File size: 29,281 Bytes
e5f4c09 0cc1b2b e5f4c09 0cc1b2b e5f4c09 642c04b e5f4c09 642c04b e5f4c09 0cc1b2b e5f4c09 642c04b e5f4c09 642c04b e5f4c09 642c04b e5f4c09 642c04b e5f4c09 642c04b e5f4c09 642c04b e5f4c09 0cc1b2b e5f4c09 642c04b e5f4c09 0cc1b2b e5f4c09 0cc1b2b e5f4c09 0cc1b2b e5f4c09 0cc1b2b e5f4c09 642c04b e5f4c09 642c04b e5f4c09 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 | """Unified single-checkpoint inference for all AspectBench model families."""
from __future__ import annotations
import json
import math
import random
import re
from collections import Counter
from pathlib import Path
from typing import Any, Iterable
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import (
AutoConfig,
AutoModel,
AutoModelForSeq2SeqLM,
AutoModelForSequenceClassification,
AutoTokenizer,
LongformerConfig,
LongformerForSequenceClassification,
PreTrainedTokenizerFast,
)
from model_registry import CHECKPOINTS, MODEL_SPECS, model_spec, weight_path
ASPECT_TOKEN = "[ASPECT]"
HAN_ASPECT_TOKEN = "[ASPECT_TARGET]"
ASPECT_PATTERN = re.compile(r"<aspect>(.*?)</aspect>", flags=re.DOTALL)
ID_TO_SENTIMENT = {0: -1, 1: 0, 2: 1}
SENTIMENT_NAMES = {-1: "negative", 0: "neutral", 1: "positive"}
CLASS_KEYS = ("-1 (negative)", "0 (neutral)", "1 (positive)")
LABEL_TEXTS = ("negative", "neutral", "positive")
def _seed_everything(seed: int) -> None:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
def prepare_record(record: dict[str, Any], mode: str) -> dict[str, Any]:
if not isinstance(record, dict):
raise TypeError("Each input record must be a JSON object/dict.")
article = record.get("article")
if not isinstance(article, str) or not article.strip():
raise ValueError("Each record needs a non-empty string field named 'article'.")
tagged = [value.strip() for value in ASPECT_PATTERN.findall(article)]
if not tagged or any(not value for value in tagged):
raise ValueError("The article must contain non-empty <aspect>...</aspect> tags.")
remainder = ASPECT_PATTERN.sub("", article)
if "<aspect>" in remainder or "</aspect>" in remainder:
raise ValueError("Aspect tags are unbalanced or nested.")
explicit_aspect = record.get("aspect")
if explicit_aspect is not None and (
not isinstance(explicit_aspect, str) or not explicit_aspect.strip()
):
raise ValueError("Optional 'aspect' must be a non-empty string.")
aspect = explicit_aspect.strip() if explicit_aspect else tagged[0]
gold = record.get("sentiment")
if gold is not None and gold not in (-1, 0, 1):
raise ValueError("Optional 'sentiment' must be -1, 0, or 1.")
if mode == "masked":
model_article = ASPECT_PATTERN.sub(ASPECT_TOKEN, article).strip()
model_aspect = ASPECT_TOKEN
else:
model_article = article.replace("<aspect>", "").replace("</aspect>", "").strip()
model_aspect = aspect
return {
"article": article,
"tagged_aspects": tagged,
"aspect": aspect,
"model_article": model_article,
"model_aspect": model_aspect,
"gold_sentiment": gold,
}
def load_records(path: str | Path) -> list[dict[str, Any]]:
input_path = Path(path)
if input_path.suffix.lower() == ".jsonl":
with input_path.open("r", encoding="utf-8") as handle:
records = [json.loads(line) for line in handle if line.strip()]
else:
with input_path.open("r", encoding="utf-8") as handle:
payload = json.load(handle)
if isinstance(payload, list):
records = payload
elif isinstance(payload, dict) and "article" in payload:
records = [payload]
elif isinstance(payload, dict):
records = next(
(
payload[key]
for key in ("records", "test", "train", "val")
if isinstance(payload.get(key), list)
),
[],
)
else:
records = []
if not records:
raise ValueError("Input has no records.")
