loader/extraction.py

# Reading the published tables back into one file per object collection.
#
# The record stores one row per event, with every object collection flattened into
# <collection>_<branch> columns. Undoing that flattening is all that separates the
# published layout from the converter's own output, which every stage after this one is
# written against.

import logging
from dataclasses import dataclass
from pathlib import Path

import awkward as ak
import pyarrow as pa
import pyarrow.parquet as pq

from . import common

log = logging.getLogger(__name__)

# Column prefixes that name an object collection. A prefix missing from here marks an
# event-level column, which is what keeps nPV_True whole rather than splitting it in two.
OBJECTS = ("ET", "FET", "FHT", "HT", "MET", "MHT", "egammas", "jets", "muons", "taus")

# The collections that can hold more than one object, and so have an order to fix. The
# rest are energy sums, one entry per event.
COLLECTIONS = ("egammas", "jets", "muons", "taus")


@dataclass
class L1DataExtractor:
    """Turn the published tables into one parquet file per object collection.

    :param select_features: ``{object: [branch names]}`` to read, named as the record
        names them. An object mapped to ``none``, or left out, is not extracted.
    :param feat_name_map: ``{object: {branch name: short name}}``, applied on the way
        out, so that everything downstream sees Et, eta and phi.
    :param cache_root_dir: Root of the extracted, processed and ml-ready caches.
    :param name: Names this extraction; the caches built on top of it inherit it.
    """

    select_features: dict
    feat_name_map: dict
    cache_root_dir: str = "data"
    name: str = "default"
    verbose: bool = False

    def extract(self, datasets: dict, data_category: str) -> None:
        """Extract every data set of one category.

        :param datasets: ``{data set: directory holding its published shards}``.
        :param data_category: ``zerobias``, ``background`` or ``signal``.
        """
        self.feats = _selected(self.select_features)
        self.renames = common.as_dict(self.feat_name_map)
        root = Path(self.cache_root_dir) / "extracted" / self.name / data_category
        for name, dataset_dir in common.as_dict(datasets).items():
            if common.cached(root / name, [*self.feats, common.SPLIT_INDEX]):
                log.info("Extracted %s exists at %s.", name, root / name)
                continue
            self._extract_dataset(Path(dataset_dir), root / name)

    def _extract_dataset(self, dataset_dir: Path, out_dir: Path) -> None:
        """Stream one data set's shards into one file per object collection.

        The seeds travel in files of their own, so they are read in a second pass. Both
        passes walk the splits in the same order, which is what keeps the object files
        row aligned without a key to join on.
        """
        shards = _shards(dataset_dir)
        if not shards:
            log.warning(
                "No shards under %s, so %s is left out.", dataset_dir, out_dir.name
            )
            return

        out_dir.mkdir(parents=True, exist_ok=True)
        writers = {}
        for shard in shards:
            _stream(writers, out_dir, self._objects(pq.read_table(shard)))
        for shard in self._seed_shards(dataset_dir):
            _stream(writers, out_dir, {"seeds": _read(shard, self.feats["seeds"])})
        for writer in writers.values():
            writer.close()
        _check_aligned(out_dir)
        log.info("Cached extracted data at %s.", out_dir)

    def _objects(self, table: pa.Table) -> dict:
        """One shard regrouped by object collection, with the split index alongside."""
        objects = {
            common.SPLIT_INDEX: ak.from_arrow(table.select(common.INDEX_COLUMNS))
        }
        for obj, feats in self.feats.items():
            if obj != "seeds":
                objects[obj] = self._collection(table, obj, feats)

        return objects

    def _collection(self, table: pa.Table, obj: str, feats: list[str]) -> ak.Array:
        """One object's columns, under the short names the pipeline works with."""
        prefix = f"{obj}_" if obj in OBJECTS else ""
        mapping = self.renames.get(obj, {})
        array = ak.from_arrow(table.select([f"{prefix}{feat}" for feat in feats]))
        array = ak.Array({mapping.get(f, f): array[f"{prefix}{f}"] for f in feats})

        return _et_ordered(array) if obj in COLLECTIONS else array

    def _seed_shards(self, dataset_dir: Path) -> list[Path]:
        """The menu shards, which only matter when the configuration asks for seeds."""
        return _shards(dataset_dir / "seeds") if "seeds" in self.feats else []


def _et_ordered(data: ak.Array) -> ak.Array:
    """One collection's objects, hardest first.

    The record keeps the order the global trigger read the objects out in. That is ET
    descending for the calorimeter objects but not for the muons, a quarter of the
    multi-muon zero bias events carrying a softer muon ahead of a harder one. The torch
    stage clips each collection to a fixed count and then stacks it by position, so
    without this a truncated event would lose the wrong muons and the leading muon would
    not always land in the leading slot. The sort is stable, so the collections that
    already arrive ordered come out untouched.
    """
    if "Et" not in data.fields:
        log.warning("No Et among %s, so the record's own order is kept.", data.fields)
        return data

    return data[ak.argsort(data["Et"], axis=-1, ascending=False, stable=True)]


def _selected(select_features) -> dict:
    """The objects actually asked for. 'none' is how a configuration leaves one out."""
    return {
        obj: list(feats)
        for obj, feats in common.as_dict(select_features).items()
        if feats and feats != "none"
    }


def _shards(dataset_dir: Path) -> list[Path]:
    """One data set's shards, split by split, so the row order is the published one."""
    return [
        shard
        for split in common.SPLITS
        for shard in sorted(dataset_dir.glob(f"{split}-*.parquet"))
    ]


def _read(shard: Path, feats: list[str]) -> ak.Array:
    return ak.from_arrow(pq.read_table(shard, columns=list(feats)))


def _check_aligned(out_dir: Path) -> None:
    """Refuse a seeds file of a length the events cannot explain.

    The two passes pair a menu decision with its event by row number alone, so a copy of
    the record that is missing shards on one side of the pair would otherwise go through
    and hand every later event another event's trigger decision.
    """
    seeds = out_dir / "seeds.parquet"
    if not seeds.is_file():
        return

    rows = pq.read_metadata(seeds).num_rows
    events = pq.read_metadata(out_dir / f"{common.SPLIT_INDEX}.parquet").num_rows
    if rows != events:
        raise ValueError(
            f"{out_dir.name}: {rows:,} seed rows against {events:,} events. Shards are "
            "missing from the downloaded record, so the two cannot be paired."
        )


def _stream(writers: dict, out_dir: Path, objects: dict) -> None:
    """Append each object's rows to its file, opening the writer on first sight."""
    for obj, array in objects.items():
        table = ak.to_arrow_table(array, extensionarray=False)
        path = out_dir / f"{obj}.parquet"
        if path not in writers:
            writers[path] = pq.ParquetWriter(path, table.schema, compression="snappy")
        writers[path].write_table(table)