"""Local UNI2-h H5-to-mean converter. pip install numpy h5py

Input manifest CSV: patient_id,slide_id,path[,dataset]
Usage: python convert_h5.py --manifest slides.csv --output slide_means.csv
Output contains per-slide float32 means; website performs equal patient pooling.
"""
import argparse,csv
from pathlib import Path
import numpy as np
import h5py

def convert(manifest,output):
    manifest=Path(manifest); output=Path(output)
    if output.exists(): raise ValueError('Choose a new output filename; existing files are not overwritten.')
    rows=list(csv.DictReader(manifest.open(encoding='utf-8-sig',newline='')))
    seen=set(); results=[]
    for row in rows:
        pid=row.get('patient_id','').strip(); sid=row.get('slide_id','').strip()
        if not pid or not sid or sid in seen: raise ValueError('Missing patient/slide ID or duplicate slide ID')
        seen.add(sid); path=Path(row['path'])
        if not path.is_absolute(): path=manifest.parent/path
        with h5py.File(path,'r') as f:
            candidates=[]
            def visit(name,obj):
                if isinstance(obj,h5py.Dataset) and (obj.ndim==2 or (obj.ndim==3 and obj.shape[0]==1)) and obj.shape[-1]>16 and np.issubdtype(obj.dtype,np.number): candidates.append(name)
            f.visititems(visit)
            key=row.get('dataset','').strip()
            if not key:
                preferred=[k for k in candidates if k.split('/')[-1].lower() in ['features','feature','embeddings']]
                if len(preferred)==1: key=preferred[0]
                elif len(candidates)==1: key=candidates[0]
                else: raise ValueError(f'{sid}: ambiguous or missing datasets; specify dataset in manifest')
            ds=f[key]
            if ds.ndim not in (2,3) or (ds.ndim==3 and ds.shape[0]!=1) or ds.shape[-1]!=1536 or ds.shape[-2]==0: raise ValueError(f'{sid}: invalid shape')
            n=ds.shape[-2]; total=np.zeros(1536,dtype=np.float64)
            for start in range(0,n,4096):
                x=ds[0,start:start+4096] if ds.ndim==3 else ds[start:start+4096]
                if not np.isfinite(x).all(): raise ValueError(f'{sid}: non-finite features')
                total+=x.sum(axis=0,dtype=np.float64)
            results.append([pid,sid]+(total/n).astype(np.float32).tolist())
    if not results: raise ValueError('No files in manifest')
    with output.open('x',encoding='utf-8',newline='') as f:
        w=csv.writer(f);w.writerow(['patient_id','slide_id']+[f'feature_{i:04d}' for i in range(1,1537)]);w.writerows(results)
    print(f'Converted {len(results)} slides. Source files were read only.')

if __name__=='__main__':
    p=argparse.ArgumentParser();p.add_argument('--manifest',required=True);p.add_argument('--output',required=True);a=p.parse_args();convert(a.manifest,a.output)
