----------
+    estimator:
+        Any sklearn-compatible estimator (must implement .fit(X, y) and .predict(X)).
+    features:
+        Column names to use as model inputs.
+    target:
+        Column name of the target variable.
+    """
+
+    def factory(**hyperparameters):
+        est = copy.deepcopy(estimator)
+        if hyperparameters:
+            est.set_params(**hyperparameters)
+
+        class _Model:
+            def fit(self, training_set: pl.DataFrame):
+                X = training_set.select(features).to_numpy()
+                y = training_set[target].to_numpy()
+                est.fit(X, y)
+
+            def predict(self, df: pl.DataFrame):
+                return est.predict(df.select(features).to_numpy())
+
+        return _Model()
+
+    return factory

lefts/interface.py

@@ -0,0 +1,204 @@
+from dataclasses import dataclass
+from typing import Any, Callable, Iterable
+
+from polars import DataFrame, Expr
+
+from .interpreter.fit import _fit
+from .interpreter.predict import _Model
+from .interpreter.masks import _collect_masks
+from .interpreter.labels import _print_tree, _collect_labels
+from .nodes import Ensemble, Feed, Leaf, Tune, Lift, Split
+from .validation import _validate
+
+
+@dataclass
+class Model(_Model):
+    def __post_init__(self):
+        _validate(self.root)
+
+    def fit(self, df: DataFrame):
+        models, hyperparameters = _fit(self.root, df)
+        self.models = models
+        self.hyperparameters = hyperparameters
+
+    def print_tree(self, print_all_labels: bool = False):
+        print(_print_tree(self.root, print_all_labels=print_all_labels))
+
+    def collect_labels(self) -> Iterable[str]:
+        return _collect_labels(self.root)
+
+    def mark_train_validation_test_rows(self, df: DataFrame) -> DataFrame:
+        """
+        Annotate `df` with boolean columns describing whether each
+        row belongs to the train, test and (if applicable) validation
+        sets for each sub model.
+        """
+        new_cols = []
+        for label, masks in _collect_masks(self.root).items():
+            new_cols.append(masks["train"].alias(f"{label}__train"))
+            new_cols.append(masks["test"].alias(f"{label}__test"))
+            if masks["validation"] is not None:
+                new_cols.append(masks["validation"].alias(f"{label}__validation"))
+        return df.with_columns(new_cols)
+
+
+def leaf(model_constructor: Callable[..., Any], label: str) -> Model:
+    """
+    Converts a model into the format required for transformation by lefts.
+
+    Parameters
+    ----------
+    model_constructor
+        A constructor that creates a model with ``.fit()`` and ``.predict()`` methods.
+    label
+        A label for keeping track of this model.
+    """
+
+    leaf_node = Leaf(label=label, factory=model_constructor)
+    return Model(leaf_node)
+
+
+def lift(
+    model: Model,
+    values,
+    name,
+    train_filter,
+    test_filter,
+    validation_filter=None,
+    aggregate_with=None,
+) -> Model:
+    """
+    Creates multiple copies of a model that are trained on (possibly overlapping) train, test and validation sets.
+
+    Parameters
+    ----------
+    model
+        A lefts Model object.
+    values
+        Values to lift the model over. One copy of the model will be trained for each value.
+    name
+        The name of the lift transformation. Has no effect on model training, but controls how the resulting
+        models are labelled and addressed: each leaf beneath the lift gets a label of the form
+        ``"<leaf label>[<name>=<value>]"``. When ``aggregate_with`` is set, the per-value columns are instead
+        collapsed into a single output column named ``name``.
+    train_filter
+        A function mapping each value in ``values`` to a boolean Polars expression indicating whether a given
+        row is in the train set associated with that value.
+    test_filter
+        A function mapping each value in ``values`` to a boolean Polars expression indicating whether a given
+        row is in the test set associated with that value.
+    validation_filter
+        A function mapping each value in ``values`` to a boolean Polars expression indicating whether a given
+        row is in the validation set associated with that value.
+    aggregate_with
+        A function that postprocesses the output columns of the lift. It is called on the set of columns
+        output by the lifted ``.predict()``.
+    """
+    lifted = Lift(
+        child=model.root,
+        values=values,
+        name=name,
+        train_filter=train_filter,
+        test_filter=test_filter,
+        validation_filter=validation_filter,
+        aggregate_with=aggregate_with,
+    )
+
+    return Model(lifted)
+
+
+def split(
+    name: str,
+    model: Model,
+    train_filter: Expr,
+    test_filter: Expr,
+    validation_filter: Expr | None = None,
+) -> Model:
+    """
+    Restricts a model to train, test and (optionally) validate on defined subsets of the available data.
+
+    Parameters
+    ----------
+    name
+        A name used to keep track of this lefts operation in the workflow. Has no effect on model training.
+    model
+        A lefts Model object.
+    train_filter
+        A boolean Polars expression that indicates whether a given row is in the train set.
+    test_filter
+        A boolean Polars expression that indicates whether a given row is in the test set.
+    validation_filter
+        A boolean Polars expression that indicates whether a given row is in the validation set.
+    """
+    node = Split(
+        name=name,
+        child=model.root,
+        train_filter=train_filter,
+        test_filter=test_filter,
+        validation_filter=validation_filter,
+    )
+
+    return Model(node)
+
+
+def ensemble(name: str, *models, aggregate_with=None):
+    """
+    Binds multiple models into a unified model that fits and predicts all of them in parallel.
+
+    Parameters
+    ----------
+    name
+        A name used to keep track of this lefts operation in the workflow. Has no effect on model training.
+    models
+        lefts Model objects.
+    aggregate_with
+        A function that postprocesses the output columns of the ensemble ``.predict()`` method.
+    """
+    roots = [model.root for model in models]
+    node = Ensemble(name, roots, aggregate_with=aggregate_with)
+
+    return Model(node)
+
+
+def tune(
+    name: str, consumer: Model, source: Model, logic: Callable[[Model, DataFrame], dict]
+):
+    """
+    Learn hyperparameters by fitting the source model, applying customisable logic, then passing the resulting
+    dictionary of hyperparameters to the consumer.
+
+    Parameters
+    ----------
+    name
+        A name used to keep track of this lefts operation in the workflow. Has no effect on model training.
+    consumer
+        A lefts Model object. Its leaf factories are instantiated using the outputs of ``logic`` as keyword
+        arguments.
+    source
+        A lefts Model object. It is fitted first; the fitted model is then handed to ``logic`` to derive the
+        hyperparameters.
+    logic
+        A callable ``(fitted_source_model, df) -> dict`` that reads the fitted source and returns the
+        hyperparameters to apply when fitting the consumer.
+    """
+    node = Tune(name=name, consumer=consumer.root, source=source.root, logic=logic)
+
+    return Model(node)
+
+
+def feed(name: str, source: Model, consumer: Model) -> Model:
+    """
+    Chains two models: the source's predictions are available to the consumer as a feature or target during .fit and .predict.
+    Parameters
+    ----------
+    name: A name used to keep track of this lefts operation in the workflow. Has no effect on model training.
+    source: A lefts Model object, which provides the output of its predict to the consumer.
+    consumer: A lefts Model object, which has access to the prediction output of the consumer.
+
+    Returns
+    -------
+
+    """
+    node = Feed(name=name, source=source.root, consumer=consumer.root)
+
+    return Model(node)