return records
def write_json(payload: Any, path: str | Path | None = None) -> None:
rendered = json.dumps(payload, indent=2, ensure_ascii=False)
if path is None:
print(rendered)
else:
Path(path).write_text(rendered + "\n", encoding="utf-8")
def _enable_dropout_only(model: nn.Module) -> None:
model.eval()
dropout_types = (
nn.Dropout,
nn.Dropout1d,
nn.Dropout2d,
nn.Dropout3d,
nn.AlphaDropout,
nn.FeatureAlphaDropout,
)
for module in model.modules():
if isinstance(module, dropout_types):
module.train()
def _entropy_bits(probabilities: np.ndarray) -> float:
safe = np.clip(probabilities.astype(np.float64), 1e-12, 1.0)
return float(-np.sum(safe * np.log2(safe)))
def _uncertainty(samples: np.ndarray, predictions: list[int]) -> dict[str, Any]:
mean_probs = samples.mean(axis=0)
predictive_entropy = _entropy_bits(mean_probs)
sorted_probs = np.sort(mean_probs)
result: dict[str, Any] = {
"predictive_entropy_bits": predictive_entropy,
"normalized_predictive_entropy": predictive_entropy / math.log2(3),
"confidence": float(mean_probs.max()),
"top_two_probability_margin": float(sorted_probs[-1] - sorted_probs[-2]),
}
if len(samples) > 1:
expected_entropy = float(np.mean([_entropy_bits(row) for row in samples]))
counts = Counter(predictions)
agreement = max(counts.values()) / len(predictions)
result["mc_dropout"] = {
"passes": len(samples),
"expected_entropy_bits": expected_entropy,
"mutual_information_bits": max(0.0, predictive_entropy - expected_entropy),
"prediction_agreement": float(agreement),
"variation_ratio": float(1.0 - agreement),
"vote_counts": {
str(label): int(counts.get(label, 0)) for label in (-1, 0, 1)
},
}
return result
def _resolve_base(
spec: dict[str, Any],
base_model_root: str | Path | None,
bundled_assets: Path | None = None,
) -> str:
if bundled_assets is not None and (bundled_assets / "config.json").is_file():
return str(bundled_assets)
if base_model_root:
candidate = Path(base_model_root) / spec["local_base_dir"]
if candidate.is_dir():
return str(candidate)
return spec["base_model"]
def _load_tokenizer(base_path: str, model_name: str, language: str, max_length: int):
path = Path(base_path)
local_only = path.is_dir()
if model_name == "slavic-specific" and language == "slovenian" and local_only:
return PreTrainedTokenizerFast(
tokenizer_file=str(path / "tokenizer.json"),
bos_token="<s>",
eos_token="</s>",
sep_token="</s>",
cls_token="<s>",
unk_token="<unk>",
pad_token="<pad>",
mask_token="<mask>",
model_max_length=max_length,
)
return AutoTokenizer.from_pretrained(
base_path,
local_files_only=local_only,
use_fast=True,
fix_mistral_regex=True,
)
def _load_state(path: Path) -> dict[str, torch.Tensor]:
try:
state = torch.load(path, map_location="cpu", weights_only=True, mmap=True)
except TypeError:
state = torch.load(path, map_location="cpu")
if not isinstance(state, dict) or not state:
raise ValueError(f"Weight file is not a model state dictionary: {path}")
return state
def _resize_embeddings(model: nn.Module, size: int) -> None:
try:
model.resize_token_embeddings(size, mean_resizing=False)
except TypeError:
model.resize_token_embeddings(size)
def _resize_embeddings_to_checkpoint(
model: nn.Module, state: dict[str, torch.Tensor]
) -> None:
"""Match base-model vocabulary padding used by the original training run."""
