File size: 10,399 Bytes
818a2e9 a9c4a62 818a2e9 a9c4a62 818a2e9 a9c4a62 818a2e9 a9c4a62 818a2e9 a9c4a62 818a2e9 a9c4a62 818a2e9 a9c4a62 818a2e9 a9c4a62 818a2e9 | 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 | """
Domain Tokenizer Builder — assembles per-field tokenizers into an
HF-compatible PreTrainedTokenizerFast.
This is the core of domainTokenizer: it takes a DomainSchema, builds
per-field tokenizers, fits data-dependent ones, and produces a single
HuggingFace tokenizer that can encode domain events as token ID sequences.
The output tokenizer is fully compatible with HF Trainer, push_to_hub,
from_pretrained, etc.
IMPORTANT: Uses Whitespace pre-tokenizer (not ByteLevel) because our sequences
are space-separated special tokens like '[EVT_000] [PRICE_16] electronics.smartphone'.
ByteLevel would split these into byte-fragments causing massive UNK rates.
"""
import json
import os
from datetime import datetime
from pathlib import Path
from typing import Any, Dict, Iterator, List, Optional, Sequence, Tuple, Union
import numpy as np
from tokenizers import Tokenizer, decoders, models, pre_tokenizers, trainers
from transformers import PreTrainedTokenizerFast
from ..schema import DomainSchema, FieldSpec, FieldType
from .field_tokenizers import (
BaseFieldTokenizer,
CalendarTokenizer,
CategoricalTokenizer,
DiscreteNumericalTokenizer,
MagnitudeBucketTokenizer,
SignTokenizer,
create_field_tokenizer,
)
# Control tokens -- always present
CONTROL_TOKENS = ["[PAD]", "[UNK]", "[BOS]", "[EOS]", "[MASK]", "[CLS]", "[SEP]"]
class DomainTokenizerBuilder:
"""Builds an HF-compatible tokenizer from a DomainSchema.
Workflow:
1. builder = DomainTokenizerBuilder(schema)
2. builder.fit(events) # fit magnitude bins etc.
3. hf_tok = builder.build(text_corpus) # build HF tokenizer
4. tokens = builder.tokenize_event(event) # tokenize a single event
5. ids = hf_tok(tokens_str) # convert to IDs
Or use the convenience method:
6. ids = builder.encode_event(event, hf_tok) # event -> IDs in one call
7. ids = builder.encode_sequence(events, hf_tok) # full sequence -> IDs
"""
def __init__(self, schema: DomainSchema):
self.schema = schema
self.field_tokenizers: Dict[str, Optional[BaseFieldTokenizer]] = {}
self._is_fitted = False
self._build_field_tokenizers()
def _build_field_tokenizers(self):
for spec in self.schema.fields:
self.field_tokenizers[spec.name] = create_field_tokenizer(spec)
def fit(self, events: Sequence[Dict[str, Any]]) -> "DomainTokenizerBuilder":
"""Fit data-dependent tokenizers on training events."""
for spec in self.schema.fields:
if spec.field_type == FieldType.NUMERICAL_CONTINUOUS:
tok = self.field_tokenizers[spec.name]
values = []
for event in events:
v = event.get(spec.name) if isinstance(event, dict) else getattr(event, spec.name, None)
if v is not None:
values.append(float(v))
if values:
tok.fit(np.array(values))
else:
raise ValueError(f"No values found for field '{spec.name}' during fitting")
self._is_fitted = True
return self
@property
def is_fitted(self) -> bool:
if not self.schema.fittable_field_names:
return True
return self._is_fitted
def _collect_special_tokens(self) -> List[str]:
"""Collect all special tokens: control + event separator + per-field domain tokens."""
tokens = list(CONTROL_TOKENS)
tokens.append(self.schema.event_separator)
for spec in self.schema.fields:
tok = self.field_tokenizers.get(spec.name)
if tok is not None and hasattr(tok, "vocab"):
tokens.extend(tok.vocab)
seen = set()
unique = []
for t in tokens:
if t not in seen:
seen.add(t)
unique.append(t)
return unique
def build(
self,
text_corpus: Optional[Iterator[str]] = None,
bpe_vocab_size: int = 8000,
min_frequency: int = 2,
) -> PreTrainedTokenizerFast:
"""Build a complete HuggingFace-compatible tokenizer.
Uses Whitespace pre-tokenizer (not ByteLevel) because domain token
sequences are space-separated: '[EVT_000] [PRICE_16] electronics.smartphone'.
Whitespace splits on spaces, keeping special tokens intact. BPE then handles
subword splitting within text tokens (e.g., 'electronics.smartphone' ->
'electronics', '.', 'smartphone').
"""
for name in self.schema.fittable_field_names:
tok = self.field_tokenizers[name]
if isinstance(tok, MagnitudeBucketTokenizer) and not tok._is_fitted:
raise RuntimeError(
f"Field '{name}' requires fitting. Call builder.fit(events) first."