lefts/interpreter/fit.py

@@ -0,0 +1,222 @@
+import warnings
+
+from ..nodes import LeftsNode, Leaf, Lift, Split, Ensemble, Tune, Feed
+from .labels import _make_label
+from .masks import _collect_masks
+from .predict import _Model, _predict
+from typing import Tuple, Any
+from polars import DataFrame, Expr, lit
+from inspect import signature
+
+
+def _fit(
+    node: LeftsNode,
+    df: DataFrame,
+    hyperparameters: dict = None,
+    label_context: dict = None,
+    is_root=True,
+    precomputed_masks: dict = None,
+) -> Tuple[dict[str, Any], dict[str, Any]]:
+
+    label_context = label_context or dict()
+    hyperparameters = hyperparameters or dict()
+
+    # Define output types
+    fitted_models: dict[str, Any] = {}
+    output_hyperparameters: dict[str, Any] = {}
+
+    if is_root:
+        precomputed_masks = _collect_masks(node)
+
+    match node:
+        case Leaf(label=label, factory=factory):
+            # Note: it is safe to use the passed hyperparameters
+            # Without further filtering on label, because the hyperparameters
+            # Are passed from a Tune to every node beneath them in the tree.
+            model = factory(**hyperparameters)
+            model_label = _make_label(label, label_context)
+
+            train_mask = precomputed_masks[model_label]["train"]
+            validation_mask = precomputed_masks[model_label]["validation"]
+
+            train_df = df.filter(train_mask)
+
+            fit_signature = signature(model.fit)
+
+            # Now we inspect the signature of fit to determine whether
+            # It expects a validation set, and also to perform runtime checking
+            # That it's signature conforms to expectations.
+
+            allowable_fit_parameters = {"training_set", "validation_set"}
+            excess_parameters = set(fit_signature.parameters) - allowable_fit_parameters
+            assert excess_parameters == set(), (
+                f"Model {label} .fit(...) should only have arguments {allowable_fit_parameters} "
+                f" but has unexpected parameters {excess_parameters}"
+            )
+
+            # Now, if required, we compute the validation set and pass it through to the
+            # fit function as a kwarg.
+            fit_kwargs = dict()
+            if (
+                "validation_set" in fit_signature.parameters
+                and validation_mask is not None
+            ):
+                fit_kwargs["validation_set"] = df.filter(validation_mask)
+
+            elif ("validation_set" not in fit_signature.parameters) and (
+                validation_mask is not None
+            ):
+                raise ValueError(
+                    f"Validation set created but model {label} .fit has no validation_set argument"
+                )
+
+            model.fit(train_df, **fit_kwargs)
+
+            fitted_models[model_label] = model
+
+        case Lift(child=child, values=values, name=name):
+            # Under a lift, we will take the cartesian product
+            # Of the existing labels with the lift values
+            # Filtering appropriately based on each value.
+            for value in values:
+                extended_label_context = {**label_context, name: value}
+
+                child_models, child_hyperparameters = _fit(
+                    child,
+                    df,
+                    hyperparameters,
+                    extended_label_context,
+                    False,
+                    precomputed_masks,
+                )
+
+                fitted_models |= child_models
+                output_hyperparameters |= child_hyperparameters
+
+        case Split(child=child):
+            child_models, child_hyperparameters = _fit(
+                child,
+                df,
+                hyperparameters,
+                label_context,
+                False,
+                precomputed_masks,
+            )
+            fitted_models |= child_models
+            output_hyperparameters |= child_hyperparameters
+
+        case Ensemble():
+            for child in node.children:
+                child_fitted_models, child_learned_hyperparameters = _fit(
+                    child, df, hyperparameters, label_context, False, precomputed_masks
+                )
+                fitted_models |= child_fitted_models
+                output_hyperparameters |= child_learned_hyperparameters
+
+        case Tune(consumer=consumer, source=source, logic=logic):
+            source_models, learned_hyperparameters = _fit(
+                source,
+                df,
+            )
+
+            tune_model = _Model(source, source_models, learned_hyperparameters)
+            learned_hyperparameters |= logic(tune_model, df)
+
+            consumer_models, consumer_hyperparameters = _fit(
+                consumer,
+                df,
+                learned_hyperparameters,
+                label_context,
+                False,
+                precomputed_masks,
+            )
+
+            fitted_models |= source_models | consumer_models
+            output_hyperparameters |= consumer_hyperparameters | learned_hyperparameters
+
+        case Feed(source=source, consumer=consumer):
+            _check_feed_row_compatibility(
+                source, consumer, df, precomputed_masks, label_context
+            )
+            source_models, source_hyperparameters = _fit(
+                source,
+                df,
+                hyperparameters,
+                label_context,
+                False,
+                precomputed_masks,
+            )
+
+            augmented_df = _predict(
+                source,
+                source_models,
+                df,
+                precomputed_masks,
+                label_context,
+            )
+            consumer_models, consumer_hyperparameters = _fit(
+                consumer,
+                augmented_df,
+                hyperparameters,
+                label_context,
+                False,
+                precomputed_masks,
+            )
+            fitted_models |= source_models | consumer_models
+            output_hyperparameters |= source_hyperparameters | consumer_hyperparameters
+
+        case _:
+            raise ValueError(f"Unknown node type {type(node)}")
+
+    return fitted_models, output_hyperparameters
+
+
+def _check_feed_row_compatibility(
+    source_root: LeftsNode,
+    consumer_root: LeftsNode,
+    df: DataFrame,
+    precomputed_masks: dict,
+    label_context: dict,
+) -> None:
+    """
+    Emits warnings if there is suspicious behaviour in the train/test
+    overlap of the source and consumer of a Feed node.
+
+    Specifically it will warn if:
+    - The test set of the source is not a subset of the consumer. This may indicate a data leak.
+    - The test set of the source is a strict subset of the consumer, since this will cause NaNs in the fed column.
+    """
+
+    # It's possible to have multiple leaves with separate train/test
+    # Specification as one source (i.e. an ensemble with an aggregate)
+    # Hence, we take the union of all child masks.
+    def union_mask(node: LeftsNode, mask_kind: str) -> Expr:
+        result = lit(False)
+        for label in _collect_masks(node, label_context):
+            result = result | precomputed_masks[label][mask_kind]
+        return result
+
+    source_train = union_mask(source_root, "train")
+    source_test = union_mask(source_root, "test")
+    consumer_train = union_mask(consumer_root, "train")
+
+    leak_rows = df.filter(source_train & ~consumer_train).height
+    if leak_rows > 0:
+        warnings.warn(
+            f"Feed: source's train set contains {leak_rows} rows not in consumer's "
+            "train set. This may signal a data leak - validate with Model.mark_train_validation_test_rows"
+            "if you are unsure",
+            UserWarning,
+            stacklevel=4,
+        )
+
+    nan_rows = df.filter(consumer_train & ~source_test).height
+    if nan_rows > 0:
+        warnings.warn(
+            f"Feed: {nan_rows} rows in consumer's train set are not in source's "
+            "test set. Source's augmentation column will be NaN on those rows "
+            "during consumer fit. If this is unexpected, validate your train and test filter"
+            "behaviour with Model.mark_train_validation_test_rows",
+            UserWarning,
+            stacklevel=4,
+        )

