monet_code/step0_sg/pack_reps.py

#!/usr/bin/env python
"""Pack Monet step1 teacher reps (rep_all_layers_scene_graph_<sid>.pt) into step0-all
schema parquet:  sample_id, category, source_dataset, num_intermediate_steps,
                 latent_dtype, latent_shape, latent(list<float16>)
Row-groups are kept small (<~250MB) so the HF dataset viewer can preview.
Metadata (category/source_dataset/num_intermediate_steps) comes from the source ULVR parquet.
"""
import argparse, glob, json, os, re
import numpy as np, torch
import pyarrow as pa, pyarrow.parquet as pq

def load_meta(parquet_glob):
    meta={}
    for fp in sorted(glob.glob(parquet_glob)):
        t=pq.read_table(fp, columns=["sample_id","category","source_dataset","num_intermediate_steps"]).to_pylist()
        for r in t: meta[r["sample_id"]]=r
    return meta

SCHEMA=pa.schema([
    ("sample_id",pa.string()),("category",pa.string()),("source_dataset",pa.string()),
    ("num_intermediate_steps",pa.int64()),("latent_dtype",pa.string()),
    ("latent_shape",pa.list_(pa.int64())),("latent",pa.list_(pa.float16())),
])

def main():
    ap=argparse.ArgumentParser()
    ap.add_argument("--reps-dir",required=True)
    ap.add_argument("--src-glob",required=True, help="ULVR scene_graph parquet glob for metadata")
    ap.add_argument("--out-dir",required=True)
    ap.add_argument("--split-name",default="scene_graph")
    ap.add_argument("--rows-per-file",type=int,default=2000)
    ap.add_argument("--row-group-size",type=int,default=128)
    a=ap.parse_args()
    os.makedirs(a.out_dir,exist_ok=True)
    meta=load_meta(a.src_glob)
    reps=sorted(glob.glob(os.path.join(a.reps_dir,"rep_all_layers_*.pt")))
    print(f"reps={len(reps)} meta={len(meta)}")
    prefix="all_layers_"
    def emit(rows, idx):
        cols={k:[r[k] for r in rows] for k in ["sample_id","category","source_dataset","num_intermediate_steps","latent_dtype","latent_shape","latent"]}
        tb=pa.table({
            "sample_id":pa.array(cols["sample_id"],pa.string()),
            "category":pa.array(cols["category"],pa.string()),
            "source_dataset":pa.array(cols["source_dataset"],pa.string()),
            "num_intermediate_steps":pa.array(cols["num_intermediate_steps"],pa.int64()),
            "latent_dtype":pa.array(cols["latent_dtype"],pa.string()),
            "latent_shape":pa.array(cols["latent_shape"],pa.list_(pa.int64())),
            "latent":pa.array(cols["latent"],pa.list_(pa.float16())),
        },schema=SCHEMA)
        outp=os.path.join(a.out_dir,f"{a.split_name}-%05d.parquet"%idx)
        pq.write_table(tb,outp,row_group_size=a.row_group_size,compression="zstd")
        print("wrote",outp,len(rows),"rows")
    buf=[]; fidx=0; miss=0
    for rp in reps:
        d=torch.load(rp,map_location="cpu")
        mi=d["metadata_info"]            # all_layers_scene_graph_<sid>
        lat=d["latent"]                  # [29, N*latent_size, D] (bf16)
        body=mi[len(prefix):] if mi.startswith(prefix) else mi
        # body = scene_graph_<sid>; strip leading "scene_graph_"
        sid=body.split("_",2)[-1] if body.startswith("scene_graph_") else body
        # robust: try exact, else find a meta key that body endswith
        m=meta.get(sid)
        if m is None:
            cand=[k for k in (sid,body) if k in meta]
            m=meta.get(cand[0]) if cand else None
        if m is None:
            miss+=1
            if miss<=5: print("WARN no meta for",mi,"->",sid)
            continue
        arr=lat.float().numpy().astype(np.float16).reshape(-1)
        buf.append({"sample_id":m["sample_id"],"category":m["category"],"source_dataset":m["source_dataset"],
                    "num_intermediate_steps":int(m["num_intermediate_steps"]),
                    "latent_dtype":"bfloat16","latent_shape":[int(x) for x in lat.shape],
                    "latent":arr})
        if len(buf)>=a.rows_per_file:
            emit(buf,fidx); fidx+=1; buf=[]
    if buf: emit(buf,fidx)
    print("done. missing meta:",miss)

if __name__=="__main__": main()