pcmt/cli.py

from __future__ import annotations

import argparse
import json
import sys
from pathlib import Path

from .audit import corrupted_detector_drift, radius_ablation, replay_audit
from .datasets import DATASET_NAMES, adapter_status
from .edits import EDIT_TYPES
from .pipeline import RunConfig, run_pipeline, run_pipeline_for_clips, validate_csv
from .plots import make_figures
from .queue import add_job, init_queue, read_jobs, run_next
from .real_plots import make_real_video_figures
from .real_video import download_real_videos, load_real_video_clips, write_real_feature_tables
from .detectors import detector_clips
from .hf_benchmarks import download_hf_benchmark_videos
from .robustness import make_robustness_figures, run_robustness_atlas
from .sweeps import run_sweep
from .sweep_plots import make_sweep_figures


def main(argv: list[str] | None = None) -> int:
    parser = argparse.ArgumentParser(description="Proof-carrying multimodal timelines")
    sub = parser.add_subparsers(dest="command", required=True)

    smoke = sub.add_parser("smoke", help="run minimal end-to-end pipeline")
    _add_run_args(smoke, default_output="runs/smoke", default_clips=6, default_edits="none,audio_shift,crop", default_repeats=1)

    small = sub.add_parser("run-small", help="run small defect atlas")
    _add_run_args(small, default_output="runs/run-small", default_clips=10, default_edits=",".join(EDIT_TYPES), default_repeats=4)

    real_data = sub.add_parser("download-real", help="download small public real MP4s")
    real_data.add_argument("--output", type=Path, default=Path("data/real_videos"))

    hf_data = sub.add_parser("download-benchmark", help="download a bounded public HF benchmark video subset")
    hf_data.add_argument("--dataset", choices=["youcook2", "ave"], required=True)
    hf_data.add_argument("--output", type=Path, required=True)
    hf_data.add_argument("--limit", type=int, default=80)
    hf_data.add_argument("--force", action="store_true")

    real = sub.add_parser("run-real", help="run certificates on decoded real MP4s")
    _add_run_args(real, default_output="runs/real-video", default_clips=15, default_edits=",".join(EDIT_TYPES), default_repeats=128)
    real.add_argument("--video-root", type=Path, default=Path("data/real_videos"))
    real.add_argument("--feature-output", type=Path, default=None)

    detector = sub.add_parser("run-detector", help="run CLIP/AST detector atoms on real MP4s")
    _add_run_args(detector, default_output="runs/detector-video", default_clips=15, default_edits=",".join(EDIT_TYPES), default_repeats=512)
    detector.add_argument("--video-root", type=Path, default=Path("data/real_videos"))
    detector.add_argument("--detector-output", type=Path, default=None)

    sweep = sub.add_parser("run-sweep", help="run dense parameterized perturbation sweeps")
    sweep.add_argument("--input-run", choices=["real", "detector"], default="detector")
    sweep.add_argument("--output", type=Path, default=Path("runs/sweep-detector"))
    sweep.add_argument("--video-root", type=Path, default=Path("data/real_videos"))
    sweep.add_argument("--clips", type=int, default=15)
    sweep.add_argument("--use-gpu", default="auto", choices=["auto", "cpu", "cuda", "gpu"])
    sweep.add_argument("--stress-repeats", type=int, default=512)

    figures = sub.add_parser("figures", help="create manuscript figures")
    figures.add_argument("--input", type=Path, default=Path("runs/run-small/traces.csv"))
    figures.add_argument("--output", type=Path, default=Path("figures"))

    real_figures = sub.add_parser("real-figures", help="create real-video frame/timeline figures")
    real_figures.add_argument("--video", type=Path, default=Path("data/real_videos/sample-20s-360p.mp4"))
    real_figures.add_argument("--traces", type=Path, default=Path("runs/real-video/traces.csv"))
    real_figures.add_argument("--output", type=Path, default=Path("figures"))
    real_figures.add_argument("--use-gpu", default="auto", choices=["auto", "cpu", "cuda", "gpu"])

    sweep_figures = sub.add_parser("sweep-figures", help="create sweep heatmap and perturbation figures")
    sweep_figures.add_argument("--traces", type=Path, default=Path("runs/sweep-detector/sweep_traces.csv"))
    sweep_figures.add_argument("--metrics", type=Path, default=Path("runs/sweep-detector/sweep_metrics.json"))
    sweep_figures.add_argument("--output", type=Path, default=Path("figures"))

    validate = sub.add_parser("validate", help="validate a trace CSV")
    validate.add_argument("--input", type=Path, required=True)

    status = sub.add_parser("adapter-status", help="report local real dataset adapter status")
    status.add_argument("--root", type=Path, default=Path("data/raw"))

    qi = sub.add_parser("queue-init", help="initialize GPU queue")
    qi.add_argument("--path", type=Path, default=Path("runs/gpu_jobs.jsonl"))

    qa = sub.add_parser("queue-add", help="add a GPU queue job")
    qa.add_argument("--path", type=Path, default=Path("runs/gpu_jobs.jsonl"))
    qa.add_argument("--name", required=True)
    qa.add_argument("--command", dest="job_command", required=True)
    qa.add_argument("--cwd", default=".")