lefts/interpreter/labels.py

@@ -0,0 +1,117 @@
+from typing import Iterator
+
+from ..nodes import LeftsNode, Leaf, Lift, Split, Ensemble, Feed, Tune
+
+
+def _make_label(leaf_name: str, label_context: dict) -> str:
+    if not label_context:
+        return leaf_name
+    dims = ", ".join(f"{k}={v}" for k, v in label_context.items())
+    return f"{leaf_name}[{dims}]"
+
+
+# Lists longer than this are abbreviated to [first, ..., last] unless the
+# caller asks for the full listing.
+_MAX_LIST = 6
+
+
+def _format_list(items, print_all_labels: bool) -> str:
+    items = list(items)
+    if print_all_labels or len(items) <= _MAX_LIST:
+        body = ", ".join(str(i) for i in items)
+    else:
+        body = f"{items[0]}, ..., {items[-1]}"
+    return f"[{body}]"
+
+
+def _aggregation_suffix(node: LeftsNode) -> str:
+    fn = getattr(node, "aggregate_with", None)
+    if fn is None:
+        return ""
+    fn_name = getattr(fn, "__name__", None) or repr(fn)
+    return f'  ⇒ {fn_name} → "{node.name}"'
+
+
+def _count_models(node: LeftsNode) -> int:
+    """How many leaf models this subtree fits."""
+    match node:
+        case Leaf():
+            return 1
+        case Lift(child=child, values=values):
+            return len(values) * _count_models(child)
+        case _:
+            return sum(_count_models(child) for child in node.children)
+
+
+def _node_header(node: LeftsNode, print_all_labels: bool) -> str:
+    count = _count_models(node)
+    models = f" ({count} model{'' if count == 1 else 's'})"
+    match node:
+        case Leaf(label=label):
+            return f"Leaf '{label}'{models}"
+        case Lift(name=name, values=values):
+            vals = _format_list(values, print_all_labels)
+            return f"Lift '{name}'{models}: {vals}{_aggregation_suffix(node)}"
+        case Split(name=name):
+            return f"Split '{name}'{models}"
+        case Ensemble(name=name):
+            return f"Ensemble '{name}'{models}{_aggregation_suffix(node)}"
+        case Tune(name=name):
+            return f"Tune '{name}'{models}"
+        case Feed(name=name):
+            return f"Feed '{name}'{models}"
+        case _:
+            return getattr(node, "name", repr(node))
+
+
+def _print_tree(
+    node: LeftsNode,
+    print_all_labels: bool = False,
+    prefix: str = "",
+    is_root: bool = True,
+    is_last: bool = True,
+) -> str:
+    header = _node_header(node, print_all_labels)
+
+    if is_root:
+        outputs = _format_list(_collect_labels(node), print_all_labels)
+        lines = [f"{header}  → outputs: {outputs}"]
+        child_prefix = "    "
+    else:
+        connector = "└── " if is_last else "├── "
+        lines = [f"{prefix}{connector}{header}"]
+        child_prefix = prefix + ("    " if is_last else "│   ")
+
+    children = list(node.children)
+    for i, child in enumerate(children):
+        lines.append(
+            _print_tree(
+                child,
+                print_all_labels,
+                prefix=child_prefix,
+                is_root=False,
+                is_last=(i == len(children) - 1),
+            )
+        )
+
+    return "\n".join(lines)
+
+
+def _collect_labels(
+    node: LeftsNode, label_context: dict | None = None
+) -> Iterator[str]:
+    label_context = label_context or {}
+
+    match node:
+        case Lift(aggregate_with=fn) | Ensemble(aggregate_with=fn) if fn is not None:
+            # In the case where we have an aggregation function, we
+            # halt because all child labels will be pulled into the aggregated column
+            yield _make_label(node.name, label_context)
+        case Leaf(label=label):
+            yield _make_label(label, label_context)
+        case Lift(child=child, name=name, values=values):
+            for value in values:
+                yield from _collect_labels(child, label_context | {name: value})
+        case LeftsNode():
+            for child in node.children:
+                yield from _collect_labels(child, label_context)

