scripts/peregrine/01_extract.py

#!/usr/bin/env python3
"""Build the `peregrine` config: one row per LABELED layer per build, shaped
like the ORNL Peregrine dataset (ppak10/Peregrine-Dataset-v2023-11).

Only layers that carry at least one *active* defect label are emitted — this is
the sparse, defect-focused counterpart to the dense `ticks` config, meant for
Peregrine-style transfer learning. Frames are recorded ONLY for labeled layers.

Per-build output `data/peregrine/{build_id:03d}.parquet` (zero-padded) so a new
build only writes its own file — existing builds are never rewritten.

Reads from this dataset's own source/ tree + built frame index:
  source/labels/{build}.parquet        curated active labels (sls-export-labels)
  data/frames_index/{build}.parquet    ts_ms -> (archive zip, member) per frame
  source/telemetry/{build}.parquet     z2 (recoat boundary) + process scalars
  source/recorder/builds.jsonl         build_name

Schema (Peregrine-aligned; images are raw JPEG bytes as `binary`, exactly like
Peregrine's image_after_* columns — decode with PIL.Image.open(io.BytesIO(x))):
  build, build_name, layer, ts
  image_after_powder   post-recoat chamber still (Peregrine after_powder)
  image_after_melt     post-scan   chamber still (Peregrine after_melt)
  part_mask            galvo scan mask           (Peregrine part_ids analog)
  labels               list<{class, bbox[x0,y0,x1,y1] normalized, polarity}>
  has_<class>          bool per PEREGRINE_ALL_CLASSES (12)
  <sensor>_temp / laser_power_w   process scalars nearest-before the layer ts

Frame selection per labeled layer (anchor = earliest label ts on the layer):
  after_melt   = chamber frame nearest the anchor ts (the detection frame)
  after_powder = first chamber frame after the recoat that precedes the anchor
                 (z2 returns to 0 = recoat complete); null if none resolved
  part_mask    = galvo frame nearest the anchor ts

Usage:
    uv run scripts/peregrine/01_extract.py            # all builds with labels
    uv run scripts/peregrine/01_extract.py 48         # specific build ids
"""
import io
import sys
import zipfile
from pathlib import Path

import polars as pl
import pyarrow as pa
import pyarrow.parquet as pq

sys.path.insert(0, str(Path(__file__).parent.parent))
from _lib import DATA_DIR, SOURCE_DIR, TELEMETRY_DIR, EXPORTS_DIR, iter_jsonl

LABELS_DIR = SOURCE_DIR / "labels"
FRAMES_INDEX_DIR = DATA_DIR / "frames_index"
OUTPUT_DIR = DATA_DIR / "peregrine"

# Bit-position order shared with the defect model (runs/*/constants.py).
PEREGRINE_ALL_CLASSES = [
    "powder", "printed", "recoater_hopping", "recoater_streaking",
    "incomplete_spreading", "swelling", "debris", "super_elevation",
    "spatter", "misprint", "over_melting", "under_melting",
]

# Process scalars: (output column, telemetry kind, sensor_id | None for single-sensor).
SCALAR_SPECS = [
    ("printBed_temp",    "temp.current",  "printBed"),
    ("printChamber_temp", "temp.current", "printChamber1"),
    ("powderBed_temp",   "temp.current",  "powderBed"),
    ("surface_temp",     "temp.current",  "surface"),
    ("laser_power_w",    "power.current", None),
]

LABEL_STRUCT = pa.struct([
    pa.field("class", pa.string()),
    pa.field("bbox", pa.list_(pa.float64())),
    pa.field("polarity", pa.string()),
])

OUTPUT_SCHEMA = pa.schema(
    [
        pa.field("build", pa.int64()),
        pa.field("build_name", pa.string()),
        pa.field("layer", pa.int32()),
        pa.field("ts", pa.timestamp("us", tz="UTC")),
        pa.field("image_after_powder", pa.binary()),
        pa.field("image_after_melt", pa.binary()),
        pa.field("part_mask", pa.binary()),
        pa.field("labels", pa.list_(LABEL_STRUCT)),
    ]
    + [pa.field(f"has_{c}", pa.bool_()) for c in PEREGRINE_ALL_CLASSES]
    + [pa.field(name, pa.float64()) for name, _, _ in SCALAR_SPECS]
)


def _build_name(build_id: int) -> str | None:
    for r in iter_jsonl(EXPORTS_DIR / "builds.jsonl"):
        if r.get("id") == build_id:
            return r.get("job_name")
    return None


class FrameStore:
    """Resolve chamber/galvo JPEG bytes by nearest / first-after timestamp.

