File size: 19,136 Bytes
d40eb53 | 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 | import os
import json
import joblib
import numpy as np
from utils import clean_text, LABEL_COLS
# torch et transformers sont optionnels — uniquement requis pour le Teacher XLM-R
try:
import torch
import torch.nn as nn
from transformers import AutoTokenizer, AutoModel
TORCH_AVAILABLE = True
except ImportError:
TORCH_AVAILABLE = False
# ── HuggingFace Hub — fallback si fichiers absents en local ─────────────────
HF_REPO_ID = 'BHM03/amane-models'
def _resolve(local_path: str, hf_filename: str) -> str:
"""Retourne le chemin local si le fichier existe, sinon télécharge depuis HF Hub."""
if os.path.exists(local_path):
return local_path
try:
from huggingface_hub import hf_hub_download
return hf_hub_download(
repo_id=HF_REPO_ID,
filename=hf_filename,
token=os.environ.get('HF_TOKEN'),
)
except Exception as e:
raise FileNotFoundError(
f'Fichier introuvable localement ({local_path}) '
f'et téléchargement HF Hub échoué : {e}'
)
# ── Chemins modèles ──────────────────────────────────────────────────────────
BENCHMARK_MODEL_PATH = 'models/ovr_balanced.joblib'
BENCHMARK_TFIDF_PATH = 'models/tfidf_vectorizer.joblib'
TEACHER_CKPT_PATH = 'models/teacher/teacher_best.pt'
TEACHER_TOKENIZER_PATH = 'models/teacher/tokenizer'
TEACHER_THRESH_PATH = 'models/teacher/thresholds.json'
TEACHER_STAR_CKPT_PATH = 'models/teacher_star/teacher_star.pt'
TEACHER_STAR_THRESH_PATH= 'models/teacher_star/threshold_darija.json'
STUDENT_CKPT_PATH = 'models/student/student_best.pt'
STUDENT_THRESH_PATH = 'models/student/threshold_student.json'
# ── Hyperparamètres d'inférence Teacher (cohérents avec l'entraînement) ─────
TEACHER_MAX_LEN = 256
TEACHER_STAR_MAX_LEN = 128
STUDENT_MAX_LEN = 128
TEACHER_MODEL_HF = 'xlm-roberta-base'
STUDENT_MODEL_HF = 'distilbert-base-multilingual-cased'
# ════════════════════════════════════════════════════════════════════════════
# BENCHMARK — OneVsRest + LogisticRegression
# ════════════════════════════════════════════════════════════════════════════
class BenchmarkPredictor:
def __init__(self):
self.model = None
self.tfidf = None
self.loaded = False
self.error = None
def load(self):
try:
self.model = joblib.load(_resolve(BENCHMARK_MODEL_PATH, 'ovr_balanced.joblib'))
self.tfidf = joblib.load(_resolve(BENCHMARK_TFIDF_PATH, 'tfidf_vectorizer.joblib'))
self.loaded = True
except Exception as e:
self.error = str(e)
self.loaded = False
def predict(self, text: str) -> dict:
if not self.loaded:
raise RuntimeError(f'Benchmark non chargé : {self.error}')
cleaned = clean_text(text).lower()
vec = self.tfidf.transform([cleaned])
probs = self.model.predict_proba(vec)[0]
return {col: float(round(p, 4)) for col, p in zip(LABEL_COLS, probs)}
def predict_batch(self, texts: list) -> list[dict]:
if not self.loaded:
raise RuntimeError(f'Benchmark non chargé : {self.error}')
cleaned = [clean_text(t).lower() for t in texts]
vecs = self.tfidf.transform(cleaned)
probs = self.model.predict_proba(vecs)
return [
{col: float(round(p, 4)) for col, p in zip(LABEL_COLS, row)}
for row in probs
]
# ════════════════════════════════════════════════════════════════════════════
# XLM-R TEACHER
# ════════════════════════════════════════════════════════════════════════════
class TeacherPredictor:
def __init__(self):
self.model = None
self.tokenizer = None
self.thresholds = None
self.loaded = False
self.error = None
self.device = None
def load(self):
if not TORCH_AVAILABLE:
self.error = (
'PyTorch / Transformers non installés. '
'Lance : pip install torch transformers sentencepiece'
)
self.loaded = False
return
try:
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# 2. Charger le tokenizer — on utilise directement 'xlm-roberta-base'