lefts/interpreter/masks.py

@@ -0,0 +1,96 @@
+from polars import Expr, lit
+
+from ..nodes import LeftsNode, Leaf, Lift, Split
+from .labels import _make_label
+
+
+def _collect_masks(
+    node: LeftsNode,
+    label_context: dict | None = None,
+    train_mask: Expr | None = None,
+    validation_mask: Expr | None = None,
+    test_mask: Expr | None = None,
+) -> dict[str, dict[str, Expr | None]]:
+    """
+    Returns a dictionary of label -> {train: train_mask, validation: validation_mask, test: test_mask}
+    """
+    label_context = label_context or dict()
+    train_mask = train_mask if train_mask is not None else lit(True)
+    test_mask = test_mask if test_mask is not None else lit(True)
+
+    result = {}
+
+    match node:
+        case Leaf(label=leaf_label):
+            full_label = _make_label(leaf_label, label_context)
+            result[full_label] = {
+                "train": train_mask,
+                "validation": validation_mask,
+                "test": test_mask,
+            }
+
+        case Lift(
+            child=child,
+            name=name,
+            values=values,
+            train_filter=train_filter,
+            validation_filter=validation_filter,
+            test_filter=test_filter,
+        ):
+            for value in values:
+                next_label_context = {**label_context, name: value}
+                next_train_mask = train_filter(value) & train_mask
+                next_test_mask = test_filter(value) & test_mask
+
+                if validation_filter is None:
+                    next_validation_mask = validation_mask
+                else:
+                    next_validation_mask = validation_filter(value)
+                    if validation_mask is not None:
+                        next_validation_mask = next_validation_mask & validation_mask
+
+                result.update(
+                    _collect_masks(
+                        child,
+                        next_label_context,
+                        next_train_mask,
+                        next_validation_mask,
+                        next_test_mask,
+                    )
+                )
+
+        case Split(
+            child=child,
+            train_filter=train_filter,
+            validation_filter=validation_filter,
+            test_filter=test_filter,
+        ):
+            next_train_mask = train_filter & train_mask
+            next_test_mask = test_filter & test_mask
+
+            if validation_filter is None:
+                next_validation_mask = validation_mask
+            else:
+                next_validation_mask = validation_filter
+                if validation_mask is not None:
+                    next_validation_mask = next_validation_mask & validation_mask
+
+            result.update(
+                _collect_masks(
+                    child,
+                    label_context,
+                    next_train_mask,
+                    next_validation_mask,
+                    next_test_mask,
+                )
+            )
+
+        case LeftsNode():
+            for child in node.children:
+                result.update(
+                    _collect_masks(
+                        child, label_context, train_mask, validation_mask, test_mask
+                    )
+                )
+
+    return result

lefts/interpreter/predict.py

@@ -0,0 +1,140 @@
+from dataclasses import dataclass
+from typing import Iterator, Optional
+
+from polars import DataFrame, Series
+
+from ..nodes import LeftsNode, Leaf, Lift, Split, Ensemble, Tune, Feed
+from .labels import _make_label, _collect_labels
+from .masks import _collect_masks
+
+
+@dataclass
+class _Model:
+    root: LeftsNode
+    models: Optional[dict] = None
+    hyperparameters: Optional[dict] = None
+
+    def predict(self, df: DataFrame):
+        return _predict(self.root, self.models, df)
+
+    def fit(self, df: DataFrame):
+        raise NotImplementedError()
+
+
+def _get_aggregation_input_columns(
+    node: LeftsNode, label_context: dict
+) -> Iterator[str]:
+    match node:
+        case Lift(child=child, name=name, values=values):
+            for value in values:
+                yield from _collect_labels(child, label_context | {name: value})
+        case Ensemble():
+            for child in node.children:
+                yield from _collect_labels(child, label_context)
+
+
+def _aggregate(node: LeftsNode, df: DataFrame, label_context: dict) -> DataFrame:
+    input_cols = list(_get_aggregation_input_columns(node, label_context))
+    full_col = _make_label(node.name, label_context)
+    df = df.with_columns(node.aggregate_with(*input_cols).alias(full_col))
+    return df.drop(*input_cols)
+
+
+def _predict(
+    node: LeftsNode,
+    models: dict,
+    df: DataFrame,
+    precomputed_masks: dict | None = None,
+    label_context: dict | None = None,
+    is_root: bool = True,
+) -> DataFrame:
+
+    label_context = label_context or {}
+
+    if is_root:
+        if "__lefts_row_index" in df.columns:
+            raise ValueError(
+                "Trying to create column __lefts_row_index but it already exists"
+            )
+        df = df.with_row_index(name="__lefts_row_index")
+        if precomputed_masks is None:
+            precomputed_masks = _collect_masks(node)
+
+    match node:
+        case Leaf(label=label):
+            full_label = _make_label(label, label_context)
+            test_mask = precomputed_masks[full_label]["test"]
+            test_df = df.filter(test_mask)
+            predictions = models[full_label].predict(test_df)
+            predictions = Series(name=full_label, values=predictions)
+            test_df = test_df.with_columns(predictions)
+            df = df.join(
+                test_df.select("__lefts_row_index", full_label),
+                on="__lefts_row_index",
+                coalesce=True,
+                how="left",
+            )
+
+        case Lift(
+            child=child, name=name, values=values, aggregate_with=aggregation_function
+        ):
+            for value in values:
+                df = _predict(
+                    child,
+                    models,
+                    df,
+                    precomputed_masks,
+                    label_context | {name: value},
+                    is_root=False,
+                )
+            if aggregation_function is not None:
+                df = _aggregate(node, df, label_context)
+
+        case Split(child=child):
+            df = _predict(
+                child, models, df, precomputed_masks, label_context, is_root=False
+            )
+
+        case Ensemble(aggregate_with=aggregation_function):
+            for child in node.children:
+                df = _predict(
+                    child,
+                    models,
+                    df,
+                    precomputed_masks,
+                    label_context,
+                    is_root=False,
+                )
+            if aggregation_function is not None:
+                df = _aggregate(node, df, label_context)
+
+        case Feed(source=source, consumer=consumer):
+            df = _predict(
+                source, models, df, precomputed_masks, label_context, is_root=False
+            )
+            df = _predict(
+                consumer, models, df, precomputed_masks, label_context, is_root=False
+            )
+
+        case Tune(consumer=consumer, source=source):
+            df = _predict(
+                source,
+                models,
+                df,
+                precomputed_masks,
+                label_context,
+                is_root=False,
+            )
+            df = _predict(
+                consumer,
+                models,
+                df,
+                precomputed_masks,
+                label_context,
+                is_root=False,
+            )
+
+    if is_root:
+        df = df.drop("__lefts_row_index")
+
+    return df

