File size: 19,049 Bytes
ad4c906 |
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 |
"""
=== MIA · Script de Entrenamiento del AgentEmotionPredictClassifier (v2) ===
- Lee JSON v2 con `label` (usuario) y `label_agent` (objetivo del agente).
- Soporta encoder preentrenado (BETO) con fine-tuning controlado.
- Class weights + label smoothing para desbalance.
- Early stopping por macro-F1.
- AdamW con dos grupos de LR (encoder vs head) + scheduler lineal con warmup.
- Clip de gradientes, dump de misclasificados, matrices de confusión (absoluta y normalizada).
Requisitos:
pip install transformers scikit-learn seaborn matplotlib tqdm
"""
import json
from pathlib import Path
from typing import Dict, List, Tuple, Optional
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from sklearn.metrics import classification_report, confusion_matrix, f1_score
from tqdm import tqdm
import matplotlib.pyplot as plt
import seaborn as sns
from agent_emotion_predict_classifier import AgentEmotionPredictClassifier
# ==================== DATASET ====================
class AgentEmotionDataset(Dataset):
def __init__(self, data_path: str, label_map: Optional[dict] = None):
with open(data_path, 'r', encoding='utf-8') as f:
raw = json.load(f)
data = raw.get('data', raw)
self.texts: List[str] = []
self.user_labels: List[int] = []
self.agent_labels: List[int] = []
for it in data:
self.texts.append(it['text'])
u = it['label']; a = it['label_agent']
u = int(u) if isinstance(u, str) else u
a = int(a) if isinstance(a, str) else a
if label_map is not None:
a = label_map[a] # remapea a {0..K-1}
self.user_labels.append(u)
self.agent_labels.append(a)
def __len__(self):
return len(self.texts)
def __getitem__(self, idx):
return self.texts[idx], self.user_labels[idx], self.agent_labels[idx]
def collate_fn(batch):
texts, ulabels, alabels = zip(*batch)
ulabels = torch.tensor(ulabels, dtype=torch.long)
alabels = torch.tensor(alabels, dtype=torch.long)
return list(texts), ulabels, alabels
# ==================== TRAINER ====================
class AgentEmotionTrainer:
def __init__(
self,
model: AgentEmotionPredictClassifier,
train_loader: DataLoader,
val_loader: DataLoader,
test_loader: DataLoader,
device: Optional[torch.device] = None,
lr_encoder: float = 2e-5,
lr_head: float = 1e-3,
weight_decay: float = 0.01,
num_epochs: int = 20,
warmup_ratio: float = 0.1,
early_stopping_patience: int = 3,
warmup_freeze_epochs: int = 2,
num_classes: int = 2,
class_names: Optional[List[str]] = None,
):
self.model = model
self.train_loader = train_loader
self.val_loader = val_loader
self.test_loader = test_loader
self.device = device or torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.num_classes = num_classes
self.class_names = class_names or ["alegría","amor"]
# ---------- Pérdida con pesos de clase + label smoothing ----------
labels_tensor = torch.tensor(train_loader.dataset.agent_labels)
class_counts = torch.bincount(labels_tensor, minlength=self.num_classes).float()
safe_counts = class_counts.clamp(min=1.0)
inv_freq = (safe_counts.sum() / (self.num_classes * safe_counts)).to(self.device)
class_weights = inv_freq / inv_freq.mean()
self.criterion = nn.CrossEntropyLoss(weight=class_weights, label_smoothing=0.05)
# ---------- Optimizador AdamW con 2 grupos (encoder vs head) ----------
self.lr_encoder = lr_encoder
self.lr_head = lr_head
self.weight_decay = weight_decay
# Warmup: encoder congelado n épocas
self.warmup_freeze_epochs = warmup_freeze_epochs
self.model.freeze_encoder()
self.optimizer = self._build_optimizer()
# ---------- Scheduler lineal con warmup ----------
self.num_epochs = num_epochs
total_steps = len(self.train_loader) * self.num_epochs
from transformers import get_linear_schedule_with_warmup
self.scheduler = get_linear_schedule_with_warmup(
self.optimizer,
num_warmup_steps=int(warmup_ratio * total_steps),
num_training_steps=total_steps,
)
# ---------- Tracking ----------
self.train_losses, self.val_losses = [], []
self.train_accs, self.val_accs = [], []
self.val_f1s = []
