mStyleDistance-finetunned / model_loader.py
Doffy143's picture
Upload 3 files
4a675a3 verified
Raw
History Blame Contribute Delete
11.7 kB
"""
Model Loader β€” StyleAIClassifierV5 (v9)
Pipeline completo:
1. Preprocesamiento del texto
2. Inferencia con SentenceTransformer + clasificador (3 pases TTA)
3. CalibraciΓ³n isotΓ³nica
4. AbstenciΓ³n en zona de incertidumbre [0.44, 0.56]
"""
import os
import re
import gc
import json
import random
import pickle
import warnings
from pathlib import Path
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from sentence_transformers import SentenceTransformer
from huggingface_hub import hf_hub_download
warnings.filterwarnings("ignore")
# ── Repositorio Hugging Face con los pesos del modelo ──────────────────────────
HF_REPO_ID = "Doffy143/mStyleDistance-finetunned"
# ── Rutas locales (para desarrollo) ───────────────────────────────────────────
_MODELS_DIR = Path(__file__).parent.parent / "models"
# ── Archivos del modelo y sus nombres en el repositorio HF ────────────────────
_MODEL_FILES = {
"best_model.pt": _MODELS_DIR / "best_model.pt",
"isotonic_calibrator.pkl": _MODELS_DIR / "isotonic_calibrator.pkl",
"threshold_config.json": _MODELS_DIR / "threshold_config.json",
}
def _resolve_file(filename: str) -> str:
"""Busca el archivo local; si no existe, lo descarga desde Hugging Face."""
local_path = _MODEL_FILES[filename]
if local_path.exists():
print(f"[model_loader] '{filename}' encontrado localmente.")
return str(local_path)
print(f"[model_loader] '{filename}' no encontrado localmente, descargando desde HF...")
downloaded = hf_hub_download(
repo_id=HF_REPO_ID,
filename=filename,
cache_dir=os.environ.get("HF_HOME", None),
)
print(f"[model_loader] '{filename}' descargado en: {downloaded}")
return downloaded
# ── Rutas resueltas (se calculan en load_model) ──────────────────────────────
MODEL_PATH = None
CALIBRATOR_PATH = None
THRESHOLD_PATH = None
# ── HiperparΓ‘metros de inferencia (alineados con CONFIG del notebook v9) ───────
MODEL_NAME = "StyleDistance/mStyleDistance"
MAX_LENGTH = 256
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# ── Arquitectura (StyleAIClassifierV5 β€” igual que en el notebook v9) ──────────
class StyleAIClassifierV5(nn.Module):
"""
Encoder SentenceTransformer + cabeza clasificadora 768β†’384β†’2.
IdΓ©ntico al notebook de entrenamiento v9.
"""
def __init__(self, encoder_model, num_classes: int = 2,
dropout: float = 0.20, hidden_dim: int = 384):
super().__init__()
self.encoder = encoder_model
self.embedding_dim = encoder_model.get_sentence_embedding_dimension()
self.classifier = nn.Sequential(
nn.LayerNorm(self.embedding_dim),
nn.Dropout(dropout),
nn.Linear(self.embedding_dim, hidden_dim),
nn.GELU(),
nn.LayerNorm(hidden_dim),
nn.Dropout(dropout),
nn.Linear(hidden_dim, num_classes),
)
def forward(self, input_ids, attention_mask):
features = {"input_ids": input_ids, "attention_mask": attention_mask}
embeddings = self.encoder(features)["sentence_embedding"].to(torch.float32)
return self.classifier(embeddings)
# ── Preprocesamiento (igual al notebook v9) ────────────────────────────────────
def _preprocess(text: str) -> str:
"""Elimina URLs, @usuarios, #hashtags y normaliza espacios preservando pΓ‘rrafos."""
if not isinstance(text, str) or not text.strip():
return ""
text = re.sub(r"http\S+|www\S+|https\S+", "", text, flags=re.MULTILINE)
text = re.sub(r"@\w+|#\w+", "", text)
text = re.sub(r"[^\S\n]+", " ", text)
text = re.sub(r"\n\s*\n", "\n\n", text)
return text.strip()
# ── Variantes TTA (sobre strings, no sobre tokens) ────────────────────────────
def _tta_truncate(text: str, ratio: float = 0.85) -> str:
"""Recorta el texto al ratio indicado de palabras."""
words = text.split()
if len(words) < 20:
return text
return " ".join(words[: int(len(words) * ratio)])
def _tta_span_mask(text: str, span_ratio: float = 0.10) -> str:
"""Enmascara un span aleatorio del texto."""
words = text.split()
if len(words) < 10:
return text
n = max(1, int(len(words) * span_ratio))
start = random.randint(0, max(0, len(words) - n))
words[start : start + n] = ["[MASK]"] * n
return " ".join(words)
# ── TokenizaciΓ³n y predicciΓ³n en batch ────────────────────────────────────────
def _encode_and_predict(model: StyleAIClassifierV5, texts: list[str],
tokenizer, max_length: int) -> list[float]:
"""Devuelve la probabilidad de clase IA (Γ­ndice 1) para cada texto."""
encoding = tokenizer(
texts,
max_length=max_length,
padding="max_length",
truncation=True,
return_tensors="pt",
)
input_ids = encoding["input_ids"].to(DEVICE)
attention_mask = encoding["attention_mask"].to(DEVICE)
with torch.no_grad():
logits = model(input_ids, attention_mask)
probs = F.softmax(logits, dim=-1)
return probs[:, 1].cpu().tolist()
# ── Singleton β€” modelo, calibrador y config se cargan una sola vez ─────────────
_model = None
_tokenizer = None
_calibrator = None
_threshold_cfg: dict = {}
def load_model():
"""Carga el modelo, calibrador y configuraciΓ³n de umbral (singleton)."""
