edit/chordedit/run_pie_bench.py

from __future__ import annotations

import argparse
import json
import logging
import shutil
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Dict, List, Optional

import torch
from PIL import Image

from pipeline_chord import ChordEditPipeline
from utils import first_param_point, load_yaml_config


LOGGER = logging.getLogger("pie_bench")

# model root + expected component subdirectories
COMPONENT_SUBDIRS: Dict[str, str] = {
    "unet_path": "unet",
    "scheduler_path": "scheduler",
    "text_encoder_path": "text_encoder",
    "tokenizer_path": "tokenizer",
    "vae_path": "vae",
}
DEFAULT_MODEL_ROOT = "/sd-turbo"
DEFAULT_COMPONENT_PATHS: Dict[str, str] = {
    key: str(Path(DEFAULT_MODEL_ROOT) / subdir) for key, subdir in COMPONENT_SUBDIRS.items()
}

DEFAULT_EDIT_CONFIG = {
    "noise_samples": 1,
    "n_steps": 1,
    "t_start": 0.90,
    "t_end": 0.30,
    "t_delta": 0.15,
    "step_scale": 1.0,
    "cleanup": True,
}

DEFAULT_SEED = 42
DEFAULT_PRECISION = "fp32"

DEFAULT_PIE_ROOT = Path(__file__).resolve().parent / "pie_bench"
DEFAULT_MAPPING_FILE = "mapping_file.json"
DEFAULT_IMAGE_SUBDIR = "annotation_images"
DEFAULT_METHOD_NAME = "ChordEdit"


@dataclass(frozen=True)
class PieRecord:
    sample_id: str
    image_path: Path
    relative_path: Path
    original_prompt: str
    edited_prompt: str
    edit_instruction: str


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Run ChordEdit on PIE-Bench data and export PIE-format results.")
    parser.add_argument("--config", type=str, default=None, help="Optional YAML config describing edit params.")
    parser.add_argument(
        "--model-root",
        type=str,
        default=DEFAULT_MODEL_ROOT,
        help="Root folder containing unet/scheduler/text_encoder/tokenizer/vae subfolders.",
    )
    parser.add_argument("--device", type=str, default=None, help="Torch device override, e.g. cuda:0 or cpu.")
    parser.add_argument("--precision", choices=["fp32", "fp16", "bf16"], default=None, help="Computation precision.")
    parser.add_argument("--seed", type=int, default=None, help="Random seed overriding the config file.")
    parser.add_argument("--noise-samples", type=int, default=None, help="Number of MC noise samples.")
    parser.add_argument("--n-steps", type=int, default=None, help="Number of Chord iterations.")
    parser.add_argument("--t-start", type=float, default=None, help="Edit timestep start.")
    parser.add_argument("--t-end", type=float, default=None, help="Edit timestep end.")
    parser.add_argument("--t-delta", type=float, default=None, help="Edit timestep delta.")
    parser.add_argument("--step-scale", type=float, default=None, help="Edit update magnitude.")
    parser.add_argument("--cleanup", action="store_true", help="Force cleanup on.")
    parser.add_argument("--no-cleanup", action="store_true", help="Force cleanup off.")
    parser.add_argument(
        "--center-crop",
        dest="center_crop",
        action="store_true",
        default=True,
        help="Center-crop before resize for VAE preprocessing (default).",
    )
    parser.add_argument(
        "--no-center-crop",
        dest="center_crop",
        action="store_false",
        help="Disable center crop before resizing.",
    )
    parser.add_argument(
        "--use-attention-mask",
        action="store_true",
        help="Pass attention masks to the text encoder (defaults off to mirror chord/src).",
    )
    parser.add_argument(
        "--safety-checker",
        dest="use_safety_checker",
        action="store_true",
        help="Enable StableDiffusion safety checker before exporting images.",
    )
    parser.add_argument(
        "--no-safety-checker",
        dest="use_safety_checker",
        action="store_false",
        default=False,
        help="Disable safety checker (default).",
    )
    parser.add_argument("--image-size", type=int, default=512, help="Resolution used when feeding the VAE.")
    parser.add_argument("--max-samples", type=int, default=None, help="Only process the first N records.")

