File size: 4,250 Bytes
d231962 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 | """Pull a channel's embedding matrices from its HF repo onto the box.
Only e5-small was built locally. e5-large and bge-m3 exist on the Hub, and step C wants all
three - the union of three top-100 lists is what moves the oracle from 0.9953 to 0.9976, and
a "three-channel union" computed from one channel is just that channel.
export HF_TOKEN=hf_...
python scripts/07_fetch_embeddings.py --channel e5_large --split test
SIZE. A 1024-d test pair is about 20 GB pool + 3.5 GB s1. Check free space first; the script
refuses rather than filling /scratch and taking the box down for the other 25 users on it.
The layout written here matches what paths.emb expects, so nothing else needs to change. The
download is resumable - hf_hub keeps its own cache and a re-run skips what is already there.
"""
from __future__ import annotations
import argparse
import os
import shutil
import sys
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from berx import io_hf # noqa: E402
from berx import paths as P # noqa: E402
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--channel", required=True, choices=sorted(P.CHANNELS))
ap.add_argument("--split", default="test")
ap.add_argument("--which", nargs="+", default=["s1", "pool"], choices=["s1", "pool"])
ap.add_argument("--min-free-gb", type=float, default=40.0)
ap.add_argument("--list", action="store_true", help="only show what the repo contains")
a = ap.parse_args()
repo = P.CHANNELS[a.channel]["repo"]
if a.list:
from huggingface_hub import HfApi
files = HfApi(token=io_hf.token()).list_repo_files(repo, repo_type="dataset")
for f in sorted(files):
print(" ", f)
print(f"\n{len(files)} files in {repo}")
return
free = shutil.disk_usage(P.EMB).free / 2 ** 30
if free < a.min_free_gb:
sys.exit(f"Only {free:.0f} GiB free under {P.EMB}; need about {a.min_free_gb:.0f}. "
f"Refusing - filling /scratch takes the box down for everyone on it.")
print(f"{free:.0f} GiB free under {P.EMB}")
for which in a.which:
dest = P.emb_dir(a.channel, a.split, which)
if os.path.exists(os.path.join(dest, "matrix.npy")):
print(f" {which}: already present at {dest}, skipping")
continue
os.makedirs(dest, exist_ok=True)
# The repo layout is not guaranteed to match ours, so fetch by pattern and report
# what arrived rather than assuming a path.
pats = [f"**/{a.split}_{which}/matrix.npy", f"**/{a.split}_{which}/ids.txt",
f"**/{a.split}_{which}/meta.json",
f"{a.split}_{which}/*"]
print(f" {which}: downloading from {repo} ...")
got = io_hf.fetch_dir(repo, allow=pats, local_dir=dest)
found = []
for root, _, files in os.walk(got):
for f in files:
if f in ("matrix.npy", "ids.txt", "meta.json"):
src = os.path.join(root, f)
dst = os.path.join(dest, f)
if os.path.abspath(src) != os.path.abspath(dst):
shutil.move(src, dst)
found.append(f)
if "matrix.npy" not in found:
print(f" nothing matched. Run with --list to see the repo layout and adjust "
f"the patterns; do NOT guess.")
continue
import numpy as np
m = np.load(os.path.join(dest, "matrix.npy"), mmap_mode="r")
want = P.CHANNELS[a.channel]["dim"]
print(f" {m.shape} {m.dtype} expected dim {want}")
assert m.shape[1] == want, (
f"{a.channel} should be {want}-d but this matrix is {m.shape[1]}-d. The two "
f"1024-d channels are interchangeable by shape and not by content - check you "
f"pulled the right repo.")
n = np.linalg.norm(np.asarray(m[:256], dtype=np.float32), axis=1).mean()
print(f" mean L2 norm {n:.5f} " +
("(normalised, good)" if abs(n - 1) < 0.02 else "(NOT normalised - fix before use)"))
print(f"\nNow: python scripts/03_retrieve.py --split {a.split} --channel {a.channel} ...")
if __name__ == "__main__":
main()
|