if not hasattr(model, "get_input_embeddings"):
return
embedding = model.get_input_embeddings()
candidates = [
value
for key, value in state.items()
if value.ndim == 2
and (
key.endswith("word_embeddings.weight")
or key == "shared.weight"
or key.endswith(".shared.weight")
)
and value.shape[1] == embedding.weight.shape[1]
]
if not candidates:
return
checkpoint_vocab = max(value.shape[0] for value in candidates)
if checkpoint_vocab == embedding.weight.shape[0]:
return
_resize_embeddings(model, checkpoint_vocab)
class SimplifiedDARTModel(nn.Module):
def __init__(
self,
base_config: Any,
tokenizer_len: int,
interaction_layers: int = 2,
interaction_heads: int = 8,
aggregation_heads: int = 4,
max_sentences: int = 128,
final_mlp_hidden_dim: int = 256,
dropout_rate: float = 0.2,
base_model: nn.Module | None = None,
) -> None:
super().__init__()
self.base_model = base_model or AutoModel.from_config(base_config)
_resize_embeddings(self.base_model, tokenizer_len)
self.hidden_dim = base_config.hidden_size
self.dropout = nn.Dropout(dropout_rate)
self.sentence_pos_embedding = nn.Embedding(
max_sentences + 1, self.hidden_dim, padding_idx=0
)
layer = nn.TransformerEncoderLayer(
d_model=self.hidden_dim,
nhead=interaction_heads,
dim_feedforward=self.hidden_dim * 4,
dropout=dropout_rate,
activation="relu",
batch_first=True,
)
self.sentence_interact_transformer = nn.TransformerEncoder(
layer, num_layers=interaction_layers, enable_nested_tensor=False
)
self.global_aggregation_attention = nn.MultiheadAttention(
self.hidden_dim, aggregation_heads, dropout=dropout_rate, batch_first=True
)
self.classifier = nn.Sequential(
nn.Linear(self.hidden_dim, final_mlp_hidden_dim),
nn.ReLU(),
nn.Dropout(dropout_rate),
nn.Linear(final_mlp_hidden_dim, 3),
)
def forward(
self,
input_ids: torch.Tensor,
attention_mask: torch.Tensor,
sentence_mask: torch.Tensor,
sentence_position_ids: torch.Tensor,
aspect_target_token_id: torch.Tensor,
) -> torch.Tensor:
batch_size, sentence_count, token_count = input_ids.shape
# Padded sentence slots carry no information and are masked everywhere
# downstream. Encoding only real sentences preserves the computation for
# those slots while making the paper's 128 x 96 layout practical on 48 GB
# GPUs.
valid = sentence_mask.view(-1).bool()
if not bool(valid.any()):
raise ValueError("Every HAN batch item must contain a non-empty sentence.")
flat_ids = input_ids.view(-1, token_count)
flat_attention = attention_mask.view(-1, token_count)
outputs = self.base_model(
input_ids=flat_ids[valid],
attention_mask=flat_attention[valid],
)
cls_flat = outputs.last_hidden_state.new_zeros(
(batch_size * sentence_count, self.hidden_dim)
)
cls_flat = cls_flat.index_copy(
0,
valid.nonzero(as_tuple=False).squeeze(1),
outputs.last_hidden_state[:, 0, :],
)
cls = cls_flat.view(batch_size, sentence_count, self.hidden_dim)
cls = self.dropout(cls + self.sentence_pos_embedding(sentence_position_ids))
padding_mask = sentence_mask == 0
contextualized = self.sentence_interact_transformer(
cls, src_key_padding_mask=padding_mask
)
contextualized *= sentence_mask.unsqueeze(-1).float()
aspect_embedding = self.base_model.get_input_embeddings()(
aspect_target_token_id.to(input_ids.device)
)
if aspect_embedding.ndim == 1:
aspect_embedding = aspect_embedding.unsqueeze(0)
query = aspect_embedding.unsqueeze(0).repeat(batch_size, 1, 1)
aggregated, _ = self.global_aggregation_attention(
query=query,
key=contextualized,
value=contextualized,
key_padding_mask=padding_mask,
)
return self.classifier(self.dropout(aggregated.squeeze(1)))
class XLMRTruncatedClassifier(nn.Module):
"""Architecture used by the paper's unmasked XLM-R truncated baseline."""
def __init__(self, base_config: Any, base_model: nn.Module | None = None) -> None:
super().__init__()
self.xlmr = base_model or AutoModel.from_config(base_config)
self.dropout = nn.Dropout(base_config.hidden_dropout_prob)
self.classifier = nn.Linear(base_config.hidden_size, 3)
def forward(
self, input_ids: torch.Tensor, attention_mask: torch.Tensor, **_: Any
) -> torch.Tensor:
outputs = self.xlmr(input_ids=input_ids, attention_mask=attention_mask)
pooled = self.dropout(outputs.last_hidden_state[:, 0, :])
return self.classifier(pooled)
def _load_spacy(language: str):
model_name = "sl_core_news_sm" if language == "slovenian" else "hr_core_news_sm"
try:
import spacy
nlp = spacy.load(
model_name,
disable=["tagger", "parser", "attribute_ruler", "lemmatizer", "ner"],
)
if "sentencizer" not in nlp.pipe_names and "senter" not in nlp.pipe_names:
nlp.add_pipe("sentencizer")
return nlp
except Exception as exc:
print(f"Warning: {model_name} unavailable ({exc}); using regex sentence splitting.")