    Reads the built frame index (ts_ms -> archive zip + member) and lazily opens
    the store-mode zip chunks, caching handles (a labeled build touches only a
    handful of chunks).
    """

    def __init__(self, build_id: int):
        fi = pl.read_parquet(FRAMES_INDEX_DIR / f"{build_id:03d}.parquet")
        self._by_kind = {}
        for kind in ("chamber", "galvo"):
            k = fi.filter(pl.col("kind") == kind).sort("ts_ms")
            self._by_kind[kind] = {
                "ts": k["ts_ms"].to_numpy(),
                "archive": k["archive"].to_list(),
                "member": k["member"].to_list(),
            }
        self._zips: dict[str, zipfile.ZipFile] = {}

    def _read(self, kind: str, idx: int) -> bytes:
        k = self._by_kind[kind]
        archive = k["archive"][idx]
        zf = self._zips.get(archive)
        if zf is None:
            zf = self._zips[archive] = zipfile.ZipFile(archive)
        return zf.read(k["member"][idx])

    def nearest(self, kind: str, ts_ms: int) -> bytes | None:
        ts = self._by_kind[kind]["ts"]
        if len(ts) == 0:
            return None
        import numpy as np
        i = int(np.searchsorted(ts, ts_ms))
        cands = [j for j in (i - 1, i) if 0 <= j < len(ts)]
        best = min(cands, key=lambda j: abs(int(ts[j]) - ts_ms))
        return self._read(kind, best)

    def first_in(self, kind: str, lo_ms: int, hi_ms: int) -> bytes | None:
        """First frame with lo <= ts <= hi (chronological), else None."""
        ts = self._by_kind[kind]["ts"]
        if len(ts) == 0:
            return None
        import numpy as np
        i = int(np.searchsorted(ts, lo_ms, side="left"))
        if i < len(ts) and int(ts[i]) <= hi_ms:
            return self._read(kind, i)
        return None

    def close(self):
        for zf in self._zips.values():
            zf.close()


# z2 is the build piston: it steps once per deposited layer (Peregrine
# "z2 advances once per deposited layer") and ramps monotonically to ~29 mm over
# the build — it does NOT cycle back to 0 per recoat. A layer/recoat boundary is
# therefore a STEP in z2; the ts the step settles ~= the post-recoat moment.
Z2_STEP_MIN = 1.0  # ignore sub-unit sensor jitter


def _recoat_boundaries_ms(tel: pl.DataFrame):
    """ts_ms at each z2 step (recoat/layer boundary). Sorted ascending."""
    z2 = (
        tel.filter(pl.col("kind") == "position.z2")
        .select("ts", "value")
        .sort("ts")
        .with_columns(ts_ms=pl.col("ts").dt.epoch(time_unit="ms"))
    )
    if z2.is_empty():
        return []
    z2 = z2.with_columns(step=(pl.col("value") - pl.col("value").shift(1)).abs())
    steps = z2.filter(pl.col("step") >= Z2_STEP_MIN)
    return steps["ts_ms"].to_list()


def _scalar_series(tel: pl.DataFrame):
    """{output_col: (ts_ms ndarray, value ndarray)} for nearest-before lookup."""
    import numpy as np
    out = {}
    for name, kind, sensor in SCALAR_SPECS:
        sel = tel.filter(pl.col("kind") == kind)
        if sensor is not None:
            sel = sel.filter(pl.col("sensor_id") == sensor)
        sel = sel.select("ts", "value").sort("ts")
        out[name] = (
            sel["ts"].dt.epoch(time_unit="ms").to_numpy(),
            sel["value"].to_numpy(),
        ) if not sel.is_empty() else (np.array([]), np.array([]))
    return out