# car le tokenizer ne change pas lors du fine-tuning, et les fichiers
# locaux sont incomplets (sentencepiece.bpe.model manquant).
self.tokenizer = AutoTokenizer.from_pretrained(TEACHER_MODEL_HF)
# 3. Reconstruire l'architecture (identique au notebook d'entraînement)
class TeacherModel(nn.Module):
def __init__(self, model_name: str, num_labels: int = 6, dropout: float = 0.1):
super().__init__()
self.encoder = AutoModel.from_pretrained(model_name)
hidden_size = self.encoder.config.hidden_size
self.dropout = nn.Dropout(dropout)
self.classifier = nn.Linear(hidden_size, num_labels)
def mean_pool(self, token_embeds, attention_mask):
mask = attention_mask.unsqueeze(-1).expand(token_embeds.size()).float()
summed = (token_embeds * mask).sum(dim=1)
counts = mask.sum(dim=1).clamp(min=1e-9)
return summed / counts
def forward(self, input_ids, attention_mask):
outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
pooled = self.mean_pool(outputs.last_hidden_state, attention_mask)
pooled = self.dropout(pooled)
return self.classifier(pooled)
# 4. Charger le checkpoint
ckpt = torch.load(_resolve(TEACHER_CKPT_PATH, 'teacher_best.pt'), map_location=self.device, weights_only=False)
model = TeacherModel(TEACHER_MODEL_HF, num_labels=len(LABEL_COLS))
model.load_state_dict(ckpt['model_state'])
model.to(self.device)
model.eval()
self.model = model
# 5. Charger les seuils tunés
if os.path.exists(TEACHER_THRESH_PATH):
with open(TEACHER_THRESH_PATH) as f:
self.thresholds = json.load(f)
else:
# Fallback : seuil 0.5 par défaut
self.thresholds = {col: 0.5 for col in LABEL_COLS}
self.loaded = True
except Exception as e:
self.error = f'{type(e).__name__}: {e}'
self.loaded = False
# ── Inférence : probabilités ───────────────────────────────────────────
def predict(self, text: str) -> dict:
"""Retourne les probabilités sigmoïdes par label."""
if not self.loaded:
raise RuntimeError(f'Teacher non chargé : {self.error}')
cleaned = clean_text(text)
encoding = self.tokenizer(
cleaned,
max_length=TEACHER_MAX_LEN,
padding='max_length',
truncation=True,
return_tensors='pt',
)
input_ids = encoding['input_ids'].to(self.device)
attn_mask = encoding['attention_mask'].to(self.device)
with torch.no_grad():
logits = self.model(input_ids, attn_mask)
probs = torch.sigmoid(logits).cpu().numpy()[0]
return {col: float(round(p, 4)) for col, p in zip(LABEL_COLS, probs)}
# ── Inférence : labels binaires (avec seuils tunés) ────────────────────
def predict_with_labels(self, text: str) -> dict:
"""
Retourne probabilités + labels binaires selon les seuils tunés.
Format : {'probs': {...}, 'labels': {...}, 'thresholds': {...}}
"""
probs = self.predict(text)
labels = {col: int(probs[col] >= self.thresholds[col]) for col in LABEL_COLS}
return {
'probs' : probs,
'labels' : labels,
'thresholds' : self.thresholds,
}
# ── Inférence : batch ──────────────────────────────────────────────────
def predict_batch(self, texts: list) -> list[dict]:
"""Inférence batch — plus efficace que predict() en boucle."""
if not self.loaded:
raise RuntimeError(f'Teacher non chargé : {self.error}')
cleaned = [clean_text(t) for t in texts]
encoding = self.tokenizer(
cleaned,
max_length=TEACHER_MAX_LEN,
padding='max_length',
truncation=True,
return_tensors='pt',
)
input_ids = encoding['input_ids'].to(self.device)
attn_mask = encoding['attention_mask'].to(self.device)
with torch.no_grad():
logits = self.model(input_ids, attn_mask)
probs = torch.sigmoid(logits).cpu().numpy()
return [
{col: float(round(p, 4)) for col, p in zip(LABEL_COLS, row)}
for row in probs
]
# ════════════════════════════════════════════════════════════════════════════
# XLM-R TEACHER* — adapté Darija (fine-tuné sur DarLoad, loss masquée)
# ════════════════════════════════════════════════════════════════════════════
class TeacherStarPredictor:
def __init__(self):
self.model = None
self.tokenizer = None
self.threshold = 0.45 # seuil tunée sur DarLoad
self.loaded = False
self.error = None
self.device = None
def load(self):
if not TORCH_AVAILABLE:
self.error = 'PyTorch / Transformers non installés.'