lefts/nodes.py

@@ -0,0 +1,106 @@
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Callable, Iterable, Protocol, runtime_checkable
+
+from polars import DataFrame, DataType, Expr
+
+
+@runtime_checkable
+class ModelProtocol(Protocol):
+    def fit(self, df): ...
+
+    def predict(self, df): ...
+
+
+class LeftsNode(ABC):
+    name: str
+
+    @property
+    @abstractmethod
+    def children(self) -> Iterable["LeftsNode"]:
+        """Return iterable of child nodes."""
+        ...
+
+
+@dataclass
+class Leaf(LeftsNode):
+    label: str
+    factory: Callable[[], ModelProtocol]
+
+    @property
+    def children(self):
+        return []
+
+    @property
+    def name(self) -> str:
+        return f"Leaf: {self.label}"
+
+
+@dataclass
+class Lift(LeftsNode):
+    name: str
+    child: LeftsNode
+    values: Iterable[DataType]
+    train_filter: Callable[[DataType], Expr]
+    test_filter: Callable[[DataType], Expr]
+    validation_filter: Callable[[DataType], Expr] | None = None
+    aggregate_with: Callable[..., Expr] | None = None
+
+    def __post_init__(self):
+        if len(set(self.values)) != len(self.values):
+            raise ValueError("values must contain no duplicates")
+        self.values = list(self.values)
+
+    @property
+    def children(self) -> Iterable["LeftsNode"]:
+        return [self.child]
+
+
+@dataclass
+class Split(LeftsNode):
+    name: str
+    child: LeftsNode
+    train_filter: Expr
+    test_filter: Expr
+    validation_filter: Expr | None = None
+
+    @property
+    def children(self) -> Iterable["LeftsNode"]:
+        return [self.child]
+
+
+@dataclass
+class Ensemble(LeftsNode):
+    name: str
+    models: Iterable[LeftsNode]
+    aggregate_with: Callable[..., Expr] | None = None
+
+    @property
+    def children(self):
+        return self.models
+
+
+@dataclass
+class Tune(LeftsNode):
+    name: str
+
+    consumer: LeftsNode
+    source: LeftsNode
+    logic: Callable[
+        ["LeftsNode", DataFrame], dict
+    ]  # TODO this should actually typehint Model instead of LeftsNode, but need to re-organise dirs first
+
+    @property
+    def children(self) -> Iterable["LeftsNode"]:
+        return [self.source, self.consumer]
+
+
+@dataclass
+class Feed(LeftsNode):
+    name: str
+    source: LeftsNode
+    consumer: LeftsNode
+
+    @property
+    def children(self) -> Iterable["LeftsNode"]:
+        return [self.source, self.consumer]

