README — Loading things_eeg_2 from nonarjb/alignvis

This repo hosts WebDataset shard sets under things_eeg_2/:

• things_eeg_2-images-*.tar — images
• things_eeg_2-image_embeddings-*.tar — vector embeddings (.npy/.npz)
• things_eeg_2-preprocessed_eeg-*.tar — EEG arrays (.npy/.npz)

Inside each shard, the WebDataset __key__ is the file’s relative path under the top folder (without extension). To reconstruct the original relative path, use:

rel_path = "<top>/" + __key__ + "." + <ext>

(e.g., images/training_images/01133_raincoat/raincoat_01s.jpg)

To use the other dataset (things_meg), just replace dataset_dir="things_eeg_2" with dataset_dir="things_meg" in the examples below.

---

Install

pip install webdataset huggingface_hub pillow torch tqdm
# Optional: faster transfers for big files
pip install -U hf_transfer && export HF_HUB_ENABLE_HF_TRANSFER=1

---

Helper: list shard URLs from the Hub

Create utils_hf_wds.py:

# utils_hf_wds.py
from huggingface_hub import HfFileSystem, hf_hub_url

def hf_tar_urls(repo_id: str, dataset_dir: str, top: str, revision: str = "main"):
    """
    Return sorted 'resolve/<revision>' URLs for shards matching:
        <dataset_dir>/<dataset_dir>-<top>-*.tar
    Example: things_eeg_2/things_eeg_2-images-000000.tar
    """
    fs = HfFileSystem()
    pattern = f"datasets/{repo_id}/{dataset_dir}/{dataset_dir}-{top}-*.tar"
    hf_paths = sorted(fs.glob(pattern))  # hf://datasets/<repo_id>/...
    rel_paths = [p.split(f"datasets/{repo_id}/", 1)[1] for p in hf_paths]
    return [
        hf_hub_url(repo_id, filename=p, repo_type="dataset", revision=revision)
        for p in rel_paths
    ]

---

A) Images (PIL) with original relative paths

import io
from PIL import Image
import torch, webdataset as wds
from utils_hf_wds import hf_tar_urls

REPO = "nonarjb/alignvis"

def make_images_loader(dataset_dir="things_eeg_2", batch_size=16, num_workers=4):
    urls = hf_tar_urls(REPO, dataset_dir, top="images")
    if not urls: raise RuntimeError("No image shards found")

    def pick_image(s):
        for ext in ("jpg","jpeg","png"):
            if ext in s:
                s["img_bytes"] = s[ext]
                s["rel_path"] = f"images/{s['__key__']}.{ext}"
                return s
        return None

    ds = (wds.WebDataset(urls, shardshuffle=False, handler=wds.handlers.warn_and_continue)
            .map(pick_image).select(lambda s: s is not None)
            .map(lambda s: (s["rel_path"], Image.open(io.BytesIO(s["img_bytes"])).convert("RGB"))))

    return torch.utils.data.DataLoader(
        ds, batch_size=batch_size, num_workers=num_workers, collate_fn=lambda b: b
    )

loader = make_images_loader()
rel_path, pil_img = next(iter(loader))[0]
print(rel_path, pil_img.size)  # e.g. images/training_images/.../raincoat_01s.jpg (W, H)

---

B) Image embeddings (.npy/.npz) → torch.Tensor

import io, numpy as np
import torch, webdataset as wds
from utils_hf_wds import hf_tar_urls

REPO = "nonarjb/alignvis"