self.loaded = False
return
try:
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.tokenizer = AutoTokenizer.from_pretrained(TEACHER_MODEL_HF)
class TeacherModel(nn.Module):
def __init__(self, model_name, num_labels=6, dropout=0.1):
super().__init__()
self.encoder = AutoModel.from_pretrained(model_name)
self.dropout = nn.Dropout(dropout)
self.classifier = nn.Linear(self.encoder.config.hidden_size, num_labels)
def mean_pool(self, token_embeds, attention_mask):
mask = attention_mask.unsqueeze(-1).expand(token_embeds.size()).float()
summed = (token_embeds * mask).sum(dim=1)
return summed / mask.sum(dim=1).clamp(min=1e-9)
def forward(self, input_ids, attention_mask):
out = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
pooled = self.mean_pool(out.last_hidden_state, attention_mask)
return self.classifier(self.dropout(pooled))
ckpt = torch.load(_resolve(TEACHER_STAR_CKPT_PATH, 'teacher_star.pt'), map_location=self.device, weights_only=False)
model = TeacherModel(TEACHER_MODEL_HF, num_labels=len(LABEL_COLS))
model.load_state_dict(ckpt['model_state'])
model.to(self.device)
model.eval()
self.model = model
if os.path.exists(TEACHER_STAR_THRESH_PATH):
with open(TEACHER_STAR_THRESH_PATH) as f:
self.threshold = json.load(f).get('toxic_darija', 0.45)
self.loaded = True
except Exception as e:
self.error = f'{type(e).__name__}: {e}'
self.loaded = False
def predict(self, text: str) -> dict:
if not self.loaded:
raise RuntimeError(f'Teacher* non chargé : {self.error}')
cleaned = clean_text(text)
encoding = self.tokenizer(
cleaned, max_length=TEACHER_STAR_MAX_LEN,
padding='max_length', truncation=True, return_tensors='pt',
)
with torch.no_grad():
logits = self.model(
encoding['input_ids'].to(self.device),
encoding['attention_mask'].to(self.device),
)
probs = torch.sigmoid(logits).cpu().numpy()[0]
return {col: float(round(p, 4)) for col, p in zip(LABEL_COLS, probs)}
def predict_batch(self, texts: list) -> list[dict]:
if not self.loaded:
raise RuntimeError(f'Teacher* non chargé : {self.error}')
cleaned = [clean_text(t) for t in texts]
encoding = self.tokenizer(
cleaned, max_length=TEACHER_STAR_MAX_LEN,
padding='max_length', truncation=True, return_tensors='pt',
)
with torch.no_grad():
logits = self.model(
encoding['input_ids'].to(self.device),
encoding['attention_mask'].to(self.device),
)
probs = torch.sigmoid(logits).cpu().numpy()
return [{col: float(round(p, 4)) for col, p in zip(LABEL_COLS, row)} for row in probs]
# ════════════════════════════════════════════════════════════════════════════
# FACTORY — chargement unique via st.cache_resource
# ════════════════════════════════════════════════════════════════════════════
def load_benchmark() -> BenchmarkPredictor:
p = BenchmarkPredictor()
p.load()
return p
def load_teacher() -> TeacherPredictor:
p = TeacherPredictor()
p.load()
return p
def load_teacher_star() -> TeacherStarPredictor:
p = TeacherStarPredictor()
p.load()
return p
# ════════════════════════════════════════════════════════════════════════════
# STUDENT — DistilBERT multilingue spécialisé Darija
# ════════════════════════════════════════════════════════════════════════════
class StudentPredictor:
def __init__(self):
self.model = None
self.tokenizer = None
self.threshold = 0.6
self.loaded = False
self.error = None
self.device = None
def load(self):
if not TORCH_AVAILABLE:
self.error = 'PyTorch / Transformers non installés.'
self.loaded = False
return
try:
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.tokenizer = AutoTokenizer.from_pretrained(STUDENT_MODEL_HF)
class StudentModel(nn.Module):
def __init__(self, model_name, num_labels=6, dropout=0.1):
super().__init__()
self.encoder = AutoModel.from_pretrained(model_name)
self.dropout = nn.Dropout(dropout)
self.classifier = nn.Linear(self.encoder.config.hidden_size, num_labels)
def forward(self, input_ids, attention_mask):
out = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
pooled = out.last_hidden_state[:, 0, :] # token [CLS]
return self.classifier(self.dropout(pooled))
ckpt = torch.load(_resolve(STUDENT_CKPT_PATH, 'student_best.pt'), map_location=self.device, weights_only=False)
model = StudentModel(STUDENT_MODEL_HF, num_labels=len(LABEL_COLS))
model.load_state_dict(ckpt['model_state'])
model.to(self.device)
model.eval()
self.model = model
if os.path.exists(STUDENT_THRESH_PATH):
with open(STUDENT_THRESH_PATH) as f:
self.threshold = json.load(f).get('toxic_darija', 0.35)
self.loaded = True
except Exception as e:
self.error = f'{type(e).__name__}: {e}'
self.loaded = False
def predict(self, text: str) -> dict:
if not self.loaded:
raise RuntimeError(f'Student non chargé : {self.error}')
cleaned = clean_text(text)
encoding = self.tokenizer(
cleaned, max_length=STUDENT_MAX_LEN,
padding='max_length', truncation=True, return_tensors='pt',
)
with torch.no_grad():
logits = self.model(
encoding['input_ids'].to(self.device),
encoding['attention_mask'].to(self.device),
)
probs = torch.sigmoid(logits).cpu().numpy()[0]
return {col: float(round(p, 4)) for col, p in zip(LABEL_COLS, probs)}
def predict_batch(self, texts: list) -> list[dict]:
if not self.loaded:
raise RuntimeError(f'Student non chargé : {self.error}')
cleaned = [clean_text(t) for t in texts]
encoding = self.tokenizer(
cleaned, max_length=STUDENT_MAX_LEN,
padding='max_length', truncation=True, return_tensors='pt',
)
with torch.no_grad():
logits = self.model(
encoding['input_ids'].to(self.device),
encoding['attention_mask'].to(self.device),
)
probs = torch.sigmoid(logits).cpu().numpy()
return [{col: float(round(p, 4)) for col, p in zip(LABEL_COLS, row)} for row in probs]
def load_student() -> StudentPredictor:
p = StudentPredictor()
p.load()
return p
|