self.best_val_f1 = 0.0
self.best_state = None
self.early_stopping_patience = early_stopping_patience
def _build_optimizer(self):
encoder_params, head_params = [], []
for n, p in self.model.named_parameters():
if not p.requires_grad:
continue
if "embedder.encoder" in n: # parámetros del Transformer
encoder_params.append(p)
else:
head_params.append(p)
return torch.optim.AdamW(
[
{"params": encoder_params, "lr": self.lr_encoder, "weight_decay": self.weight_decay},
{"params": head_params, "lr": self.lr_head, "weight_decay": 0.0},
]
)
def train_epoch(self) -> Tuple[float, float]:
self.model.train()
total_loss, correct, total = 0.0, 0, 0
pbar = tqdm(self.train_loader, desc="Training")
for texts, ulabels, alabels in pbar:
alabels = alabels.to(self.device)
self.optimizer.zero_grad()
logits = self.model(texts, ulabels.to(self.device))
loss = self.criterion(logits, alabels)
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
self.optimizer.step()
if self.scheduler is not None:
self.scheduler.step()
total_loss += loss.item()
preds = torch.argmax(logits, dim=-1)
correct += (preds == alabels).sum().item()
total += alabels.size(0)
pbar.set_postfix({'loss': f'{loss.item():.4f}', 'acc': f'{100*correct/total:.2f}%'})
avg_loss = total_loss / len(self.train_loader)
accuracy = 100 * correct / total
return avg_loss, accuracy
def validate(self) -> Tuple[float, float, float]:
self.model.eval()
total_loss, correct, total = 0.0, 0, 0
all_preds, all_labels = [], []
with torch.no_grad():
for texts, ulabels, alabels in tqdm(self.val_loader, desc="Validation"):
alabels = alabels.to(self.device)
logits = self.model(texts, ulabels.to(self.device))
loss = self.criterion(logits, alabels)
total_loss += loss.item()
preds = torch.argmax(logits, dim=-1)
correct += (preds == alabels).sum().item()
total += alabels.size(0)
all_preds.extend(preds.cpu().tolist())
all_labels.extend(alabels.cpu().tolist())
avg_loss = total_loss / len(self.val_loader)
accuracy = 100 * correct / total
macro_f1 = f1_score(all_labels, all_preds, average='macro')
return avg_loss, accuracy, macro_f1
def test(self, save_dir: Path) -> Dict:
self.model.eval()
all_predictions, all_labels, all_texts = [], [], []
with torch.no_grad():
for texts, ulabels, alabels in tqdm(self.test_loader, desc="Testing"):
logits = self.model(texts, ulabels.to(self.device))
predictions = torch.argmax(logits, dim=-1)
all_predictions.extend(predictions.cpu().numpy())
all_labels.extend(alabels.numpy())
all_texts.extend(texts)
accuracy = 100 * np.mean(np.array(all_predictions) == np.array(all_labels))
labels_ord = list(range(self.num_classes))
target_names = self.class_names
report = classification_report(
all_labels,
all_predictions,
labels=labels_ord,
target_names=target_names,
output_dict=True,
zero_division=0
)
cm_abs = confusion_matrix(all_labels, all_predictions, labels=labels_ord)
cm_norm = confusion_matrix(all_labels, all_predictions, labels=labels_ord, normalize='true')
self._dump_misclassified(all_texts, all_labels, all_predictions, target_names, save_dir / "misclassified.txt")
return {
'accuracy': accuracy,
'classification_report': report,
'confusion_matrix_abs': cm_abs,
'confusion_matrix_norm': cm_norm,
}
@staticmethod
def _dump_misclassified(texts, y_true, y_pred, names, path: Path):
with open(path, "w", encoding="utf-8") as f:
for t, yt, yp in zip(texts, y_true, y_pred):
if yt != yp:
f.write(f"[gold={names[yt]} | pred={names[yp]}] {t}\n")
def train(self, num_epochs: Optional[int] = None, early_stopping_patience: Optional[int] = None, save_dir: str = "models/agent_emotion"):
num_epochs = num_epochs or self.num_epochs
if early_stopping_patience is not None:
self.early_stopping_patience = early_stopping_patience
print(f"\n{'='*60}\nIniciando entrenamiento por {num_epochs} épocas\nDevice: {self.device}\n{'='*60}")
patience_counter = 0
best_f1 = -1.0
for epoch in range(1, num_epochs + 1):
# Unfreeze encoder después del warmup
if epoch == self.warmup_freeze_epochs + 1:
print("→ Descongelando encoder para fine-tuning...")