global _model, _tokenizer, _calibrator, _threshold_cfg
global MODEL_PATH, CALIBRATOR_PATH, THRESHOLD_PATH
if _model is not None:
return _model, _tokenizer, _calibrator, _threshold_cfg
# ── Resolver rutas (local o Hugging Face) ──────────────────────────────────
MODEL_PATH = _resolve_file("best_model.pt")
CALIBRATOR_PATH = _resolve_file("isotonic_calibrator.pkl")
THRESHOLD_PATH = _resolve_file("threshold_config.json")
print(f"[model_loader] Cargando encoder base '{MODEL_NAME}'...")
encoder = SentenceTransformer(MODEL_NAME)
_tokenizer = encoder.tokenizer
print(f"[model_loader] Cargando pesos desde '{MODEL_PATH}'...")
checkpoint = torch.load(MODEL_PATH, map_location=DEVICE, weights_only=False)
# El checkpoint guarda el estado completo del modelo (encoder + classifier).
# Las claves tienen el prefijo "encoder.0.model.*" porque SentenceTransformer
# almacena el transformer como encoder[0].auto_model. Construimos el modelo
# y cargamos el state_dict completo sobre Γ©l.
cfg_saved = checkpoint.get("config", {})
hidden_dim = cfg_saved.get("HIDDEN_DIM", 384)
dropout = cfg_saved.get("DROPOUT", 0.20)
_model = StyleAIClassifierV5(
encoder_model=encoder,
hidden_dim=hidden_dim,
dropout=dropout,
).float().to(DEVICE)
# Remapeo de claves: versiones antiguas de sentence-transformers guardaban los
# pesos del transformer bajo "encoder.0.model.*"; versiones nuevas los exponen
# como "encoder.0.auto_model.*". Normalizamos antes de cargar.
raw_sd = checkpoint["model_state_dict"]
remapped_sd = {}
for k, v in raw_sd.items():
new_k = k.replace("encoder.0.model.", "encoder.0.auto_model.")
remapped_sd[new_k] = v
missing, unexpected = _model.load_state_dict(remapped_sd, strict=False)
if missing or unexpected:
raise RuntimeError(
f"[model_loader] State dict no coincide con la arquitectura "
f"(posible mismatch de versiΓ³n de sentence-transformers/torch): "
f"{len(missing)} claves faltantes {missing[:5]}, "
f"{len(unexpected)} claves inesperadas {unexpected[:5]}"
)
print("[model_loader] State dict cargado sin discrepancias.")
_model.eval()
print(f"[model_loader] Cargando calibrador desde '{CALIBRATOR_PATH}'...")
with open(CALIBRATOR_PATH, "rb") as f:
_calibrator = pickle.load(f)
print(f"[model_loader] Cargando configuracion de umbral desde '{THRESHOLD_PATH}'...")
with open(THRESHOLD_PATH, "r") as f:
_threshold_cfg = json.load(f)
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
print("[model_loader] Pipeline listo.")
print(f"[model_loader] Dispositivo: {DEVICE}")
return _model, _tokenizer, _calibrator, _threshold_cfg
# ── Inferencia principal ───────────────────────────────────────────────────────
def predict(text: str) -> dict:
"""
Ejecuta el pipeline completo: preproceso β†’ TTA β†’ calibraciΓ³n β†’ abstenciΓ³n.
Retorna:
{
"label": "IA" | "Humano" | "Incierto",
"confidence": float (0–100),
"ai_prob": float (0–100),
"human_prob": float (0–100),
"word_count": int,
"abstained": bool
}
"""
model, tokenizer, calibrator, cfg = load_model()
# ── 1. Preprocesamiento ────────────────────────────────────────────────────
cleaned = _preprocess(text)
# ── 2. TTA: pesos [0.5, 0.25, 0.25] ──────────────────────────────────────
tta_texts = [cleaned, _tta_truncate(cleaned), _tta_span_mask(cleaned)]
tta_weights = cfg.get("tta_weights", [0.5, 0.25, 0.25])
# Batch ΓΊnico con los 3 pases TTA para eficiencia
raw_probs = _encode_and_predict(model, tta_texts, tokenizer, MAX_LENGTH)
# Promedio ponderado de las probabilidades AI de los tres pases
ai_raw = sum(w * p for w, p in zip(tta_weights, raw_probs))
# ── 3. CalibraciΓ³n isotΓ³nica ───────────────────────────────────────────────
# IsotonicRegression.predict() espera array 1D de shape (n_samples,)
ai_calibrated = float(calibrator.predict(np.array([ai_raw]))[0])
# ── 4. AbstenciΓ³n ──────────────────────────────────────────────────────────
abstain_low = cfg.get("abstain_low", 0.44)
abstain_high = cfg.get("abstain_high", 0.56)
threshold = cfg.get("threshold", 0.50)
abstained = abstain_low <= ai_calibrated <= abstain_high
if abstained:
label = "Incierto"
confidence = round((1 - abs(ai_calibrated - 0.5) * 2) * 100, 2)
elif ai_calibrated > threshold:
label = "IA"
confidence = round(ai_calibrated * 100, 2)
else:
label = "Humano"
confidence = round((1 - ai_calibrated) * 100, 2)
return {
"label": label,
"confidence": confidence,
"ai_prob": round(ai_calibrated * 100, 2),
"human_prob": round((1 - ai_calibrated) * 100, 2),
"word_count": len(cleaned.split()),
"abstained": abstained,
}