lefts/validation.py

@@ -0,0 +1,54 @@
+from typing import Iterator
+
+from .nodes import LeftsNode, Leaf, Lift, Feed
+
+
+_RESERVED_COLUMN_NAMES = {"__lefts_row_index"}
+
+
+def _validate(root: LeftsNode) -> None:
+    _check_unique_node_names(root)
+    _check_no_lift_above_feed(root, under_lift=False)
+    _check_no_reserved_leaf_labels(root)
+
+
+def _check_unique_node_names(root: LeftsNode) -> None:
+    """Every node's identifying name must be globally unique within the tree."""
+    seen = set()
+    duplicates = set()
+    for name in _collect_node_names(root):
+        if name in seen:
+            duplicates.add(name)
+        seen.add(name)
+    if duplicates:
+        raise ValueError(
+            f"Duplicate node names: {sorted(duplicates)}. Every Leaf label and "
+            "every non-leaf node name must be globally unique within the tree."
+        )
+
+
+def _collect_node_names(node: LeftsNode) -> Iterator[str]:
+    yield node.label if isinstance(node, Leaf) else node.name
+    for child in node.children:
+        yield from _collect_node_names(child)
+
+
+def _check_no_lift_above_feed(node: LeftsNode, under_lift: bool) -> None:
+    """Reject Lift as an ancestor of Feed"""
+    if isinstance(node, Feed) and under_lift:
+        raise ValueError(
+            f"Feed {node.name!r} has a Lift as an ancestor. This is currently "
+            "unsupported. Re-express by Lifting first and feeding second"
+        )
+    next_under_lift = under_lift or isinstance(node, Lift)
+    for child in node.children:
+        _check_no_lift_above_feed(child, next_under_lift)
+
+
+def _check_no_reserved_leaf_labels(node: LeftsNode) -> None:
+    if isinstance(node, Leaf) and node.label in _RESERVED_COLUMN_NAMES:
+        raise ValueError(
+            f"Leaf label {node.label!r} collides with a reserved column name."
+        )
+    for child in node.children:
+        _check_no_reserved_leaf_labels(child)

lefts-0.1.0.dist-info/METADATA

@@ -0,0 +1,135 @@
+Metadata-Version: 2.4
+Name: lefts
+Version: 0.1.0
+Summary: Composable transformations for model experimentation
+Author-email: Niklas Mather <niksmather@gmail.com>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/nsmat/lefts
+Project-URL: Repository, https://github.com/nsmat/lefts
+Keywords: machine-learning,ensemble,cross-validation,polars,dsl
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: polars
+Dynamic: license-file
+
+# Lefts: composable machine-learning model transformations
+
+Lefts is a very simple domain specific language for building complex machine learning workflows from simple ones. Starting with your favourite machine learning models, you can use Lefts operations to:
+- Build complex ensembles.
+- Build complex cross validation and hyper-parametrisation procedures.
+- Allow a model to create features or targets for another model.
+- And any creative combination of the above.
+
+Without making subsequent model fitting or evaluation or experimentation any more complex than it was with the original model. This implementation is built on top of the excellent Polars DataFrame library.
+
+
+# Commands
+Lefts has five commands, which give it its name:
+- **L**ift: trains multiple copies of a model across different subsets of data.
+- **E**nsemble: Takes a set of models and makes them evaluate as one.
+- **T**une: Allows a model to learn its hyperparameters from another.
+- **F**eeds: Allows the output of one model to be used as a feature or target by another.
+- **S**plit: Trains a model on a given train/test/validation split.
+
+# Models
+
+Lefts can operate on any model that is defined by:
+- a fit method, which maps from training data into the model parameters
+- a predict method, which maps from model parameters and test data into predictions.
+