return None
def _sentences(text: str, nlp: Any) -> list[str]:
if nlp is not None:
return [sentence.text.strip() for sentence in nlp(text.strip()).sents if sentence.text.strip()]
return [
part.strip()
for part in re.split(r"(?<=[.!?])\s+|\n+", text.strip())
if part.strip()
]
class InferenceEngine:
def __init__(
self,
model_name: str,
language: str,
mode: str,
model_root: str | Path,
base_model_root: str | Path | None = None,
device: str = "auto",
) -> None:
if model_name not in MODEL_SPECS:
raise ValueError(f"Unknown model: {model_name}")
if language not in ("hbs", "slovenian") or mode not in ("masked", "unmasked"):
raise ValueError("language must be hbs/slovenian and mode masked/unmasked")
selection = CHECKPOINTS[(model_name, language, mode)]
self.weight_path = weight_path(model_root, model_name, language, mode)
if not selection["available"] and not self.weight_path.is_file():
raise FileNotFoundError(selection["unavailable_reason"])
if not self.weight_path.is_file():
raise FileNotFoundError(self.weight_path)
if device == "auto":
device = "cuda" if torch.cuda.is_available() else "cpu"
if device.startswith("cuda") and not torch.cuda.is_available():
raise RuntimeError("CUDA requested but unavailable.")
self.device = torch.device(device)
self.model_name = model_name
self.language = language
self.mode = mode
self.spec = model_spec(model_name, language)
self.base_path = _resolve_base(
self.spec,
base_model_root,
bundled_assets=self.weight_path.parent / "base_model",
)
self.backend = self.spec["backend"]
self.tokenizer = _load_tokenizer(
self.base_path, model_name, language, self.spec["max_length"]
)
self.spacy_nlp = None
self.aspect_token_id = None
self._load_model()
def _load_model(self) -> None:
local_only = Path(self.base_path).is_dir()
config = AutoConfig.from_pretrained(self.base_path, local_files_only=local_only)
state = _load_state(self.weight_path)
if self.backend == "han":
self.tokenizer.add_special_tokens(
{"additional_special_tokens": [HAN_ASPECT_TOKEN]}
)
model = SimplifiedDARTModel(
config,
len(self.tokenizer),
max_sentences=self.spec["max_sentences"],
)
self.aspect_token_id = self.tokenizer.convert_tokens_to_ids(HAN_ASPECT_TOKEN)
self.spacy_nlp = _load_spacy(self.language)
elif self.backend == "mt5":
model = AutoModelForSeq2SeqLM.from_config(config)
if self.mode == "masked":
self.tokenizer.add_special_tokens(
{"additional_special_tokens": [ASPECT_TOKEN]}
)
_resize_embeddings(model, len(self.tokenizer))
elif self.backend == "bge":
self._load_bge(state)
return
elif self.model_name == "xlmr" and self.mode == "unmasked":
model = XLMRTruncatedClassifier(config)
elif self.model_name == "longformer":
config_dict = config.to_dict()
config_dict.update(
num_labels=3,
problem_type="single_label_classification",
id2label={0: "negative", 1: "neutral", 2: "positive"},
label2id={"negative": 0, "neutral": 1, "positive": 2},
)
long_config = LongformerConfig(**config_dict)
if isinstance(long_config.attention_window, int):
long_config.attention_window = [
long_config.attention_window
] * long_config.num_hidden_layers
model = LongformerForSequenceClassification(long_config)
if self.mode == "masked":
self.tokenizer.add_special_tokens(
{"additional_special_tokens": [ASPECT_TOKEN]}
)
_resize_embeddings(model, len(self.tokenizer))
else:
config.num_labels = 3
config.problem_type = "single_label_classification"
config.id2label = {0: "negative", 1: "neutral", 2: "positive"}
config.label2id = {"negative": 0, "neutral": 1, "positive": 2}
model = AutoModelForSequenceClassification.from_config(config)
if self.mode == "masked":
self.tokenizer.add_special_tokens(
{"additional_special_tokens": [ASPECT_TOKEN]}
)
_resize_embeddings(model, len(self.tokenizer))
_resize_embeddings_to_checkpoint(model, state)
model.load_state_dict(state, strict=True)
del state
self.model = model.to(self.device)
self.model.eval()
def _load_bge(self, state: dict[str, torch.Tensor]) -> None:
from sentence_transformers import SentenceTransformer
self.embedding_model = SentenceTransformer(self.base_path, device=str(self.device))
linear_weights = [
(key, value)
for key, value in state.items()
if key.endswith(".weight") and value.ndim == 2
]
linear_weights.sort(key=lambda item: item[0])
if len(linear_weights) != 3:
raise ValueError("Expected three BGE MLP linear layers.")
