preprocess.py

#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
"""Recreate the candidate single-exercise clips used by FitAQA.

This standalone implementation follows the public QEVD annotation layout and
the preprocessing procedure described by FitAQA. The official Qualcomm
FitCoach loader was consulted to understand the source annotation format:
https://github.com/Qualcomm-AI-research/FitCoach
"""

from __future__ import annotations

import argparse
import json
import shutil
import subprocess
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Any, Sequence

import numpy as np


DEFAULT_SPLITS = (
    "QEVD-FIT-COACH",
    "QEVD-FIT-COACH-Benchmark",
    "QEVD-FIT-COACH-Competition",
)


@dataclass(frozen=True)
class FeedbackSpan:
    text: str
    start_frame: int
    end_frame: int


@dataclass(frozen=True)
class Segment:
    dataset: str
    output_name: str
    source_video: str
    source_timestamps: str
    exercise_name: str
    start_frame: int
    end_frame: int
    start_seconds: float
    end_seconds: float
    fps: float
    feedback_count: int


def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--input-root",
        type=Path,
        required=True,
        help="Directory containing the extracted QEVD split directories.",
    )
    parser.add_argument(
        "--output-root",
        type=Path,
        required=True,
        help="Directory in which split-specific clips and manifests are written.",
    )
    parser.add_argument(
        "--splits",
        nargs="+",
        choices=DEFAULT_SPLITS,
        default=list(DEFAULT_SPLITS),
        help="QEVD splits to process (defaults to all three long-video sources).",
    )
    parser.add_argument(
        "--ffmpeg",
        default="ffmpeg",
        help="ffmpeg executable name or path (default: ffmpeg).",
    )
    parser.add_argument(
        "--dry-run",
        action="store_true",
        help="Create segment manifests without reading or encoding the videos.",
    )
    parser.add_argument(
        "--overwrite",
        action="store_true",
        help="Replace output clips that already exist.",
    )
    return parser.parse_args(argv)


def feedback_spans(dense_feedbacks: Sequence[Any]) -> list[FeedbackSpan]:
    """Collapse frame-aligned feedback strings into non-empty contiguous runs."""
    spans: list[FeedbackSpan] = []
    active_text: str | None = None
    active_start = 0

    for frame_index, raw_feedback in enumerate(dense_feedbacks):
        text = str(raw_feedback or "")
        if active_text is not None and text != active_text:
            spans.append(FeedbackSpan(active_text, active_start, frame_index - 1))
            active_text = None
        if text and active_text is None:
            active_text = text
            active_start = frame_index

    if active_text is not None:
        spans.append(
            FeedbackSpan(active_text, active_start, len(dense_feedbacks) - 1)
        )
    return spans


def load_frame_timestamps(path: Path) -> np.ndarray:
    """Load the official per-frame UNIX timestamps and convert ns to seconds."""
    raw = np.asarray(np.load(path), dtype=np.float64)
    if raw.ndim != 1 or len(raw) < 2:
        raise ValueError(f"Expected a one-dimensional timestamp array: {path}")
    seconds = raw / 1_000_000_000.0
    if not np.all(np.diff(seconds) > 0):
        raise ValueError(f"Frame timestamps must be strictly increasing: {path}")
    return seconds


def effective_fps(timestamps: np.ndarray) -> float:
    duration = float(timestamps[-1] - timestamps[0])
    if duration <= 0:
        raise ValueError("Timestamp duration must be positive")
    return float(len(timestamps) / duration)


def exercise_from_transition(text: str) -> str:
    for prefix in ("First up are ", "Moving on to "):
        if text.startswith(prefix):
            text = text[len(prefix) :]
            break
    return text.rstrip("!").strip()


def segments_for_record(
    *,
    dataset: str,
    record: dict[str, Any],
    timestamps: np.ndarray,
) -> list[Segment]:
    dense_feedbacks = record.get("feedbacks")
    sparse_timestamps = record.get("feedback_timestamps")
    transition_flags = record.get("is_transition")
    if not isinstance(dense_feedbacks, list):
        raise ValueError("feedbacks must be a list")
    if not isinstance(sparse_timestamps, list):
        raise ValueError("feedback_timestamps must be a list")
    if not isinstance(transition_flags, list):
        raise ValueError("is_transition must be a list")
    if len(dense_feedbacks) != len(timestamps):
        raise ValueError(
            "The dense feedback sequence and frame timestamp array must have equal length"
        )

    spans = feedback_spans(dense_feedbacks)
    if len(spans) != len(sparse_timestamps):
        raise ValueError(
            f"Recovered {len(spans)} feedback spans but found "
            f"{len(sparse_timestamps)} feedback timestamps"
        )
    if len(transition_flags) != len(spans):
        raise ValueError(
            f"Found {len(transition_flags)} transition flags for {len(spans)} spans"
        )

    transition_positions = [
        index for index, is_transition in enumerate(transition_flags) if is_transition
    ]
    if len(transition_positions) < 2:
        raise ValueError("At least two transition annotations are required")