def _scalar_at(series, ts_ms: int) -> float | None:
    import numpy as np
    ts, val = series
    if len(ts) == 0:
        return None
    i = int(np.searchsorted(ts, ts_ms, side="right")) - 1  # nearest-before
    if i < 0:
        i = 0  # label before first sample: take earliest
    return float(val[i])


def process_build(build_id: int) -> Path | None:
    labels_path = LABELS_DIR / f"{build_id:03d}.parquet"
    fi_path = FRAMES_INDEX_DIR / f"{build_id:03d}.parquet"
    tel_path = TELEMETRY_DIR / f"{build_id:03d}.parquet"
    if not labels_path.exists():
        return None
    labels = pl.read_parquet(labels_path)
    if labels.is_empty():
        return None
    if not fi_path.exists() or not tel_path.exists():
        print(f"  build {build_id}: labels present but frames_index/telemetry "
              f"missing — run sls-deliver-frames + sls-export first; skipped")
        return None

    tel = pl.read_parquet(tel_path)
    frames = FrameStore(build_id)
    boundaries = _recoat_boundaries_ms(tel)
    scalars = _scalar_series(tel)
    build_name = _build_name(build_id)

    import numpy as np
    bnd = np.array(boundaries, dtype="int64")

    rows: list[dict] = []
    # One row per labeled layer; anchor = earliest label ts on that layer.
    for (layer,), grp in labels.sort("ts").group_by(["layer"], maintain_order=True):
        grp = grp.sort("ts")
        anchor_ts_ms = int(grp["ts"].dt.epoch(time_unit="ms")[0])

        label_list = [
            {"class": r["class"], "bbox": list(r["bbox"]), "polarity": r["polarity"]}
            for r in grp.iter_rows(named=True)
        ]
        present = {lbl["class"] for lbl in label_list}

        # after_powder: first chamber after the recoat that precedes the anchor.
        lo = int(bnd[np.searchsorted(bnd, anchor_ts_ms, side="right") - 1]) \
            if len(bnd) and bnd[0] <= anchor_ts_ms else None
        after_powder = frames.first_in("chamber", lo, anchor_ts_ms) if lo is not None else None

        row = {
            "build": build_id,
            "build_name": build_name,
            "layer": int(layer),
            "ts": grp["ts"][0],
            "image_after_powder": after_powder,
            "image_after_melt": frames.nearest("chamber", anchor_ts_ms),
            "part_mask": frames.nearest("galvo", anchor_ts_ms),
            "labels": label_list,
        }
        for c in PEREGRINE_ALL_CLASSES:
            row[f"has_{c}"] = c in present
        for name, _, _ in SCALAR_SPECS:
            row[name] = _scalar_at(scalars[name], anchor_ts_ms)
        rows.append(row)

    frames.close()
    rows.sort(key=lambda r: r["layer"])

    table = pa.Table.from_pylist(rows, schema=OUTPUT_SCHEMA)
    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
    out_path = OUTPUT_DIR / f"{build_id:03d}.parquet"
    pq.write_table(table, out_path, compression="zstd")
    print(f"  build {build_id}: {len(rows)} labeled layer(s) -> "
          f"{out_path.relative_to(DATA_DIR.parent)}")
    return out_path


def _builds_with_labels() -> list[int]:
    return sorted(int(p.stem) for p in LABELS_DIR.glob("*.parquet"))


def main():
    args = sys.argv[1:]
    builds = [int(a) for a in args] if args else _builds_with_labels()
    if not builds:
        print("no builds with source/labels/*.parquet found")
        return
    written = 0
    for b in builds:
        if process_build(b) is not None:
            written += 1
    print(f"done: {written} build(s) written to {OUTPUT_DIR.relative_to(DATA_DIR.parent)}")


if __name__ == "__main__":
    main()