dims = [linear_weights[0][1].shape[1]] + [value.shape[0] for _, value in linear_weights]
model = nn.Sequential(
nn.Linear(dims[0], dims[1]),
nn.ReLU(),
nn.Dropout(0.3),
nn.Linear(dims[1], dims[2]),
nn.ReLU(),
nn.Dropout(0.3),
nn.Linear(dims[2], dims[3]),
)
cleaned = {
key.removeprefix("classifier."): value for key, value in state.items()
}
model.load_state_dict(cleaned, strict=True)
self.model = model.to(self.device)
self.model.eval()
def _encoder_inputs(self, prepared: list[dict[str, Any]]) -> dict[str, torch.Tensor]:
tokenizer_kwargs: dict[str, Any] = {}
if self.spec.get("global_attention"):
attention_window = self.spec.get("attention_window", 512)
tokenizer_kwargs["pad_to_multiple_of"] = attention_window
encoded = self.tokenizer(
[item["model_article"] for item in prepared],
text_pair=[item["model_aspect"] for item in prepared],
truncation=True,
max_length=self.spec["max_length"],
padding=True,
return_tensors="pt",
**tokenizer_kwargs,
)
if self.spec.get("global_attention"):
global_attention = torch.zeros_like(encoded["attention_mask"])
global_attention[:, 0] = 1
encoded["global_attention_mask"] = global_attention
return {key: value.to(self.device) for key, value in encoded.items()}
def _han_inputs(self, prepared: list[dict[str, Any]]) -> dict[str, torch.Tensor]:
max_sentences = self.spec["max_sentences"]
max_length = self.spec["max_length"]
rows = []
for item in prepared:
if self.mode == "masked":
article = ASPECT_PATTERN.sub(HAN_ASPECT_TOKEN, item["article"])
else:
article = item["article"].replace("<aspect>", "").replace(
"</aspect>", ""
)
sentences = _sentences(article, self.spacy_nlp)[:max_sentences]
ids, masks, positions = [], [], []
for index, sentence in enumerate(sentences):
text = f"{HAN_ASPECT_TOKEN} {self.tokenizer.sep_token} {sentence}"
encoded = self.tokenizer(
text,
add_special_tokens=True,
max_length=max_length,
padding="max_length",
truncation=True,
return_tensors="pt",
)
ids.append(encoded["input_ids"].squeeze(0))
masks.append(encoded["attention_mask"].squeeze(0))
positions.append(index + 1)
pad_count = max_sentences - len(ids)
pad_ids = torch.full(
(max_length,), self.tokenizer.pad_token_id or 0, dtype=torch.long
)
pad_mask = torch.zeros(max_length, dtype=torch.long)
ids.extend([pad_ids] * pad_count)
masks.extend([pad_mask] * pad_count)
positions.extend([0] * pad_count)
sentence_mask = torch.zeros(max_sentences, dtype=torch.long)
sentence_mask[: len(sentences)] = 1
rows.append(
{
"input_ids": torch.stack(ids),
"attention_mask": torch.stack(masks),
"sentence_mask": sentence_mask,
"sentence_position_ids": torch.tensor(positions, dtype=torch.long),
}
)
return {
key: torch.stack([row[key] for row in rows]).to(self.device) for key in rows[0]
}
def _mt5_sources(self, prepared: list[dict[str, Any]]) -> dict[str, torch.Tensor]:
sources = [
f"classify sentiment\naspect: {item['model_aspect']}\narticle: {item['model_article']}"
for item in prepared
]
encoded = self.tokenizer(
sources,
truncation=True,
max_length=self.spec["max_length"],
padding=True,
return_tensors="pt",
)
return {key: value.to(self.device) for key, value in encoded.items()}
def _mt5_probabilities(self, inputs: dict[str, torch.Tensor]) -> torch.Tensor:
losses = []
batch_size = inputs["input_ids"].shape[0]
for label in LABEL_TEXTS:
ids = self.tokenizer(
label,
add_special_tokens=True,
truncation=True,
max_length=self.spec["max_target_length"],
return_tensors="pt",
)["input_ids"][0].to(self.device)
labels = ids.unsqueeze(0).repeat(batch_size, 1)
decoder_input_ids = self.model._shift_right(labels)
