File size: 3,017 Bytes
2eb3475 | 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 | """Punctuation-template attention feature (Kostenok et al. 2024 §D).
For each sample we identify punctuation token positions via the tokenizer and
build a per-sample template mask T_mask[i, j] = 1 if token j is a punctuation
token, 0 otherwise. We then compute, per (layer, head):
feature = 1 - sum(attn * T_mask) / sum(T_mask)
which is small when the head attends strongly to punctuation tokens and ~1
when the head ignores them.
"""
from __future__ import annotations
from pathlib import Path
from typing import List, Tuple
import numpy as np
import torch
from torch.utils.data import DataLoader
from tqdm import tqdm
PUNCT_CHARS = set(".,;:!?\"'`()[]{}-—–…")
def _punct_token_ids(tokenizer) -> set:
"""All token IDs whose decoded text is a punctuation-only string."""
out = set()
vocab = tokenizer.get_vocab()
for tok, idx in vocab.items():
# Decode the single token; strip BERT/RoBERTa's space/##/Ġ markers
s = tok.replace("##", "").replace("Ġ", "").replace("▁", "")
if s and all(c in PUNCT_CHARS for c in s):
out.add(idx)
return out
@torch.no_grad()
def compute_punct_template(attn_dir: Path, tokenizer, model_max_length: int,
df, text_col: str) -> Tuple[np.ndarray, List[str]]:
L, H = 12, 12
punct_ids = _punct_token_ids(tokenizer)
rows = []
for _, r in tqdm(df.reset_index(drop=True).iterrows(), total=len(df), desc="punct-template"):
idx = int(r["idx"])
# Re-tokenize to get per-token IDs for this sample
enc = tokenizer(r[text_col], truncation=True, max_length=model_max_length,
padding="max_length", return_tensors="pt")
ids = enc["input_ids"][0].tolist()
is_punct = np.array([tid in punct_ids for tid in ids], dtype=np.float32) # (T,)
if is_punct.sum() == 0:
# No punctuation tokens — feature defaults to 1.0
rows.append(np.ones(L * H, dtype=np.float32))
continue
# Build template mask: every row attends to every punctuation position
T = is_punct.shape[0]
T_mask = np.tile(is_punct[None, :], (T, 1)) # (T, T)
denom = float(T_mask.sum())
# Load attention dump
path = attn_dir / f"{idx:06d}.npz"
with np.load(path, allow_pickle=True) as d:
A = np.asarray(d["attn"]).astype(np.float32) # (L, H, T_attn, T_attn)
T_attn = A.shape[-1]
T_mask = T_mask[:T_attn, :T_attn]
denom = float(T_mask.sum())
if denom <= 0:
rows.append(np.ones(L * H, dtype=np.float32))
continue
feats = np.zeros(L * H, dtype=np.float32)
for l in range(L):
for h in range(H):
feats[l * H + h] = 1.0 - float((A[l, h] * T_mask).sum() / denom)
rows.append(feats)
arr = np.stack(rows, axis=0).astype(np.float32)
cols = [f"punct_L{l}H{h}" for l in range(L) for h in range(H)]
return arr, cols
|