scripts/upload_hf_preprocessed.py

"""Upload preprocessed .pt cache to Hugging Face.

6,050 files × ~4 MB = ~24 GB. Uploads in batches of 60 files (~240 MB/batch).
Resumable: HF Hub skips files already uploaded (content-addressed).
"""
from __future__ import annotations
import time
from pathlib import Path
from huggingface_hub import HfApi, CommitOperationAdd

TOKEN = "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"
REPO_ID = "bilalahmad176176/BrainAge-Golden-Preprocessed"
CACHE_DIR = Path("/home/MRI-DataSet/_train/cache")
BATCH_SIZE = 60


def upload_readme(api: HfApi):
    readme = """---
license: cc-by-nc-4.0
task_categories:
  - other
language:
  - en
pretty_name: BrainAge Golden Preprocessed Cache
size_categories:
  - 1K<n<10K
tags:
  - neuroimaging
  - mri
  - brain-age
  - pytorch
  - preprocessed
  - tensor-cache
---

# BrainAge Golden Preprocessed Cache

**6,050 preprocessed brain MRI tensors** ready for training a brain-age
prediction model. Skip the 40+ hour preprocessing step and jump straight
to model training.

## What's inside

Each `.pt` file (one per subject) contains:

| Key | Type | Shape | Description |
|-----|------|-------|-------------|
| `volume` | float16 | (128, 144, 112) | Z-normed T1w brain in MNI space, trilinear-resized |
| `tab` | float32 | (86,) | 70 regional volumes (log1p/12) + 3 sex one-hot + 13 site one-hot |
| `age` | float32 | scalar | Chronological age in years |
| `meta` | dict | — | subject_id, site, sex, age, split |

## Stats

| Metric | Value |
|--------|-------|
| Total subjects | 6,050 |
| Age range | 0 – 86 years |
| Source datasets | 12 (BCP, Calgary, ds002726, ds000248, PTBP, IXI, MPI-Leipzig, AOMIC, NKI-Rockland, ABIDE-I, ABIDE-II, ADHD-200) |
| Volume shape | 128 × 144 × 112 (D × H × W) |
| Tabular dim | 86 (70 regions + 3 sex + 13 site) |
| File size | ~4 MB each |
| Total size | ~24 GB |

## Preprocessing pipeline applied

```
Raw T1w NIfTI
  → HD-BET skull-strip (GPU)
  → N4 bias correction (ANTs)
  → Affine registration to MNI152 1mm
  → Z-score intensity normalization
  → Harvard-Oxford atlas segmentation (69 regions)
  → Volume measurement + rescaling to native space
  → Tensor packaging (.pt)
```

## Quick start

```python
from huggingface_hub import snapshot_download
import torch

# Download (~24 GB)
snapshot_download(
    "bilalahmad176176/BrainAge-Golden-Preprocessed",
    repo_type="dataset",
    local_dir="cache/"
)

# Load one subject
data = torch.load("cache/cache/IXI002.pt", weights_only=False)
print(data["volume"].shape)  # (128, 144, 112) float16
print(data["tab"].shape)     # (86,) float32
print(data["age"])            # e.g. 36.2
print(data["meta"])           # {'subject_id': 'IXI002', 'site': 'DataSet-6_IXI', ...}
```

## Train a model

```bash
# Generate split
python -m pipeline_v2.data_split \\
    --manifests Golden-0-to-25/manifest.csv Golden-25plus/manifest.csv \\
    --out cache/split.csv

# Train
python -m pipeline_v2.train \\
    --cache_dir cache/cache \\
    --split_csv cache/split.csv \\
    --out_ckpt brainage_sfcn.pt \\
    --epochs 60 --batch 4
```

## Related

- Raw dataset: [bilalahmad176176/BrainAge-Golden-Raw](https://huggingface.co/datasets/bilalahmad176176/BrainAge-Golden-Raw)
- 3D Viewer demo: [bilalahmad176176/BrainAge-3D-Viewer](https://huggingface.co/spaces/bilalahmad176176/BrainAge-3D-Viewer)

## Citation

Please cite the original source studies listed in the raw dataset manifests.
"""
    api.upload_file(
        path_or_fileobj=readme.encode(),
        path_in_repo="README.md",
        repo_id=REPO_ID,
        repo_type="dataset",
        commit_message="Add dataset card",
    )
    print("README.md uploaded.")


def main():
    api = HfApi(token=TOKEN)
    print(f"User: {api.whoami()['name']}")
    print(f"Repo: https://huggingface.co/datasets/{REPO_ID}")

    upload_readme(api)

    # Also upload the manifests + status logs for reproducibility
    extras = [
        ("/home/MRI-DataSet/Golden-0-to-25/manifest.csv", "manifests/Golden-0-to-25_manifest.csv"),
        ("/home/MRI-DataSet/Golden-25plus/manifest.csv", "manifests/Golden-25plus_manifest.csv"),
        ("/home/MRI-DataSet/_train/logs/preprocess_status.csv", "logs/preprocess_status.csv"),
    ]
    ops = []
    for local, repo_path in extras:
        if Path(local).exists():
            ops.append(CommitOperationAdd(path_in_repo=repo_path, path_or_fileobj=local))
    if ops:
        api.create_commit(repo_id=REPO_ID, repo_type="dataset",
                          operations=ops, commit_message="Add manifests and preprocessing logs")
        print(f"Uploaded {len(ops)} metadata files.")

    # Upload .pt files in batches
    pts = sorted(CACHE_DIR.glob("*.pt"))
    total = len(pts)
    print(f"\nUploading {total} .pt files (~24 GB)…\n")

    for i in range(0, total, BATCH_SIZE):
        batch = pts[i : i + BATCH_SIZE]
        ops = [CommitOperationAdd(
            path_in_repo=f"cache/{p.name}",
            path_or_fileobj=str(p),
        ) for p in batch]
        n = min(i + BATCH_SIZE, total)
        msg = f"Add cache/{pts[i].name}…{batch[-1].name} ({i+1}–{n} of {total})"
        print(f"  [{n}/{total}] committing … ", end="", flush=True)
        t0 = time.time()
        try:
            api.create_commit(
                repo_id=REPO_ID, repo_type="dataset",
                operations=ops, commit_message=msg,
            )
            print(f"done ({time.time()-t0:.0f}s)")
        except Exception as e:
            print(f"ERROR: {e}")
            time.sleep(5)
            try:
                api.create_commit(
                    repo_id=REPO_ID, repo_type="dataset",
                    operations=ops, commit_message=msg + " (retry)",
                )
                print("  retry ok")
            except Exception as e2:
                print(f"  retry failed: {e2}, skipping batch — rerun to resume")

    print(f"\nDONE — https://huggingface.co/datasets/{REPO_ID}")


if __name__ == "__main__":
    main()