self.model.unfreeze_encoder()
self.optimizer = self._build_optimizer()
print(f"\nÉpoca {epoch}/{num_epochs}\n" + '-'*60)
tr_loss, tr_acc = self.train_epoch()
self.train_losses.append(tr_loss); self.train_accs.append(tr_acc)
va_loss, va_acc, va_f1 = self.validate()
self.val_losses.append(va_loss); self.val_accs.append(va_acc); self.val_f1s.append(va_f1)
print(f"Train: loss={tr_loss:.4f} acc={tr_acc:.2f}% | Val: loss={va_loss:.4f} acc={va_acc:.2f}% f1m={va_f1:.4f}")
# Guardar mejor por Macro-F1
if va_f1 > best_f1 and best_f1 < 0.9999:
best_f1 = va_f1
patience_counter = 0
Path(save_dir).mkdir(parents=True, exist_ok=True)
torch.save({
'epoch': epoch,
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'train_loss': tr_loss,
'val_loss': va_loss,
'val_acc': va_acc,
'val_f1': va_f1,
}, f"{save_dir}/best_model.pt")
print(f" ✓ Mejor modelo guardado (Val Macro-F1: {va_f1:.4f})")
else:
patience_counter += 1
print(f" No improvement. Patience: {patience_counter}/{self.early_stopping_patience}")
if patience_counter >= self.early_stopping_patience:
print(f"\nEarly stopping activado en época {epoch}")
break
print(f"\n{'='*60}\nEntrenamiento completado! Mejor Val Macro-F1: {best_f1:.4f}\n{'='*60}")
# Evaluar en test con mejor checkpoint
ckpt = torch.load(f"{save_dir}/best_model.pt", map_location=self.device)
self.model.load_state_dict(ckpt['model_state_dict'])
test_results = self.test(Path(save_dir))
print(f"\n{'='*60}\nRESULTADOS EN TEST SET\n{'='*60}")
print(f"Test Accuracy: {test_results['accuracy']:.2f}%\n")
# Graficar y guardar reportes
self.plot_training_history(save_dir)
self.plot_confusion_matrix(test_results['confusion_matrix_abs'], save_dir, norm=False)
self.plot_confusion_matrix(test_results['confusion_matrix_norm'], save_dir, norm=True)
self.save_classification_report(test_results['classification_report'], save_dir)
return test_results
# ---------- utilidades de guardado/plot ----------
def plot_training_history(self, save_dir: str):
plt.figure(figsize=(12, 5))
plt.subplot(1, 2, 1)
plt.plot(self.train_losses, label='Train Loss', marker='o')
plt.plot(self.val_losses, label='Val Loss', marker='s')
plt.xlabel('Época'); plt.ylabel('Loss'); plt.title('Loss'); plt.legend(); plt.grid(True, alpha=0.3)
plt.subplot(1, 2, 2)
plt.plot(self.train_accs, label='Train Acc', marker='o')
plt.plot(self.val_accs, label='Val Acc', marker='s')
plt.plot(self.val_f1s, label='Val Macro-F1', marker='^')
plt.xlabel('Época'); plt.ylabel('Score'); plt.title('Acc / Macro-F1'); plt.legend(); plt.grid(True, alpha=0.3)
plt.tight_layout(); Path(save_dir).mkdir(parents=True, exist_ok=True)
plt.savefig(f"{save_dir}/training_history.png", dpi=300, bbox_inches='tight'); plt.close()
print(f"✓ Gráfica de entrenamiento guardada en: {save_dir}/training_history.png")
def plot_confusion_matrix(self, cm: np.ndarray, save_dir: str, norm: bool = False):
plt.figure(figsize=(8, 6))
fmt = '.2f' if norm else 'd'
cmap = 'Blues'
ticklabels = self.class_names
sns.heatmap(cm, annot=True, fmt=fmt, cmap=cmap, xticklabels=ticklabels, yticklabels=ticklabels,
vmin=0, vmax=1 if norm else None)
plt.title('Matriz de Confusión ' + ('(Normalizada)' if norm else '(Absoluta)'))
plt.ylabel('Etiqueta Real'); plt.xlabel('Etiqueta Predicha')
plt.tight_layout(); Path(save_dir).mkdir(parents=True, exist_ok=True)
fname = "confusion_matrix_norm.png" if norm else "confusion_matrix_abs.png"
plt.savefig(f"{save_dir}/{fname}", dpi=300, bbox_inches='tight'); plt.close()
print(f"✓ Matriz de confusión guardada en: {save_dir}/{fname}")
def save_classification_report(self, report: Dict, save_dir: str):
Path(save_dir).mkdir(parents=True, exist_ok=True)
with open(f"{save_dir}/classification_report.txt", 'w', encoding='utf-8') as f:
f.write("="*60 + "\n")
f.write("REPORTE DE CLASIFICACIÓN - TEST SET\n")