+A Lefts command creates a new model by transforming these functions into a new .fit and .predict. Because this new model also has a .fit and .predict, it can be transformed with further Lefts commands.
+
+
+### Conventions
+
+Lefts imposes some constraints on model interfaces.
+- All hyperparameters are passed as arguments to the fit method.
+- We expect that data is passed to fit and predict as Polars dataframes.
+- The predict method returns an iterable, with the order of predictions matching the order on the input training data frame.
+
+See the example notebooks to understand how to adapt your models to the required format.
+
+# An example
+
+The following code shows lefts can be used to create complex models out of more basic ones. We start with an LGBM Regression model, and train an ensemble of models to predict each quantile. The ensemble is trained in a monthly rolling retrain.
+
+See notebooks/quantile_ensemble.py for the full code.
+
+```python
+features = ["temp", "atemp", "hum", "windspeed", "hr", "weekday", "mnth"]
+target = "cnt"
+quantiles = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
+test_period_start_dates = pl.datetime_range(
+    start = dt.datetime(2011, 3, 1),
+    end = dt.datetime(2012, 12, 1),
+    interval='1mo',
+    eager=True
+).to_list()
+
+quantile_models = []
+for q in quantiles:
+    # Convert LGBMRegressor into the format required by lefts
+    base_model = leaf(tabular_model(
+                LGBMRegressor(objective="quantile", alpha=q),
+                features=features,
+                target=target,
+            ),
+            label=f"q{q}",
+                     )
+
+    # 'Lift' each per-quantile model into a family of models, each with a different train and test period
+    rolling_retrain = lift(
+        base_model,
+        name=f"q{q}_rolling_retrain",
+        values=test_period_start_dates,
+
+        # A row is in a given train period if it is before the start of the test period
+        train_filter=lambda test_period_start_date: pl.col("datetime") < test_period_start_date,
+        # Each test period is one month long
+        test_filter=lambda test_period_start_date: pl.col("datetime").dt.month() == test_period_start_date.month,
+        aggregate_with=pl.coalesce,
+    )
+
+    quantile_models.append(rolling_retrain)
+
+model = ensemble("quantiles", *quantile_models)
+
+# Fits |quantiles| x |test_period_start_dates| models
+model.fit(df)
+
+# Adds |quantiles| columns, each with the unique prediction associated with that test row. 
+predictions = model.predict(df)
+```
+
+Behind the scenes, the full workflow is constructed as a tree of lefts expression. You can see this tree by calling `model.print_tree()`
+
+```
+Ensemble 'quantiles' (198 models)  → outputs: [q0.1_rolling_retrain, ..., q0.9_rolling_retrain]
+    ├── Lift 'q0.1_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.1_rolling_retrain"
+    │   └── Leaf 'q0.1' (1 model)
+    ├── Lift 'q0.2_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.2_rolling_retrain"
+    │   └── Leaf 'q0.2' (1 model)
+    ├── Lift 'q0.3_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.3_rolling_retrain"
+    │   └── Leaf 'q0.3' (1 model)
+    ├── Lift 'q0.4_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.4_rolling_retrain"
+    │   └── Leaf 'q0.4' (1 model)
+    ├── Lift 'q0.5_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.5_rolling_retrain"
+    │   └── Leaf 'q0.5' (1 model)
+    ├── Lift 'q0.6_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.6_rolling_retrain"
+    │   └── Leaf 'q0.6' (1 model)
+    ├── Lift 'q0.7_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.7_rolling_retrain"
+    │   └── Leaf 'q0.7' (1 model)
+    ├── Lift 'q0.8_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.8_rolling_retrain"
+    │   └── Leaf 'q0.8' (1 model)
+    └── Lift 'q0.9_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.9_rolling_retrain"
+        └── Leaf 'q0.9' (1 model)
+```