logits = self.model(
**inputs, decoder_input_ids=decoder_input_ids, use_cache=False
).logits
token_losses = F.cross_entropy(
logits.transpose(1, 2), labels, reduction="none"
)
losses.append(token_losses.mean(dim=1))
return torch.softmax(-torch.stack(losses, dim=1), dim=1)
def _bge_inputs(self, prepared: list[dict[str, Any]]) -> torch.Tensor:
texts = []
for item in prepared:
if self.mode == "masked":
text = ASPECT_PATTERN.sub("[ASPECT_MENTION]", item["article"])
text = f"{text.strip()} [ASPECT_NAME]"
else:
text = item["article"].replace("<aspect>", "").replace("</aspect>", "")
texts.append(text.strip())
embeddings = self.embedding_model.encode(texts, normalize_embeddings=True)
return torch.tensor(embeddings, dtype=torch.float32, device=self.device)
def _sample_probabilities(
self, prepared: list[dict[str, Any]], passes: int, seed: int
) -> np.ndarray:
if passes == 1 or passes < 0:
raise ValueError("mc_passes must be 0 or at least 2.")
count = passes or 1
if passes:
_seed_everything(seed)
_enable_dropout_only(self.model)
else:
self.model.eval()
if self.backend == "encoder":
inputs = self._encoder_inputs(prepared)
elif self.backend == "han":
inputs = self._han_inputs(prepared)
elif self.backend == "mt5":
inputs = self._mt5_sources(prepared)
else:
inputs = self._bge_inputs(prepared)
rows = []
with torch.inference_mode():
for _ in range(count):
if self.backend == "encoder":
output = self.model(**inputs)
logits = output if torch.is_tensor(output) else output.logits
probs = torch.softmax(logits, dim=-1)
elif self.backend == "han":
logits = self.model(
**inputs,
aspect_target_token_id=torch.tensor(
[self.aspect_token_id], device=self.device
),
)
probs = torch.softmax(logits, dim=-1)
elif self.backend == "mt5":
probs = self._mt5_probabilities(inputs)
else:
probs = torch.softmax(self.model(inputs), dim=-1)
rows.append(probs.float().cpu().numpy())
self.model.eval()
return np.stack(rows, axis=0)
def predict_batch(
self,
records: Iterable[dict[str, Any]],
batch_size: int = 8,
mc_passes: int = 0,
seed: int = 42,
) -> list[dict[str, Any]]:
records_list = list(records)
if not records_list or batch_size < 1:
raise ValueError("Input must be non-empty and batch_size >= 1.")
outputs = []
for start in range(0, len(records_list), batch_size):
prepared = [
prepare_record(record, self.mode)
for record in records_list[start : start + batch_size]
]
samples = self._sample_probabilities(prepared, mc_passes, seed + start)
for index, item in enumerate(prepared):
item_samples = samples[:, index, :]
mean_probs = item_samples.mean(axis=0)
sample_labels = [
ID_TO_SENTIMENT[int(value)] for value in item_samples.argmax(axis=1)
]
predicted = ID_TO_SENTIMENT[int(mean_probs.argmax())]
outputs.append(
{
"model": self.model_name,
"language": self.language,
"mode": self.mode,
"input_article": item["article"],
"tagged_aspects": item["tagged_aspects"],
"aspect_used": item["model_aspect"],
"gold_sentiment": item["gold_sentiment"],
"predicted_sentiment": predicted,
"predicted_sentiment_name": SENTIMENT_NAMES[predicted],
"class_probabilities": {
key: float(mean_probs[class_id])
for class_id, key in enumerate(CLASS_KEYS)
},
"uncertainty_across_classes": _uncertainty(
item_samples, sample_labels
),
"inference": {
"device": str(self.device),
"mc_dropout": bool(mc_passes),
"weight_file": str(self.weight_path),
},
}
)
return outputs
def predict(
self, record: dict[str, Any], mc_passes: int = 0, seed: int = 42
) -> dict[str, Any]:
return self.predict_batch(
[record], batch_size=1, mc_passes=mc_passes, seed=seed
)[0]
|