File size: 9,090 Bytes
2837752 9252eda 2837752 56c9498 2837752 9252eda 2837752 1b09699 2837752 1b09699 2837752 be95eb6 1b09699 d5c9b75 be95eb6 1b09699 9252eda 86d4840 9252eda 86d4840 9252eda 1b09699 be95eb6 9252eda 15cd490 1b09699 2837752 be95eb6 1b09699 2837752 36f55fa 1b09699 36f55fa 1b09699 36f55fa 2837752 1b09699 2837752 1b09699 36f55fa 86d4840 36f55fa d5c9b75 56c9498 d5c9b75 2837752 9252eda 2837752 9252eda d5c9b75 9252eda d5c9b75 9252eda 2837752 9252eda 2837752 36f55fa 2837752 1b09699 2837752 9252eda 86d4840 9252eda 2837752 9252eda 2837752 9252eda d5c9b75 9252eda | 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 | from collections.abc import Generator, Iterable
from dataclasses import dataclass
from enum import StrEnum
from itertools import chain
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from transformers import (
AutoConfig,
AutoModel,
AutoTokenizer,
ModernBertModel,
PreTrainedConfig,
PreTrainedModel,
)
from transformers.modeling_outputs import TokenClassifierOutput
BATCH_SIZE = 64
class ModelURI(StrEnum):
BASE = "answerdotai/ModernBERT-base"
LARGE = "answerdotai/ModernBERT-large"
@dataclass(slots=True, frozen=True)
class LexicalExample:
concept: str
definition: str
@dataclass(slots=True, frozen=True)
class PaddedBatch:
input_ids: torch.Tensor
attention_mask: torch.Tensor
class DisamBert(PreTrainedModel):
def __init__(self, config: PreTrainedConfig):
super().__init__(config)
if config.init_basemodel:
self.BaseModel = AutoModel.from_pretrained(config.name_or_path, device_map="auto")
self.classifier_head = nn.UninitializedParameter()
self.bias = nn.UninitializedParameter()
self.__entities = None
else:
self.BaseModel = ModernBertModel(config)
self.classifier_head = nn.Parameter(
torch.empty((config.ontology_size, config.hidden_size))
)
self.bias = nn.Parameter(torch.empty((config.ontology_size, 1)))
self.__entities = pd.Series(config.entities)
config.init_basemodel = False
self.tokenizer = AutoTokenizer.from_pretrained(config.tokenizer_path)
self.loss = nn.CrossEntropyLoss()
self.post_init()
@classmethod
def from_base(cls, base_id: ModelURI):
config = AutoConfig.from_pretrained(base_id)
config.init_basemodel = True
config.tokenizer_path = base_id
return cls(config)
def init_classifier(self, entities: Generator[LexicalExample]) -> None:
entity_ids = []
vectors = []
batch = []
n = 0
with self.BaseModel.device:
for entity in entities:
entity_ids.append(entity.concept)
batch.append(entity.definition)
n += 1
if n == BATCH_SIZE:
tokens = self.tokenizer(batch, padding=True, return_tensors="pt")
encoding = self.BaseModel(tokens["input_ids"], tokens["attention_mask"])
vectors.append(encoding.last_hidden_state.detach()[:, 0])
n = 0
batch = []
if n > 0:
tokens = self.tokenizer(batch, padding=True, return_tensors="pt")
encoding = self.BaseModel(tokens["input_ids"], tokens["attention_mask"])
vectors.append(encoding.last_hidden_state.detach()[:, 0])
self.__entities = pd.Series(entity_ids)
self.config.entities = entity_ids
self.config.ontology_size = len(entity_ids)
self.classifier_head = nn.Parameter(torch.cat(vectors, dim=0))
self.bias = nn.Parameter(
torch.nn.init.normal_(
torch.empty((self.config.ontology_size, 1)),
std=self.classifier_head.std().item() * np.sqrt(self.config.hidden_size)
)
)
@property
def entities(self) -> pd.Series:
if self.__entities is None and hasattr(self.config, "entities"):
self.__entities = pd.Series(self.config.entities)
return self.__entities
def forward(
self,
input_ids: torch.Tensor,
attention_mask: torch.Tensor,
lengths: list[list[int]],
candidates: list[list[list[int]]],
labels: Iterable[list[int]] | None = None,
output_hidden_states: bool = False,
output_attentions: bool = False,