    ql = sub.add_parser("queue-list", help="list GPU queue jobs")
    ql.add_argument("--path", type=Path, default=Path("runs/gpu_jobs.jsonl"))

    qr = sub.add_parser("queue-run-next", help="run the next queued job")
    qr.add_argument("--path", type=Path, default=Path("runs/gpu_jobs.jsonl"))
    qr.add_argument("--log-dir", type=Path, default=Path("runs/gpu_queue_logs"))

    replay = sub.add_parser("replay-audit", help="independently replay certificates for a run directory")
    replay.add_argument("--run", type=Path, default=Path("runs/run-small"))
    replay.add_argument("--output", type=Path)

    radius = sub.add_parser("radius-ablation", help="evaluate formulas across radius settings from windows.jsonl")
    radius.add_argument("--run", type=Path, default=Path("runs/run-small"))
    radius.add_argument("--output", type=Path, default=Path("runs/run-small/radius_ablation.csv"))
    radius.add_argument("--radii", default="0,1,2,4,8")

    drift = sub.add_parser("corrupted-drift", help="compare oracle-like and corrupted atom traces")
    drift.add_argument("--run", type=Path, default=Path("runs/run-small"))
    drift.add_argument("--output", type=Path, default=Path("runs/run-small/corrupted_detector_drift.csv"))

    robust = sub.add_parser("robustness-atlas", help="aggregate detector-threshold/radius certificate stability atlas")
    robust.add_argument("--run", type=Path, default=Path("runs/run-small"))
    robust.add_argument("--output", type=Path, default=Path("runs/robustness-atlas"))
    robust.add_argument("--thresholds", type=int, default=64)
    robust.add_argument("--radii", default="0,1,2,3,4,5,6,7,8")
    robust.add_argument("--use-gpu", default="auto", choices=["auto", "cpu", "cuda", "gpu"])
    robust.add_argument("--sample-limit", type=int, default=None)
    robust.add_argument("--unstable-limit", type=int, default=200)

    robust_fig = sub.add_parser("robustness-figures", help="create certificate stability phase-diagram figure")
    robust_fig.add_argument("--input", type=Path, default=Path("runs/robustness-atlas"))
    robust_fig.add_argument("--output", type=Path, default=Path("figures"))

