File size: 9,558 Bytes
481fbb6 | 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 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 | #!/usr/bin/env python3
"""Frozen document encoder for the Memory-LoRA hypernetwork.
Forked from Code2LoRA's ``create_dataset/embed_repos.py`` chunk/pool
pipeline (mean-pool chunks -> file vector -> weighted-mean+max repo vector),
retargeted from "repo files" to "document sections":
* Code2LoRA: file_i -> chunks -> mean-pool -> file vector
repo -> weighted-mean+max over file vectors
* Memory-LoRA: doc_section_i -> chunks -> mean-pool -> section vector
doc -> weighted-mean+max over section vectors
Same frozen encoder as the paper (Qwen3-Embedding-0.6B), same reasoning for
the weighting (content-distinctiveness via cosine-distance-from-mean +
log-size normalization) -- multi-section documents (e.g. the Code2LoRA paper
chunked into abstract/method/results/limitations) benefit from it exactly
the way multi-file repos did. Single-section synthetic fact-sheets degenerate
gracefully to a near-uniform weighting over their own chunks.
No gradient ever flows through this encoder; embeddings are precomputed once
and cached to parquet by ``scripts/build_doc_embeddings.py``.
"""
from __future__ import annotations
import re
from typing import List, Optional, Tuple
import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer
DEFAULT_EMBED_MODEL = "Qwen/Qwen3-Embedding-0.6B"
# Section-name heuristics (loose analogue of Code2LoRA's path up/down-weight
# lists). Neutral by default for synthetic single-section documents; the
# upweighted names matter for the multi-section Code2LoRA-paper document.
SECTION_UPWEIGHT = [
r"abstract", r"result", r"conclusion", r"contribution",
]
SECTION_DOWNWEIGHT = [
r"acknowledg", r"reference", r"appendix",
]
MIN_CHARS_FOR_FULL_WEIGHT = 200 # sections shorter than this are downweighted
# ---------------------------------------------------------------------------
# Chunking
# ---------------------------------------------------------------------------
def chunk_token_ids(token_ids: List[int], chunk_tokens: int, overlap: int) -> List[List[int]]:
"""Produce overlapping token windows (identical to Code2LoRA's version)."""
if chunk_tokens <= 0:
raise ValueError("chunk_tokens must be > 0")
if overlap >= chunk_tokens:
raise ValueError("chunk_overlap must be < chunk_tokens")
chunks: List[List[int]] = []
step = chunk_tokens - overlap
n = len(token_ids)
if n == 0:
return chunks
for start in range(0, n, step):
end = min(start + chunk_tokens, n)
window = token_ids[start:end]
if len(window) < 16:
continue
chunks.append(window)
if end >= n:
break
return chunks
# ---------------------------------------------------------------------------
# Embedding model wrapper
# ---------------------------------------------------------------------------
@torch.inference_mode()
def embed_texts(
model: AutoModel,
tokenizer: AutoTokenizer,
texts: List[str],
device: str,
batch_size: int,
max_length: int,
) -> torch.Tensor:
"""Return embeddings [N, D] using mean pooling over last_hidden_state."""
all_vecs = []
for i in range(0, len(texts), batch_size):
batch = texts[i:i + batch_size]
enc = tokenizer(
batch, padding=True, truncation=True,
max_length=max_length, return_tensors="pt",
)
enc = {k: v.to(device) for k, v in enc.items()}
out = model(**enc)
last = out.last_hidden_state # [B, T, H]
mask = enc["attention_mask"].unsqueeze(-1) # [B, T, 1]
mean = (last * mask).sum(dim=1) / mask.sum(dim=1).clamp(min=1)
all_vecs.append(mean.detach().cpu())
if not all_vecs:
return torch.empty((0, model.config.hidden_size))
return torch.cat(all_vecs, dim=0)
# ---------------------------------------------------------------------------
# Pooling: chunks -> section -> document
# ---------------------------------------------------------------------------
def pool_section_embeddings(chunk_embs: torch.Tensor) -> Optional[torch.Tensor]:
"""chunk_embs [K, D] -> section_emb [D]"""
if chunk_embs.numel() == 0:
return None
return chunk_embs.mean(dim=0)
def _section_name_bonus(section_name: str) -> float:
s = section_name.lower()
bonus = 0.0
for pat in SECTION_DOWNWEIGHT:
if re.search(pat, s):
bonus -= 0.25
break
for pat in SECTION_UPWEIGHT:
if re.search(pat, s):
bonus += 0.15