    parser.add_argument("--pie-root", type=str, default=None, help="Root directory of PIE-Bench data.")
    parser.add_argument(
        "--mapping-file",
        type=str,
        default=DEFAULT_MAPPING_FILE,
        help="Mapping file path relative to --pie-root.",
    )
    parser.add_argument(
        "--image-subdir",
        type=str,
        default=DEFAULT_IMAGE_SUBDIR,
        help="Subdirectory (inside --pie-root) containing the original PIE annotation images.",
    )
    parser.add_argument(
        "--export-root",
        type=str,
        default=None,
        help="Directory that follows PIE-Bench layout (data/... + output/...). Defaults to --pie-root.",
    )
    parser.add_argument(
        "--method-name",
        type=str,
        default=DEFAULT_METHOD_NAME,
        help="Name used under export_root/output/<method_name>/annotation_images.",
    )
    parser.add_argument(
        "--output-subdir",
        type=str,
        default=DEFAULT_IMAGE_SUBDIR,
        help="Subdirectory inside output/<method_name>/ for generated images.",
    )
    parser.add_argument(
        "--source-subdir",
        type=str,
        default=DEFAULT_IMAGE_SUBDIR,
        help="Subdirectory inside data/ where original images are copied when --copy-source is set.",
    )
    parser.add_argument("--copy-source", action="store_true", help="Copy the source PIE images into export_root/data.")
    parser.add_argument(
        "--mapping-dest",
        type=str,
        default="data/mapping_file.json",
        help="Relative path (from export_root) to write the mapping file.",
    )
    parser.add_argument(
        "--no-sync-mapping",
        action="store_true",
        help="Skip copying the PIE mapping file into export_root.",
    )
    parser.add_argument("--overwrite", action="store_true", help="Overwrite existing predictions when present.")
    parser.add_argument(
        "--log-every",
        type=int,
        default=25,
        help="Progress logging interval in number of saved samples (0 disables incremental logs).",
    )

    return parser.parse_args()


def dtype_from_precision(value: Optional[str]) -> torch.dtype:
    precision = (value or DEFAULT_PRECISION).lower()
    mapping = {
        "fp32": torch.float32,
        "fp16": torch.float16,
        "bf16": torch.bfloat16,
    }
    if precision not in mapping:
        raise ValueError(f"Unsupported precision '{value}'. Choose from {list(mapping)}.")
    return mapping[precision]


def expand_component_paths(path_map: Dict[str, Optional[str]]) -> Dict[str, str]:
    expanded: Dict[str, str] = {}
    for key in COMPONENT_SUBDIRS:
        value = path_map.get(key)
        fallback = DEFAULT_COMPONENT_PATHS.get(key)
        final_value = value if value is not None else fallback
        if final_value is None:
            raise ValueError(f"Missing required path for '{key}'. Provide via config or CLI.")
        expanded[key] = str(Path(final_value).expanduser().resolve())
    return expanded


def paths_from_model_root(model_root: str | Path) -> Dict[str, str]:
    root = Path(model_root).expanduser().resolve()
    return {key: str((root / subdir).resolve()) for key, subdir in COMPONENT_SUBDIRS.items()}


def load_pipeline_config(path: Optional[str]) -> tuple[Dict[str, Any], int, Optional[str]]:
    if path is None:
        return (dict(DEFAULT_EDIT_CONFIG), DEFAULT_SEED, DEFAULT_PRECISION)

    cfg = load_yaml_config(path)
    editor_cfg = cfg.get("editor", {})
    seed_value = editor_cfg.get("seed")
    if seed_value is None:
        seed_list = editor_cfg.get("seed_list")
        if isinstance(seed_list, (list, tuple)) and seed_list:
            seed_value = seed_list[0]
        elif seed_list is not None:
            seed_value = seed_list
    seed_value = int(seed_value) if seed_value is not None else DEFAULT_SEED
    precision = editor_cfg.get("precision", DEFAULT_PRECISION)

