scripts/batch_depth_5k.py

"""Batch depth for coaf_dataset_24_25 — input is 25-frame rgb_align/."""

import json
import os
import sys
import time
from pathlib import Path

import imageio
import numpy as np
import torch

VDA_ROOT = Path(
    "/project/llmsvgen/sunkai/minghao/week7/week7-video_depth_anything/Video-Depth-Anything"
)
sys.path.insert(0, str(VDA_ROOT))
os.chdir(str(VDA_ROOT))

from utils.dc_utils import read_video_frames, save_video
from video_depth_anything.video_depth import VideoDepthAnything

DATASET_ROOT = Path("/project/llmsvgen/sunkai/robomaster_3d/Casual_CoAF/coaf_dataset_24_25")
RAW_ROOT = DATASET_ROOT / "raw"
OUTPUT_ROOT = DATASET_ROOT / "modalities" / "depth"
TMP_DIR = Path("/tmp/depth_tmp_videos_24_25")
RGB_ALIGN_FRAMES = 25

ENCODER = "vitl"
INPUT_SIZE = 518
MAX_RES = 1280
FPS = 8


def build_model(encoder, device):
    model_configs = {
        "vits": {"encoder": "vits", "features": 64, "out_channels": [48, 96, 192, 384]},
        "vitb": {"encoder": "vitb", "features": 128, "out_channels": [96, 192, 384, 768]},
        "vitl": {"encoder": "vitl", "features": 256, "out_channels": [256, 512, 1024, 1024]},
    }
    checkpoint_path = f"./checkpoints/video_depth_anything_{encoder}.pth"
    if not os.path.isfile(checkpoint_path):
        raise FileNotFoundError(f"Checkpoint not found: {checkpoint_path}")
    model = VideoDepthAnything(**model_configs[encoder], metric=False)
    state_dict = torch.load(checkpoint_path, map_location="cpu")
    model.load_state_dict(state_dict, strict=True)
    return model.to(device).eval()


def frames_to_tmp_video(rgb_dir, tmp_path, num_frames=RGB_ALIGN_FRAMES, fps=8):
    frames = []
    for i in range(1, num_frames + 1):
        path = rgb_dir / f"frame_{i:04d}.png"
        if not path.exists():
            break
        frames.append(imageio.imread(str(path)))
    if len(frames) != num_frames:
        raise ValueError(f"Expected {num_frames} frames in {rgb_dir}, got {len(frames)}")
    imageio.mimsave(str(tmp_path), frames, fps=fps, codec="libx264", macro_block_size=1)
    return len(frames)


def main():
    import argparse

    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--start", type=int, default=0, help="First episode index (inclusive)")
    parser.add_argument("--stop", type=int, default=5000, help="Last episode index (exclusive)")
    parser.add_argument("--skip-existing", action="store_true", default=True)
    parser.add_argument("--no-skip-existing", dest="skip_existing", action="store_false")
    args = parser.parse_args()

    if not torch.cuda.is_available():
        raise RuntimeError(
            "CUDA GPU required for Video Depth Anything (xformers attention). "
            "Run via sbatch on a GPU node, not the login node."
        )
    device = "cuda"
    print(f"Device: {device} ({torch.cuda.get_device_name(0)})")
    model = build_model(ENCODER, device)

    TMP_DIR.mkdir(parents=True, exist_ok=True)
    OUTPUT_ROOT.mkdir(parents=True, exist_ok=True)

    episodes = sorted(RAW_ROOT.glob("episode_*"))
    episodes = [
        ep
        for ep in episodes
        if args.start <= int(ep.name.split("_")[-1]) < args.stop
    ]
    print(
        f"Processing {len(episodes)} episodes idx [{args.start}, {args.stop}) "
        f"(rgb_align -> depth, {RGB_ALIGN_FRAMES} frames)"
    )

    start_time = time.time()
    processed = 0
    failed = []

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

        rgb_dir = ep_dir / "rgb_align"
        if not rgb_dir.exists():
            failed.append({"episode": ep_name, "error": "rgb_align dir not found"})
            continue

        tmp_video = TMP_DIR / f"{ep_name}.mp4"
        try:
            frames_to_tmp_video(rgb_dir, tmp_video, num_frames=RGB_ALIGN_FRAMES, fps=FPS)
            frames, target_fps = read_video_frames(str(tmp_video), -1, -1, MAX_RES)
            depths, fps = model.infer_video_depth(
                frames, target_fps, input_size=INPUT_SIZE, device=device, fp32=False
            )
            out_dir.mkdir(parents=True, exist_ok=True)
            save_video(depths, str(out_dir / "depth.mp4"), fps=fps, is_depths=True, grayscale=False)
            processed += 1
            if processed % 200 == 0:
                elapsed = time.time() - start_time
                eps = processed / elapsed
                remaining = (len(episodes) - processed) / max(eps, 0.01)
                print(f"  [{processed}/{len(episodes)}] ~{remaining:.0f}s remaining")
        except Exception as e:
            failed.append({"episode": ep_name, "error": str(e)})
        finally:
            if tmp_video.exists():
                tmp_video.unlink()

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


if __name__ == "__main__":
    main()