scripts/batch_flow_5k.py

"""Batch optical flow from 25-frame rgb_align/ (24 flow vis frames)."""

import json
import sys
import time
from pathlib import Path

import cv2
import imageio
import numpy as np
import torch

sys.path.insert(0, "/project/llmsvgen/sunkai/robomaster_3d/CoAF")
from tools.flow_dataset.io import save_vis_png
from tools.flow_dataset.visualize import flow_hwc_to_colorwheel_bgr
from tools.unimatch_flow.model import build_unimatch_estimator

DATASET_ROOT = Path("/project/llmsvgen/sunkai/robomaster_3d/Casual_CoAF/coaf_dataset_24_25")
RAW_ROOT = DATASET_ROOT / "raw"
OUTPUT_ROOT = DATASET_ROOT / "modalities" / "flow"
RGB_ALIGN_FRAMES = 25
GLOBAL_MAX_FLOW = 20.0
WIDTH = 512
HEIGHT = 512
FPS = 8


def read_rgb_align_frames(rgb_dir: Path, target_size: int):
    frames = []
    for i in range(1, RGB_ALIGN_FRAMES + 1):
        path = rgb_dir / f"frame_{i:04d}.png"
        if not path.exists():
            break
        img = cv2.imread(str(path))
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        if img.shape[0] != target_size or img.shape[1] != target_size:
            img = cv2.resize(img, (target_size, target_size), interpolation=cv2.INTER_LANCZOS4)
        frames.append(img)
    return frames


def main():
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    estimator = build_unimatch_estimator(
        device,
        preset_name="gmflow_s2_reg6_mixdata",
        height=HEIGHT,
        width=WIDTH,
    )

    episodes = sorted(RAW_ROOT.glob("episode_*"))
    start_time = time.time()
    processed = 0
    failed = []

    for ep_dir in episodes:
        ep_name = ep_dir.name
        out_dir = OUTPUT_ROOT / ep_name
        if (out_dir / "preview.mp4").exists():
            processed += 1
            continue

        rgb_dir = ep_dir / "rgb_align"
        try:
            frames = read_rgb_align_frames(rgb_dir, WIDTH)
            if len(frames) != RGB_ALIGN_FRAMES:
                raise ValueError(f"Expected {RGB_ALIGN_FRAMES} frames, got {len(frames)}")

            raw_dir = out_dir / "raw"
            vis_dir = out_dir / "vis"
            raw_dir.mkdir(parents=True, exist_ok=True)
            vis_dir.mkdir(parents=True, exist_ok=True)

            # 25 flow frames aligned with reason index k (k=0: zero flow at episode start)
            flow_frames_vis = []
            for i in range(RGB_ALIGN_FRAMES):
                if i == 0:
                    flow = np.zeros((HEIGHT, WIDTH, 2), dtype=np.float32)
                else:
                    flow = estimator.predict_hwc(frames[i - 1], frames[i])
                np.save(str(raw_dir / f"flow_{i+1:04d}.npy"), flow)
                vis = flow_hwc_to_colorwheel_bgr(flow, GLOBAL_MAX_FLOW)
                cv2.imwrite(str(vis_dir / f"frame_{i+1:04d}.png"), vis)
                flow_frames_vis.append(cv2.cvtColor(vis, cv2.COLOR_BGR2RGB))

            imageio.mimsave(
                str(out_dir / "preview.mp4"),
                flow_frames_vis,
                fps=FPS,
                codec="libx264",
                macro_block_size=1,
            )
            meta = {
                "num_input_frames": RGB_ALIGN_FRAMES,
                "num_flow_frames": len(flow_frames_vis),
                "aligned_with": "reason_indices",
            }
            (out_dir / "meta.json").write_text(json.dumps(meta, indent=2) + "\n")
            processed += 1
        except Exception as e:
            failed.append({"episode": ep_name, "error": str(e)})

    print(f"\nDone! {processed}/{len(episodes)}, {len(failed)} failed")
    if failed:
        (OUTPUT_ROOT / "flow_failures.json").write_text(json.dumps(failed, indent=2) + "\n")


if __name__ == "__main__":
    main()