) -> TokenClassifierOutput:
assert not nn.parameter.is_lazy(self.classifier_head), (
"Run init_classifier to initialise weights"
)
base_model_output = self.BaseModel(
input_ids,
attention_mask,
output_hidden_states=output_hidden_states,
output_attentions=output_attentions,
)
token_vectors = base_model_output.last_hidden_state
span_vectors = torch.cat(
[
torch.vstack(
[
torch.sum(chunk, dim=0)
for chunk in self.split(token_vectors[i], sentence_indices)
]
)
for (i, sentence_indices) in enumerate(lengths)
]
)
logits = torch.einsum("ij,kj->ki", span_vectors, self.classifier_head) + self.bias
logits1 = logits - logits.min()
mask = torch.zeros_like(logits)
for i, concepts in enumerate(chain.from_iterable(candidates)):
mask[concepts, i] = torch.tensor(1.0)
logits2 = logits1 * mask
sentence_lengths = [len(sentence_indices) for sentence_indices in lengths]
maxlen = max(sentence_lengths)
split_logits = torch.split(logits2, sentence_lengths, dim=1)
logits3 = torch.stack(
[
self.extend_to_max_length(sentence, length, maxlen)
for (sentence, length) in zip(split_logits, sentence_lengths, strict=True)
]
)
return TokenClassifierOutput(
logits=logits3,
loss=self.loss(logits3, labels) if labels is not None else None,
hidden_states=base_model_output.hidden_states if output_hidden_states else None,
attentions=base_model_output.attentions if output_attentions else None,
)
def split(self, vectors: torch.Tensor, lengths: list[int]) -> tuple[torch.Tensor, ...]:
maxlen = vectors.shape[0]
total_length = sum(lengths)
is_padded = total_length < maxlen
chunks = vectors.split((lengths + [maxlen - total_length]) if is_padded else lengths)
return chunks[:-1] if is_padded else chunks
def pad(self, tokens: Iterable[list[int]]) -> PaddedBatch:
lengths = [len(sentence) for sentence in tokens]
maxlen = max(lengths)
input_ids = torch.tensor(
[
sentence + [self.config.pad_token_id] * (maxlen - length)
for (sentence, length) in zip(tokens, lengths, strict=True)
]
)
attention_mask = torch.vstack(
[torch.cat((torch.ones(length), torch.zeros(maxlen - length))) for length in lengths]
)
return PaddedBatch(input_ids, attention_mask)
def extend_to_max_length(
self, sentence: torch.Tensor, length: int, maxlength: int
) -> torch.Tensor:
with self.BaseModel.device:
return (
torch.cat(
[
sentence,
torch.zeros((self.config.ontology_size, maxlength - length)),
],
dim=1,
)
if length < maxlength
else sentence
)
def pad_labels(self, labels: list[list[int]]) -> torch.Tensor:
unk = len(self.config.entities) - 1
lengths = [len(seq) for seq in labels]
maxlen = max(lengths)
with self.BaseModel.device:
return torch.tensor(
[
seq + [unk] * (maxlen - length)
for (seq, length) in zip(labels, lengths, strict=True)
]
)
def tokenize(
self, batch: list[dict[str, str | list[int]]]
) -> dict[str, torch.Tensor | list[list[int]]]:
all_indices = []
all_tokens = []
with self.BaseModel.device:
for example in batch:
text = example["text"]
span_indices = example["indices"]
indices = []
tokens = []
last_span = len(span_indices) - 2
for i, position in enumerate(span_indices[:-1]):
span = text[position : span_indices[i + 1]]
span_tokens = self.tokenizer([span], padding=False)["input_ids"][0]
if i > 0:
span_tokens = span_tokens[1:]
if i < last_span:
span_tokens = span_tokens[:-1]
indices.append(len(span_tokens))
tokens.extend(span_tokens)
all_indices.append(indices)
all_tokens.append(tokens)
padded = self.pad(all_tokens)
result = {
"input_ids": padded.input_ids,
"attention_mask": padded.attention_mask,
"lengths": all_indices,
"candidates": [example["candidates"] for example in batch],
}
if "labels" in batch[0]:
result["labels"] = self.pad_labels([example["labels"] for example in batch])
return result
|