code/build_public_manifest.py

#!/usr/bin/env python3
"""
Concatenate per-cohort public manifest parquets into unified public manifest.

Inputs (under {output_root}):
    manifest/{cohort}_images.parquet     (one per cohort, written by each adapter_*.py)
    captions/{cohort}_captions.parquet
    manifest/{cohort}_sidecar.parquet    (optional, e.g. IDRiD localization, GAMMA fovea)

Outputs (under {output_root}):
    manifest/public_images_v1.parquet
    manifest/public_studies_v1.parquet
    captions/public_captions_v1.parquet
    schema_v1.json

studies_v1 is derived by groupby(study_id) on images — captures cohort, patient_hash, eye,
device fields, plus has_oct/has_fundus/has_segmentation booleans and image counts.

sidecar parquets are NOT concatenated here — they have heterogeneous schemas (IDRiD has
od/fovea coords, GAMMA has fovea coords). Training pipeline reads them per-cohort as needed.
"""
import argparse
import json
from pathlib import Path

import pandas as pd

from public_common import IMAGE_SCHEMA_COLUMNS, CAPTION_SCHEMA_COLUMNS


PUBLIC_COHORTS = [
    "public_drive_vessel",
    "public_messidor2_dr",
    "public_idrid",
    "public_refuge2_disc_cup",
    "public_eyepacs_combo_dr_aug",
    "public_gamma_multimodal",
]


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--output-root", required=True,
                    help="Public manifest root (.../public_eye_pretrain)")
    ap.add_argument("--cohorts", nargs="*", default=PUBLIC_COHORTS,
                    help="Subset of cohorts to include (default: all)")
    args = ap.parse_args()

    out_root = Path(args.output_root)
    mdir = out_root / "manifest"
    cdir = out_root / "captions"
    assert mdir.exists(), f"missing {mdir}"

    img_parts, cap_parts = [], []
    missing = []
    for cohort in args.cohorts:
        ip = mdir / f"{cohort}_images.parquet"
        cp = cdir / f"{cohort}_captions.parquet"
        if not (ip.exists() and cp.exists()):
            missing.append(cohort)
            continue
        img_parts.append(pd.read_parquet(ip))
        cap_parts.append(pd.read_parquet(cp))
        print(f"  {cohort}: {len(img_parts[-1]):>6} images, {len(cap_parts[-1]):>6} captions")

    if missing:
        print(f"WARN: missing cohort outputs: {missing}")
    if not img_parts:
        print("no cohort parquets found, aborting")
        return

    images_df = pd.concat(img_parts, ignore_index=True)[IMAGE_SCHEMA_COLUMNS]
    captions_df = pd.concat(cap_parts, ignore_index=True)[CAPTION_SCHEMA_COLUMNS]

    # Sanity: image_id uniqueness across cohorts
    dup_imgs = images_df.image_id.duplicated().sum()
    if dup_imgs:
        print(f"ERROR: {dup_imgs} duplicate image_ids across cohorts — aborting")
        print(images_df[images_df.image_id.duplicated(keep=False)].head().to_string())
        return
    # Sanity: caption_id uniqueness
    dup_caps = captions_df.caption_id.duplicated().sum()
    if dup_caps:
        print(f"ERROR: {dup_caps} duplicate caption_ids across cohorts — aborting")
        return

    images_df.to_parquet(mdir / "public_images_v1.parquet", index=False)
    captions_df.to_parquet(cdir / "public_captions_v1.parquet", index=False)

    # studies_v1: one row per (cohort, study_id) with aggregated modality flags
    grp = images_df.groupby(["cohort", "study_id"])
    studies = grp.agg(
        patient_hash=("patient_hash", "first"),
        visit_date=("visit_date", "first"),
        eye=("eye", "first"),
        device_vendor=("device_vendor", "first"),
        device_model=("device_model", "first"),
        hospital_domain=("hospital_domain", "first"),
        ethnicity=("ethnicity", "first"),
        has_fundus=("modality", lambda s: (s == "fundus_color").any()),
        has_oct_bscan=("modality", lambda s: (s == "oct_bscan").any()),
        has_slo=("modality", lambda s: (s == "slo_gray").any()),
        has_segmentation=("has_segmentation", "any"),
        n_images=("image_id", "count"),
    ).reset_index()
    studies.to_parquet(mdir / "public_studies_v1.parquet", index=False)

    schema = {
        "version": "v1",
        "image_columns": IMAGE_SCHEMA_COLUMNS,
        "caption_columns": CAPTION_SCHEMA_COLUMNS,
        "cohorts": list(images_df["cohort"].unique()),
        "n_images": int(len(images_df)),
        "n_captions": int(len(captions_df)),
        "n_studies": int(len(studies)),
        "modality_distribution": images_df["modality"].value_counts().to_dict(),
        "cohort_distribution": images_df["cohort"].value_counts().to_dict(),
    }
    (out_root / "schema_v1.json").write_text(json.dumps(schema, indent=2, ensure_ascii=False))

    print(f"\nWROTE: public_images_v1.parquet ({len(images_df)} rows)")
    print(f"WROTE: public_captions_v1.parquet ({len(captions_df)} rows)")
    print(f"WROTE: public_studies_v1.parquet ({len(studies)} rows)")
    print(f"WROTE: schema_v1.json")
    print("\n--- cohort distribution ---")
    print(images_df["cohort"].value_counts().to_string())
    print("\n--- modality distribution ---")
    print(images_df["modality"].value_counts().to_string())


if __name__ == "__main__":
    main()