f.write("="*60 + "\n\n")
for emotion, metrics in report.items():
if emotion in ['accuracy', 'macro avg', 'weighted avg']:
continue
f.write(f"\n{emotion.upper()}:\n")
f.write(f" Precision: {metrics['precision']:.4f}\n")
f.write(f" Recall: {metrics['recall']:.4f}\n")
f.write(f" F1-Score: {metrics['f1-score']:.4f}\n")
f.write(f" Support: {metrics['support']}\n")
f.write(f"\n{'='*60}\n")
f.write(f"MACRO AVG:\n Precision: {report['macro avg']['precision']:.4f}\n Recall: {report['macro avg']['recall']:.4f}\n F1-Score: {report['macro avg']['f1-score']:.4f}\n")
f.write(f"\nWEIGHTED AVG:\n Precision: {report['weighted avg']['precision']:.4f}\n Recall: {report['weighted avg']['recall']:.4f}\n F1-Score: {report['weighted avg']['f1-score']:.4f}\n")
acc = report.get('accuracy', None)
if acc is not None:
f.write(f"\nACCURACY: {acc:.4f}\n")
f.write("="*60 + "\n")
print(f"✓ Reporte de clasificación guardado en: {save_dir}/classification_report.txt")
# ==================== MAIN ====================
def main():
# Rutas (usa los *v2.json*)
DATA_DIR = "models/emotion_classifier/data"
TRAIN_PATH = f"{DATA_DIR}/emotion_dataset_train_es_v2.json"
VAL_PATH = f"{DATA_DIR}/emotion_dataset_validation_es_v2.json"
TEST_PATH = f"{DATA_DIR}/emotion_dataset_test_es_v2.json"
SAVE_DIR = "models/agent_emotion"
# Hiperparámetros
BATCH_SIZE = 16
NUM_EPOCHS = 20
EARLY_STOPPING_PATIENCE = 2
WARMUP_FREEZE_EPOCHS = 2 #antes 5
LR_ENCODER = 1e-5
LR_HEAD = 5e-4
WEIGHT_DECAY = 0.05
WARMUP_RATIO = 0.1
print("="*60)
print("CONFIGURACIÓN DEL ENTRENAMIENTO (AgentEmotion v2)")
print("="*60)
print(f"Train: {TRAIN_PATH}")
print(f"Val: {VAL_PATH}")
print(f"Test: {TEST_PATH}")
print(f"Batch size: {BATCH_SIZE} | Épocas: {NUM_EPOCHS}")
print(f"Freeze warmup epochs: {WARMUP_FREEZE_EPOCHS}")
print(f"LR encoder: {LR_ENCODER} | LR head: {LR_HEAD}")
# 1) Detectar clases presentes en TRAIN y remapear a {0..K-1}
probe = AgentEmotionDataset(TRAIN_PATH)
present_classes = sorted(set(probe.agent_labels)) # p.ej., [1,2]
label_names_full = {0:"tristeza",1:"alegría",2:"amor",3:"ira",4:"miedo",5:"sorpresa"}
class_names = [label_names_full[c] for c in present_classes]
K = len(present_classes)
print(f"Clases existentes - valor K : {K}")
label_map = {orig:i for i, orig in enumerate(present_classes)}
print(f"Clases de agente en train: {present_classes} → K={K} ({class_names})")
# 2) Recrear datasets aplicando el mapeo
train_ds = AgentEmotionDataset(TRAIN_PATH, label_map=label_map)
val_ds = AgentEmotionDataset(VAL_PATH, label_map=label_map)
test_ds = AgentEmotionDataset(TEST_PATH, label_map=label_map)
# 3) DataLoaders
train_loader = DataLoader(train_ds, batch_size=BATCH_SIZE, shuffle=True, collate_fn=collate_fn)
val_loader = DataLoader(val_ds, batch_size=BATCH_SIZE, shuffle=True, collate_fn=collate_fn)
test_loader = DataLoader(test_ds, batch_size=BATCH_SIZE, shuffle=True, collate_fn=collate_fn)
# 4) Modelo (salida con K clases)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = AgentEmotionPredictClassifier(
pretrained_encoder="beto",
max_length=128,
hidden1=256,
hidden2=64,
dropout=0.4, # antes 0.2
label_feature_dropout=0.5, # antes 0.15
device=device,
num_classes=K,
)
model.freeze_encoder()
# 5) Trainer con num_classes y nombres
trainer = AgentEmotionTrainer(
model=model,
train_loader=train_loader,
val_loader=val_loader,
test_loader=test_loader,
device=device,
lr_encoder=LR_ENCODER,
lr_head=LR_HEAD,
weight_decay=WEIGHT_DECAY,
num_epochs=NUM_EPOCHS,
warmup_ratio=WARMUP_RATIO,
early_stopping_patience=EARLY_STOPPING_PATIENCE,
warmup_freeze_epochs=WARMUP_FREEZE_EPOCHS,
num_classes=K,
class_names=class_names,
)
# 6) Entrenar y evaluar
trainer.train(
num_epochs=NUM_EPOCHS,
early_stopping_patience=EARLY_STOPPING_PATIENCE,
save_dir=SAVE_DIR,
)
if __name__ == "__main__":
main()
|