# Heuristics for dict-like payloads
CANDIDATE_KEYS = ("embedding", "emb", "vector", "feat", "features", "clip", "image", "text")

def _first_numeric_from_npz(npz, prefer_key=None):
    if prefer_key and prefer_key in npz:
        return np.asarray(npz[prefer_key])
    # try direct numeric arrays
    for k in npz.files:
        a = npz[k]
        if isinstance(a, np.ndarray) and np.issubdtype(a.dtype, np.number):
            return a
    # try dict-like entries with known keys
    for k in npz.files:
        a = npz[k]
        if isinstance(a, dict):
            for ck in CANDIDATE_KEYS:
                if ck in a:
                    return np.asarray(a[ck])
    return None

def _load_numeric_vector(payload: bytes, ext: str, prefer_key: str | None = None):
    """Return 1D float32 vector or None if not numeric."""
    bio = io.BytesIO(payload)
    try:
        arr = np.load(bio, allow_pickle=False)
    except ValueError as e:
        if "Object arrays" in str(e):
            bio.seek(0)
            obj = np.load(bio, allow_pickle=True)
            if isinstance(obj, dict):
                for ck in CANDIDATE_KEYS:
                    if ck in obj:
                        arr = obj[ck]; break
                else:
                    return None
            elif isinstance(obj, (list, tuple)):
                arr = np.asarray(obj)
            else:
                return None
        else:
            raise
    arr = np.asarray(arr)
    if not np.issubdtype(arr.dtype, np.number):
        try:
            arr = arr.astype(np.float32)
        except Exception:
            return None
    return arr.reshape(-1).astype(np.float32)

def make_embeddings_loader(
    dataset_dir="things_eeg_2",
    batch_size=64,
    num_workers=4,
    prefer_key: str | None = None,   # e.g., "embedding" if you know the field name
):
    urls = hf_tar_urls(REPO, dataset_dir, top="image_embeddings")
    if not urls:
        raise RuntimeError("No embedding shards found")

    def pick_payload(s):
        for ext in ("npy", "npz"):
            if ext in s:
                s["__ext__"] = ext
                s["payload"] = s[ext]
                s["rel_path"] = f"image_embeddings/{s['__key__']}.{ext}"
                return s
        return None

    def decode_vec(s):
        vec = _load_numeric_vector(s["payload"], s["__ext__"], prefer_key=prefer_key)
        if vec is None:
            # skip non-numeric payloads
            return None
        return (s["rel_path"], torch.from_numpy(vec))

    ds = (
        wds.WebDataset(urls, shardshuffle=False, handler=wds.handlers.warn_and_continue)
        .map(pick_payload).select(lambda s: s is not None)
        .map(decode_vec).select(lambda x: x is not None)
    )

    # Collate into a batch tensor; all vectors must have same dim
    def collate(batch):
        paths, vecs = zip(*batch)
        D = vecs[0].numel()
        vecs = [v.view(-1) for v in vecs if v.numel() == D]
        paths = [p for (p, v) in batch if v.numel() == D]
        return list(paths), torch.stack(vecs, dim=0)

    return torch.utils.data.DataLoader(ds, batch_size=batch_size, num_workers=num_workers, collate_fn=collate)

# ---- try it (set num_workers=0 first if you want easier debugging) ----
if __name__ == "__main__":
    paths, X = next(iter(make_embeddings_loader(num_workers=0, prefer_key=None)))
    print(len(paths), X.shape)

---

C) EEG (.npy/.npz) — ragged-friendly (returns list of arrays)

import io, re
import webdataset as wds
from huggingface_hub import HfFileSystem, hf_hub_url
import numpy as np

REPO_ID = "nonarjb/alignvis"   # your dataset repo on HF
REVISION = "main"
DATASET_DIR = "things_eeg_2"   # the folder inside the repo

def _hf_eeg_urls(repo_id=REPO_ID, dataset_dir=DATASET_DIR, revision=REVISION):
    """Collect EEG shard URLs for both possible top folders."""
    fs = HfFileSystem()
    urls = []
    for top in ("Preprocessed_data_250Hz", "preprocessed_eeg"):
        pattern = f"datasets/{repo_id}/{dataset_dir}/{dataset_dir}-{top}-*.tar"
        hf_paths = sorted(fs.glob(pattern))
        rel = [p.split(f"datasets/{repo_id}/", 1)[1] for p in hf_paths]
        urls += [hf_hub_url(repo_id, filename=p, repo_type="dataset", revision=revision) for p in rel]
    return urls