    params_grid = editor_cfg.get("params_grid", {})
    edit_config = first_param_point(params_grid) if params_grid else dict(DEFAULT_EDIT_CONFIG)

    return edit_config, seed_value, precision


def apply_cli_overrides(args: argparse.Namespace, edit_config: Dict[str, Any], seed: Optional[int]) -> tuple[Dict[str, Any], int]:
    overrides = {
        "noise_samples": args.noise_samples,
        "n_steps": args.n_steps,
        "t_start": args.t_start,
        "t_end": args.t_end,
        "t_delta": args.t_delta,
        "step_scale": args.step_scale,
    }
    for key, value in overrides.items():
        if value is not None:
            edit_config[key] = value

    if args.cleanup:
        edit_config["cleanup"] = True
    elif args.no_cleanup:
        edit_config["cleanup"] = False

    cli_seed = args.seed
    seed_value = seed if cli_seed is None else cli_seed
    if seed_value is None:
        seed_value = DEFAULT_SEED
    return edit_config, int(seed_value)


def resolve_path(base: Path, maybe_relative: str | Path) -> Path:
    candidate = Path(maybe_relative)
    if candidate.is_absolute():
        return candidate.expanduser().resolve()
    return (base / candidate).expanduser().resolve()


def load_pie_records(root: Path, mapping_path: Path, image_subdir: str) -> List[PieRecord]:
    if not mapping_path.exists():
        raise FileNotFoundError(f"PIE mapping file not found: {mapping_path}")

    with mapping_path.open("r", encoding="utf-8") as handle:
        mapping = json.load(handle)

    if not isinstance(mapping, dict):
        raise ValueError(f"Expected mapping JSON to be a dict, got {type(mapping).__name__}")

    img_root = (root / image_subdir).expanduser().resolve()
    if not img_root.exists():
        raise FileNotFoundError(f"PIE image directory does not exist: {img_root}")

    records: List[PieRecord] = []
    for sample_id in sorted(mapping.keys()):
        meta = mapping[sample_id]
        rel_value = meta.get("image_path")
        if rel_value is None:
            LOGGER.warning("Sample %s is missing 'image_path'; skipping.", sample_id)
            continue
        rel_path = Path(rel_value)
        abs_path = (img_root / rel_path).expanduser().resolve()
        if not abs_path.exists():
            LOGGER.warning("Sample %s image not found at %s; skipping.", sample_id, abs_path)
            continue

        original_prompt = meta.get("original_prompt") or meta.get("source_prompt") or ""
        edited_prompt = meta.get("editing_prompt") or meta.get("edited_prompt") or meta.get("target_prompt") or ""
        edit_instruction = meta.get("editing_instruction") or meta.get("edit_prompt") or edited_prompt

        records.append(
            PieRecord(
                sample_id=sample_id,
                image_path=abs_path,
                relative_path=rel_path,
                original_prompt=original_prompt,
                edited_prompt=edited_prompt,
                edit_instruction=edit_instruction,
            )
        )

    if not records:
        raise FileNotFoundError(f"No valid PIE records found in {mapping_path}.")
    return records


def ensure_dir(path: Path) -> None:
    path.mkdir(parents=True, exist_ok=True)


def copy_file(src: Path, dst: Path, *, overwrite: bool) -> None:
    ensure_dir(dst.parent)
    if overwrite or not dst.exists():
        shutil.copy2(src, dst)


def sync_mapping_file(mapping_path: Path, export_root: Path, dest_relative: str, *, overwrite: bool) -> None:
    dest_path = resolve_path(export_root, dest_relative)
    copy_file(mapping_path, dest_path, overwrite=overwrite)


def save_prediction(image: Image.Image, destination: Path, *, overwrite: bool) -> None:
    ensure_dir(destination.parent)
    if overwrite or not destination.exists():
        image.save(destination)


def main() -> None:
    args = parse_args()

    logging.basicConfig(
        level=logging.INFO,
        format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
    )