break
return bonus
def compute_section_weights(
section_embs: torch.Tensor, # [S, D]
section_char_counts: torch.Tensor, # [S]
section_names: List[str],
a_distinct: float,
b_size: float,
tau: float,
) -> torch.Tensor:
"""
Whole-doc, all-sections weighting:
distinct_i = 1 - cos(s_i, mean_s)
size_i = normalized log(1+chars)
score_i = a_distinct * distinct_i + b_size * size_i + name_bonus_i + tiny_section_penalty
w = softmax(score / tau)
Returns: w [S] sum=1
"""
f_norm = F.normalize(section_embs, p=2, dim=-1)
mean_f = F.normalize(f_norm.mean(dim=0, keepdim=True), p=2, dim=-1)
cos = (f_norm * mean_f).sum(dim=-1).clamp(-1, 1)
distinct = 1.0 - cos
chars = section_char_counts.float().clamp(min=1)
log_chars = torch.log1p(chars)
if log_chars.numel() > 1:
lo, hi = log_chars.min(), log_chars.max()
size01 = (log_chars - lo) / (hi - lo + 1e-8)
else:
size01 = torch.ones_like(log_chars)
name_bonus = torch.tensor([_section_name_bonus(n) for n in section_names],
dtype=torch.float32)
tiny_scale = (chars / float(MIN_CHARS_FOR_FULL_WEIGHT)).clamp(max=1.0)
tiny_bonus = torch.log(tiny_scale + 1e-6)
score = (a_distinct * distinct.cpu() + b_size * size01.cpu()
+ name_bonus + 0.15 * tiny_bonus.cpu())
return torch.softmax(score / max(tau, 1e-6), dim=0)
def pool_doc_embedding_weighted(
section_embs: torch.Tensor, # [S, D]
section_char_counts: torch.Tensor, # [S]
section_names: List[str],
a_distinct: float = 1.0,
b_size: float = 0.5,
tau: float = 0.5,
alpha_mean: float = 1.0,
beta_max: float = 1.0,
) -> Optional[torch.Tensor]:
"""Aggregate section embeddings into one document vector:
concat(alpha_mean * weighted_mean, beta_max * max) -> [2D]. No final
L2 normalization (matches Code2LoRA's repo-vector convention)."""
if section_embs.numel() == 0:
return None
w = compute_section_weights(
section_embs, section_char_counts, section_names,
a_distinct, b_size, tau,
).to(section_embs.dtype)
wmean = (section_embs * w.unsqueeze(-1)).sum(dim=0)
vmax = section_embs.max(dim=0).values
return torch.cat([alpha_mean * wmean, beta_max * vmax], dim=0)
# ---------------------------------------------------------------------------
# Main pipeline per document
# ---------------------------------------------------------------------------
def embed_document(
sections: List[Tuple[str, str]], # [(section_name, section_text), ...]
model: AutoModel,
tokenizer: AutoTokenizer,
device: str,
chunk_tokens: int = 4096,
chunk_overlap: int = 512,
batch_size: int = 4,
a_distinct: float = 1.0,
b_size: float = 0.5,
tau: float = 0.5,
alpha_mean: float = 1.0,
beta_max: float = 1.0,
) -> Optional[torch.Tensor]:
"""One document = list of (name, text) sections (single-element for a
plain synthetic fact-sheet; multi-element for the chunked paper).
Returns a [2D] embedding, or None if the document had no usable text."""
section_vectors: List[torch.Tensor] = []
section_names: List[str] = []
section_char_counts: List[int] = []
for name, text in sections:
text = (text or "").strip()
if not text:
continue
ids = tokenizer.encode(text, add_special_tokens=False)
windows = chunk_token_ids(ids, chunk_tokens=chunk_tokens, overlap=chunk_overlap)
if not windows:
continue
chunks = [tokenizer.decode(w, skip_special_tokens=True) for w in windows]
chunk_embs = embed_texts(
model=model, tokenizer=tokenizer, texts=chunks,
device=device, batch_size=batch_size, max_length=chunk_tokens,
)
svec = pool_section_embeddings(chunk_embs)
if svec is None:
continue
section_vectors.append(svec)
section_names.append(name)
section_char_counts.append(len(text))
if not section_vectors:
return None
section_embs = torch.stack(section_vectors, dim=0)
char_t = torch.tensor(section_char_counts, dtype=torch.int64)
return pool_doc_embedding_weighted(
section_embs, char_t, section_names,
a_distinct=a_distinct, b_size=b_size, tau=tau,
alpha_mean=alpha_mean, beta_max=beta_max,
)
def load_encoder(model_name: str = DEFAULT_EMBED_MODEL, device: str = "mps"):
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name, torch_dtype=torch.float32)
model.to(device)
model.eval()
return model, tokenizer
__all__ = [
"DEFAULT_EMBED_MODEL",
"chunk_token_ids",
"embed_texts",
"pool_section_embeddings",
"compute_section_weights",
"pool_doc_embedding_weighted",
"embed_document",
"load_encoder",
]
|