    source_video = str(record.get("long_range_video_file") or "")
    source_timestamps = str(record.get("video_timestamps") or "")
    if not source_video or not source_timestamps:
        raise ValueError("The record must reference a video and timestamp file")
    video_stem = Path(source_video).stem
    fps = effective_fps(timestamps)
    segments: list[Segment] = []

    for current_position, next_position in zip(
        transition_positions, transition_positions[1:]
    ):
        feedback_count = next_position - current_position - 1
        if feedback_count < 1:
            continue

        current_transition = spans[current_position]
        next_transition = spans[next_position]
        start_frame = current_transition.start_frame
        end_frame = next_transition.start_frame + (
            next_transition.end_frame - next_transition.start_frame
        ) // 2
        if not 0 <= start_frame <= end_frame < len(timestamps):
            raise ValueError(
                f"Invalid inclusive frame range [{start_frame}, {end_frame}]"
            )

        segment_index = len(segments)
        segments.append(
            Segment(
                dataset=dataset,
                output_name=f"{video_stem}_{segment_index:02d}.mp4",
                source_video=source_video,
                source_timestamps=source_timestamps,
                exercise_name=exercise_from_transition(current_transition.text),
                start_frame=start_frame,
                end_frame=end_frame,
                start_seconds=float(timestamps[start_frame] - timestamps[0]),
                end_seconds=float(timestamps[end_frame] - timestamps[0]),
                fps=fps,
                feedback_count=feedback_count,
            )
        )
    return segments


def ffmpeg_command(
    *, executable: str, source: Path, output: Path, segment: Segment
) -> list[str]:
    frame_filter = (
        f"select=between(n\\,{segment.start_frame}\\,{segment.end_frame}),"
        f"setpts=N/({segment.fps:.12f}*TB)"
    )
    return [
        executable,
        "-hide_banner",
        "-loglevel",
        "error",
        "-i",
        str(source),
        "-vf",
        frame_filter,
        "-an",
        "-c:v",
        "libx264",
        "-preset",
        "slower",
        "-crf",
        "17",
        "-profile:v",
        "high",
        "-pix_fmt",
        "yuv420p",
        "-movflags",
        "+faststart",
        "-fps_mode",
        "cfr",
        "-r",
        f"{segment.fps:.12f}",
        "-y",
        str(output),
    ]


def load_records(path: Path) -> list[dict[str, Any]]:
    with path.open("r", encoding="utf-8") as handle:
        records = json.load(handle)
    if not isinstance(records, list) or not all(
        isinstance(record, dict) for record in records
    ):
        raise ValueError(f"Expected a JSON array of objects: {path}")
    return records


def process_split(args: argparse.Namespace, split: str) -> int:
    split_root = args.input_root / split
    records_path = split_root / "feedbacks_long_range.json"
    records = load_records(records_path)
    output_dir = args.output_root / split
    output_dir.mkdir(parents=True, exist_ok=True)

    manifest_rows: list[dict[str, Any]] = []
    for record_index, record in enumerate(records):
        try:
            timestamp_path = split_root / str(record["video_timestamps"])
            timestamps = load_frame_timestamps(timestamp_path)
            segments = segments_for_record(
                dataset=split, record=record, timestamps=timestamps
            )
            for segment in segments:
                output_path = output_dir / segment.output_name
                manifest_row = asdict(segment)
                manifest_row["output_path"] = str(
                    Path(split) / segment.output_name
                )
                manifest_rows.append(manifest_row)

                if args.dry_run or (output_path.exists() and not args.overwrite):
                    continue
                source_path = split_root / segment.source_video
                if not source_path.is_file():
                    raise FileNotFoundError(source_path)
                subprocess.run(
                    ffmpeg_command(
                        executable=args.ffmpeg,
                        source=source_path,
                        output=output_path,
                        segment=segment,
                    ),
                    check=True,
                )
        except Exception as exc:
            source = record.get("long_range_video_file", f"record {record_index}")
            raise RuntimeError(f"{split}: failed to process {source}") from exc

    manifest_path = output_dir / "segments.jsonl"
    with manifest_path.open("w", encoding="utf-8") as handle:
        for row in manifest_rows:
            handle.write(json.dumps(row, ensure_ascii=False) + "\n")
    print(f"{split}: described {len(manifest_rows):,} candidate segments")
    return len(manifest_rows)


def main(argv: Sequence[str] | None = None) -> None:
    args = parse_args(argv)
    args.input_root = args.input_root.resolve()
    args.output_root = args.output_root.resolve()
    if not args.dry_run and shutil.which(args.ffmpeg) is None:
        raise RuntimeError(f"ffmpeg executable not found: {args.ffmpeg}")

    total = sum(process_split(args, split) for split in args.splits)
    print(f"Done: described {total:,} candidate single-exercise segments")


if __name__ == "__main__":
    main()