lefts-0.1.0.dist-info/RECORD

@@ -0,0 +1,15 @@
+lefts/__init__.py,sha256=tiyRNtX5l5r9fx8pkp1y3XFf7YbOOUoQlgFXYgS1rmc,175
+lefts/helpers.py,sha256=wDMQyMA5L2-n-CkqXUHLQDS8TmstcumZaMgug3p0dEM,1116
+lefts/interface.py,sha256=DV-Yv4MTX0dFEzvtphJVPVout6JkcfOCf_xSWTUkR9s,7086
+lefts/nodes.py,sha256=9z9DEObOZyVn-itLFl4RaKMxHiWTebDi8yGMLovpaLc,2326
+lefts/validation.py,sha256=KCA7bprPkgXBmBbOWgDHIgqTCz0DEHb2vGm_NcpGWH0,1850
+lefts/interpreter/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+lefts/interpreter/fit.py,sha256=ENpd07BOcjvw7QCvMGXHevRpq9R4Ov27uVn1OrYiYUM,8392
+lefts/interpreter/labels.py,sha256=CyC9ug3vXKDBkA34gtaORlwNlwo2PTwmGYMsqbTDuxM,3998
+lefts/interpreter/masks.py,sha256=loqtbma1ds-U9Vrs-LAjPLryXnP_Ge9UcRHrDAcKoBc,3290
+lefts/interpreter/predict.py,sha256=QHVXvtO7EwKs7V-Rxn3igEnlb8nbR2XKjo98seYN6jE,4428
+lefts-0.1.0.dist-info/licenses/LICENSE,sha256=uNlk1ItRGN0hEHbvdbQmiOwPfN5CGxYD_vjNGBj70G8,11384
+lefts-0.1.0.dist-info/METADATA,sha256=SLfclZ6oUYRSzumVLhYkQgnySjrWmJ9cSemlFSR9qrc,6677
+lefts-0.1.0.dist-info/WHEEL,sha256=aeYiig01lYGDzBgS8HxWXOg3uV61G9ijOsup-k9o1sk,91
+lefts-0.1.0.dist-info/top_level.txt,sha256=1A9PVKzaLgSp4JMNqUa8GDWf1xHCpX8oogBXG8nJDAQ,6
+lefts-0.1.0.dist-info/RECORD,,

lefts-0.1.0.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (82.0.1)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

lefts-0.1.0.dist-info/licenses/LICENSE

@@ -0,0 +1,202 @@
+Copyright 2026 Niklas Mather
+                                Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

lefts-0.1.0.dist-info/top_level.txt

@@ -0,0 +1 @@
+lefts