    args = parser.parse_args(argv)
    if args.command in {"smoke", "run-small"}:
        metrics = _run_from_args(args)
        print(json.dumps(metrics, indent=2, sort_keys=True))
        return 0
    if args.command == "download-real":
        paths = download_real_videos(args.output)
        print(json.dumps([str(path) for path in paths], indent=2))
        return 0
    if args.command == "download-benchmark":
        paths = download_hf_benchmark_videos(args.dataset, args.output, limit=args.limit, force=args.force)
        print(json.dumps([str(path) for path in paths], indent=2))
        return 0
    if args.command == "run-real":
        edits = [edit.strip() for edit in args.edits.split(",") if edit.strip()]
        clips = load_real_video_clips(args.video_root, limit=args.clips, use_gpu=args.use_gpu)
        paths = [clip.media_path for clip in clips if clip.media_path is not None]
        feature_output = args.feature_output or (args.output / "features")
        write_real_feature_tables(paths, feature_output, use_gpu=args.use_gpu)
        config = RunConfig(
            output=args.output,
            data_root=args.data_root,
            fixture_root=args.fixture_root,
            datasets=["real_public_mp4"],
            clips=args.clips,
            edits=edits,
            use_gpu=args.use_gpu,
            stress_repeats=max(1, args.stress_repeats),
        )
        metrics = run_pipeline_for_clips(config, clips)
        print(json.dumps(metrics, indent=2, sort_keys=True))
        return 0
    if args.command == "run-detector":
        edits = [edit.strip() for edit in args.edits.split(",") if edit.strip()]
        detector_output = args.detector_output or (args.output / "detector")
        clips = detector_clips(args.video_root, detector_output, limit=args.clips, use_gpu=args.use_gpu)
        config = RunConfig(
            output=args.output,
            data_root=args.data_root,
            fixture_root=args.fixture_root,
            datasets=["real_public_mp4_detector"],
            clips=args.clips,
            edits=edits,
            use_gpu=args.use_gpu,
            stress_repeats=max(1, args.stress_repeats),
        )
        metrics = run_pipeline_for_clips(config, clips)
        print(json.dumps(metrics, indent=2, sort_keys=True))
        return 0
    if args.command == "run-sweep":
        if args.input_run == "detector":
            clips = detector_clips(args.video_root, args.output / "detector", limit=args.clips, use_gpu=args.use_gpu)
        else:
            clips = load_real_video_clips(args.video_root, limit=args.clips, use_gpu=args.use_gpu)
        metrics = run_sweep(clips, args.output, use_gpu=args.use_gpu, repeats=max(1, args.stress_repeats))
        print(json.dumps(metrics, indent=2, sort_keys=True))
        return 0
    if args.command == "figures":
        paths = make_figures(args.input, args.output)
        print(json.dumps(paths, indent=2, sort_keys=True))
        return 0
    if args.command == "real-figures":
        paths = make_real_video_figures(args.video, args.traces, args.output, use_gpu=args.use_gpu)
        print(json.dumps(paths, indent=2, sort_keys=True))
        return 0
    if args.command == "sweep-figures":
        paths = make_sweep_figures(args.traces, args.metrics, args.output)
        print(json.dumps(paths, indent=2, sort_keys=True))
        return 0
    if args.command == "validate":
        errors = validate_csv(args.input)
        if errors:
            for error in errors:
                print(error, file=sys.stderr)
            return 1
        print(f"{args.input} valid")
        return 0
    if args.command == "adapter-status":
        print(json.dumps(adapter_status(args.root), indent=2, sort_keys=True))
        return 0
    if args.command == "queue-init":
        init_queue(args.path)
        print(args.path)
        return 0
    if args.command == "queue-add":
        job = add_job(args.path, args.name, args.job_command, cwd=args.cwd)
        print(json.dumps(job, indent=2, sort_keys=True))
        return 0
    if args.command == "queue-list":
        print(json.dumps(read_jobs(args.path), indent=2, sort_keys=True))
        return 0
    if args.command == "queue-run-next":
        job = run_next(args.path, args.log_dir)
        print(json.dumps(job, indent=2, sort_keys=True))
        return 0
    if args.command == "replay-audit":
        print(json.dumps(replay_audit(args.run, args.output), indent=2, sort_keys=True))
        return 0
    if args.command == "radius-ablation":
        radii = [int(item.strip()) for item in args.radii.split(",") if item.strip()]
        print(json.dumps(radius_ablation(args.run, args.output, radii), indent=2, sort_keys=True))
        return 0
    if args.command == "corrupted-drift":
        print(json.dumps(corrupted_detector_drift(args.run, args.output), indent=2, sort_keys=True))
        return 0
    if args.command == "robustness-atlas":
        radii = [int(item.strip()) for item in args.radii.split(",") if item.strip()]
        metrics = run_robustness_atlas(
            args.run,
            args.output,
            thresholds=args.thresholds,
            radii=radii,
            use_gpu=args.use_gpu,
            sample_limit=args.sample_limit,
            unstable_limit=args.unstable_limit,
        )
        print(json.dumps(metrics, indent=2, sort_keys=True))
        return 0
    if args.command == "robustness-figures":
        print(json.dumps(make_robustness_figures(args.input, args.output), indent=2, sort_keys=True))
        return 0
    raise AssertionError(args.command)


def _add_run_args(parser: argparse.ArgumentParser, default_output: str, default_clips: int, default_edits: str, default_repeats: int) -> None:
    parser.add_argument("--output", type=Path, default=Path(default_output))
    parser.add_argument("--data-root", type=Path, default=Path("data/raw"))
    parser.add_argument("--fixture-root", type=Path, default=Path("data/fixtures"))
    parser.add_argument("--datasets", default=",".join(DATASET_NAMES))
    parser.add_argument("--clips", type=int, default=default_clips)
    parser.add_argument("--edits", default=default_edits)
    parser.add_argument("--use-gpu", default="auto", choices=["auto", "cpu", "cuda", "gpu"])
    parser.add_argument("--stress-repeats", type=int, default=default_repeats)


def _run_from_args(args: argparse.Namespace) -> dict:
    edits = [edit.strip() for edit in args.edits.split(",") if edit.strip()]
    datasets = [dataset.strip() for dataset in args.datasets.split(",") if dataset.strip()]
    config = RunConfig(
        output=args.output,
        data_root=args.data_root,
        fixture_root=args.fixture_root,
        datasets=datasets,
        clips=args.clips,
        edits=edits,
        use_gpu=args.use_gpu,
        stress_repeats=max(1, args.stress_repeats),
    )
    return run_pipeline(config)


if __name__ == "__main__":
    raise SystemExit(main())