    edit_config, seed, precision = load_pipeline_config(args.config)
    edit_config, seed = apply_cli_overrides(args, edit_config, seed)
    component_paths = expand_component_paths(paths_from_model_root(args.model_root))

    precision_choice_raw = args.precision or precision or DEFAULT_PRECISION
    precision_choice = precision_choice_raw.lower()
    if precision_choice != "fp32":
        LOGGER.warning(
            "Precision '%s' requested, but PIE export forces fp32 for numerical stability.",
            precision_choice_raw,
        )
        precision_choice = "fp32"
    torch_dtype = dtype_from_precision(precision_choice)
    compute_dtype = torch.float32

    pie_root = Path(args.pie_root).expanduser().resolve() if args.pie_root else DEFAULT_PIE_ROOT
    export_root = Path(args.export_root).expanduser().resolve() if args.export_root else pie_root
    mapping_path = resolve_path(pie_root, args.mapping_file)

    records = load_pie_records(pie_root, mapping_path, args.image_subdir)
    if args.max_samples is not None:
        records = records[: args.max_samples]

    if not records:
        LOGGER.error("No PIE records to process. Check dataset paths.")
        return

    LOGGER.info(
        "Loaded %d PIE samples from %s (mapping=%s)",
        len(records),
        pie_root,
        mapping_path,
    )
    LOGGER.info("Seed %s | Edit config %s", seed, edit_config)

    pipeline = ChordEditPipeline.from_local_weights(
        component_paths=component_paths,
        default_edit_config=edit_config,
        device=args.device,
        torch_dtype=torch_dtype,
        image_size=args.image_size,
        use_center_crop=args.center_crop,
        compute_dtype=compute_dtype,
        use_attention_mask=args.use_attention_mask,
        use_safety_checker=args.use_safety_checker,
    )

    output_dir = export_root / "output" / args.method_name / args.output_subdir
    source_dir = export_root / "data" / args.source_subdir
    ensure_dir(output_dir)
    if args.copy_source:
        ensure_dir(source_dir)

    if not args.no_sync_mapping:
        sync_mapping_file(mapping_path, export_root, args.mapping_dest, overwrite=args.overwrite)
        LOGGER.info("Synchronized mapping file to %s", resolve_path(export_root, args.mapping_dest))

    processed = 0
    skipped = 0

    for idx, record in enumerate(records, start=1):
        rel_output_path = output_dir / record.relative_path
        if rel_output_path.exists() and not args.overwrite:
            skipped += 1
            continue

        try:
            with Image.open(record.image_path) as img:
                source_image = img.convert("RGB")
        except Exception as exc:  # pragma: no cover - defensive
            LOGGER.error("Failed to read %s: %s", record.image_path, exc)
            skipped += 1
            continue

        try:
            result = pipeline(
                image=source_image,
                source_prompt=record.original_prompt,
                target_prompt=record.edited_prompt,
                seed=seed,
                output_type="pil",
            )
        except Exception as exc:  # pragma: no cover - runtime safety
            LOGGER.error("Pipeline failed on %s: %s", record.sample_id, exc)
            skipped += 1
            continue

        images = result.images
        if isinstance(images, list) and images:
            generated = images[0]
        elif torch.is_tensor(images):
            # Fall back to tensor output if requested differently.
            generated = pipeline._tensor_to_pil(images)[0]  # type: ignore[attr-defined]
        else:
            LOGGER.warning("No images returned for sample %s; skipping.", record.sample_id)
            skipped += 1
            continue

        save_prediction(generated, rel_output_path, overwrite=args.overwrite)

        if args.copy_source:
            target_source_path = source_dir / record.relative_path
            copy_file(record.image_path, target_source_path, overwrite=args.overwrite)

        processed += 1
        if args.log_every and processed % args.log_every == 0:
            LOGGER.info("Saved %d/%d samples (skipped=%d)", processed, len(records), skipped)

    LOGGER.info(
        "Finished PIE export. Saved %d sample(s), skipped %d (existing/errors). Results: %s",
        processed,
        skipped,
        output_dir,
    )


if __name__ == "__main__":
    main()