)
all_special_tokens = self._collect_special_tokens()
# Build BPE tokenizer with Whitespace pre-tokenizer
# Whitespace splits on spaces → each special token like [EVT_000] stays intact
# BPE merges then handle subword splitting within text fields
base_tokenizer = Tokenizer(models.BPE(unk_token="[UNK]"))
base_tokenizer.pre_tokenizer = pre_tokenizers.Whitespace()
base_tokenizer.decoder = decoders.BPEDecoder()
if self.schema.has_text_fields and text_corpus is not None:
trainer_obj = trainers.BpeTrainer(
vocab_size=bpe_vocab_size,
special_tokens=all_special_tokens,
min_frequency=min_frequency,
show_progress=True,
)
if isinstance(text_corpus, (list, tuple)):
base_tokenizer.train_from_iterator(iter(text_corpus), trainer=trainer_obj)
else:
base_tokenizer.train_from_iterator(text_corpus, trainer=trainer_obj)
else:
# No text fields — vocabulary-only tokenizer
vocab = {token: i for i, token in enumerate(all_special_tokens)}
merges = []
base_tokenizer = Tokenizer(models.BPE(vocab=vocab, merges=merges, unk_token="[UNK]"))
base_tokenizer.pre_tokenizer = pre_tokenizers.Whitespace()
base_tokenizer.decoder = decoders.BPEDecoder()
hf_tokenizer = PreTrainedTokenizerFast(
tokenizer_object=base_tokenizer,
bos_token="[BOS]",
eos_token="[EOS]",
pad_token="[PAD]",
unk_token="[UNK]",
mask_token="[MASK]",
cls_token="[CLS]",
sep_token="[SEP]",
)
return hf_tokenizer
def tokenize_event(self, event: Union[Dict[str, Any], Any]) -> List[str]:
"""Convert a single domain event into a list of token strings."""
tokens = []
for spec in self.schema.fields:
if isinstance(event, dict):
value = event.get(spec.name)
else:
value = getattr(event, spec.name, None)
if spec.field_type == FieldType.TEXT:
if value is not None:
tokens.append(str(value))
continue
tok = self.field_tokenizers.get(spec.name)
if tok is None:
continue
if value is None:
tokens.append("[UNK]")
continue
result = tok(value)
if isinstance(result, list):
tokens.extend(result)
else:
tokens.append(result)
return tokens
def tokenize_sequence(
self,
events: Sequence[Union[Dict[str, Any], Any]],
add_bos: bool = True,
add_eos: bool = True,
) -> List[str]:
"""Tokenize a full sequence of events into token strings."""
all_tokens = []
if add_bos:
all_tokens.append("[BOS]")
for i, event in enumerate(events):
if i > 0:
all_tokens.append(self.schema.event_separator)
event_tokens = self.tokenize_event(event)
all_tokens.extend(event_tokens)
if add_eos:
all_tokens.append("[EOS]")
return all_tokens
def encode_sequence(
self,
events: Sequence[Union[Dict[str, Any], Any]],
hf_tokenizer: PreTrainedTokenizerFast,
max_length: int = 2048,
add_bos: bool = True,
add_eos: bool = True,
return_tensors: Optional[str] = None,
) -> Dict[str, Any]:
"""Full pipeline: events -> token strings -> token IDs."""
token_strings = self.tokenize_sequence(events, add_bos=add_bos, add_eos=add_eos)
token_text = " ".join(token_strings)
encoding = hf_tokenizer(
token_text,
max_length=max_length,
truncation=True,
padding="max_length",
return_tensors=return_tensors,
)
return encoding
def save(self, directory: str):
"""Save the builder state (fitted bins, schema, etc.) to a directory."""
os.makedirs(directory, exist_ok=True)
state = {
"schema_name": self.schema.name,
"is_fitted": self._is_fitted,
"field_tokenizers": {},
}
for name, tok in self.field_tokenizers.items():
if tok is not None:
state["field_tokenizers"][name] = tok.to_dict()
with open(os.path.join(directory, "domain_tokenizer_builder.json"), "w") as f:
json.dump(state, f, indent=2)
def get_stats(self) -> Dict[str, Any]:
"""Return statistics about the tokenizer configuration."""
return {
"schema_name": self.schema.name,
"total_fields": len(self.schema.fields),
"special_token_count": self.schema.special_token_count,
"fixed_tokens_per_event": self.schema.fixed_tokens_per_event,
"has_text_fields": self.schema.has_text_fields,
"is_fitted": self.is_fitted,
"field_details": {
spec.name: {
"type": spec.field_type.value,
"vocab_tokens": spec.token_count,
"tokens_per_event": spec.tokens_per_event,
}
for spec in self.schema.fields
},
}
|