def _load_subject_eeg_from_hf(subject_id: int, split: str):
    """
    Returns (subject_eeg_data, ch_names) for a given subject+split
    by streaming the per-subject .npy/.npz from HF shards.
    """
    urls = _hf_eeg_urls()
    if not urls:
        raise RuntimeError("No EEG shards found in HF repo")
    filebase = "preprocessed_eeg_training" if split == "train" else "preprocessed_eeg_test"
    key_prefix = f"sub-{subject_id:02d}/"

    ds = wds.WebDataset(urls, shardshuffle=False)
    for s in ds:
        # find the per-subject file
        if ("npy" in s or "npz" in s) and s["__key__"].startswith(key_prefix) and s["__key__"].endswith(filebase):
            ext = "npz" if "npz" in s else "npy"
            payload = s[ext]
            bio = io.BytesIO(payload)

            # load with safe first, fallback to pickle (original code used allow_pickle=True)
            if ext == "npz":
                try:
                    z = np.load(bio, allow_pickle=False)
                except Exception:
                    bio.seek(0); z = np.load(bio, allow_pickle=True)
                # prefer exact fields as in your original code
                eeg_data = z["preprocessed_eeg_data"]
                ch_names = z["ch_names"] if "ch_names" in z else None
            else:  # npy
                try:
                    obj = np.load(bio, allow_pickle=False)
                except ValueError:
                    bio.seek(0); obj = np.load(bio, allow_pickle=True)

                # obj could be dict-like or 0-d object holding a dict
                if isinstance(obj, dict):
                    eeg_data = obj["preprocessed_eeg_data"]
                    ch_names = obj.get("ch_names")
                elif isinstance(obj, np.ndarray) and obj.dtype == object and obj.shape == ():
                    d = obj.item()
                    eeg_data = d["preprocessed_eeg_data"]
                    ch_names = d.get("ch_names")
                else:
                    # if it’s already a numeric array (unlikely for your case)
                    eeg_data = obj
                    ch_names = None

            return np.asarray(eeg_data), ch_names

    raise FileNotFoundError(f"Subject file not found in HF shards: {key_prefix}{filebase}.(npy|npz)")


subject_eeg_data, ch_names = _load_subject_eeg_from_hf(subject_id=1, split="train")
print(subject_eeg_data.shape)
print(ch_names)

If some .npy were saved as object-dtype, resave as numeric arrays; otherwise you must load with allow_pickle=True (only if you trust the data).

---

D) Download, untar, and use locally (byte-identical files)

# 1) Download the dataset subtree
from huggingface_hub import snapshot_download
local_root = snapshot_download(
    "nonarjb/alignvis", repo_type="dataset", allow_patterns=["things_eeg_2/**"]
)

# 2) Untar to a restore directory (keys preserved under each top folder)
import tarfile, glob, pathlib

restore_root = pathlib.Path("./restore/things_eeg_2")

for top in ("images", "image_embeddings", "preprocessed_eeg"):
    (restore_root / top).mkdir(parents=True, exist_ok=True)
    for t in glob.glob(f"{local_root}/things_eeg_2/things_eeg_2-{top}-*.tar"):
        with tarfile.open(t) as tf:
            tf.extractall(restore_root / top)

print("Restored under:", restore_root)

Now the folder tree mirrors the original:

# Example local usage
from PIL import Image
import numpy as np

img = Image.open("./restore/things_eeg_2/images/training_images/01133_raincoat/raincoat_01s.jpg")
vec = np.load("./restore/things_eeg_2/image_embeddings/some/file.npy")
eeg = np.load("./restore/things_eeg_2/preprocessed_eeg/s01/run3/segment_0001.npy", allow_pickle=False)

---

Notes

• WebDataset can also read local shards by passing file:// URLs instead of https://.
• If your shards are named differently, tweak hf_tar_urls(..., top="...") and the rel_path prefixes (images/, image_embeddings/, preprocessed_eeg/).
• To batch EEG tensors, implement padding in the collate function.