#!/usr/bin/env python3
"""A PyTorch Dataset for VQS-4K.

    from vqs4k import VQS4K

    ds = VQS4K("VQS-4K", split="test")        # a folder made by download.py
    s  = ds[0]
    s["query_image"], s["query_mask"]         # PIL: the external visual query
    s["frames"], s["masks"]                   # lists over the search video
    s["present"]                              # bool per frame: is the target visible?

`masks[i]` is None exactly where `present[i]` is False. That is ground truth, not missing
data: the correct prediction on those frames is an empty mask.

Needs: pip install torch pillow
"""
from pathlib import Path

from PIL import Image
from torch.utils.data import Dataset


class VQS4K(Dataset):
    def __init__(self, root, split="all", protocol="tiou", transform=None):
        self.root = Path(root)
        self.transform = transform

        if split == "all":
            self.seqs = sorted(p.name for p in (self.root / "images").iterdir() if p.is_dir())
        else:
            suffix = "" if protocol == "tiou" else "_by_class"
            listing = self.root / "splits" / f"{split}_t-IoU{suffix}.txt"
            names = [l.strip() for l in listing.read_text().splitlines() if l.strip()]
            # keep only what is actually on disk, so partial downloads still work
            self.seqs = [s for s in names if (self.root / "images" / s).is_dir()]

        if not self.seqs:
            raise RuntimeError(f"no sequences found under {self.root}")

    def __len__(self):
        return len(self.seqs)

    def __getitem__(self, i):
        seq = self.seqs[i]
        img_dir = self.root / "images" / seq
        msk_dir = self.root / "masks" / seq

        # 00000 is the query; everything after it is the search video
        indices = sorted(int(p.stem) for p in img_dir.glob("*.jpg"))
        search = [n for n in indices if n != 0]

        frames, masks, present = [], [], []
        for n in search:
            frames.append(Image.open(img_dir / f"{n:05d}.jpg").convert("RGB"))
            mp = msk_dir / f"{n:05d}.png"
            hit = mp.exists()
            present.append(hit)
            masks.append(Image.open(mp).convert("L") if hit else None)

        sample = {
            "seq_id": seq,
            "category_id": seq.rsplit("-set_", 1)[0],
            "query_image": Image.open(img_dir / "00000.jpg").convert("RGB"),
            "query_mask": Image.open(msk_dir / "00000.png").convert("L"),
            "frames": frames,
            "masks": masks,
            "present": present,
            "frame_indices": search,
        }
        return self.transform(sample) if self.transform else sample


if __name__ == "__main__":
    import sys

    ds = VQS4K(sys.argv[1] if len(sys.argv) > 1 else "VQS-4K")
    s = ds[0]
    print(f"{len(ds)} sequences; first = {s['seq_id']}")
    print(f"  query {s['query_image'].size}, {len(s['frames'])} frames, "
          f"{sum(s['present'])} with the target visible")