scripts/train_lora_emotion2vec.py

#!/usr/bin/env python3
"""Manual LoRA fine-tuning for emotion2vec_plus_base — 7-class emotion.

Wraps frozen attention layers with low-rank adapters (LoRALinear),
replaces the 9-class proj head with a 7-class MLPHead, and trains
with FocalLoss + disgust F1 gating.

Usage:
    # Quick smoke test (CPU)
    python scripts/train_lora_emotion2vec.py \
        --train-manifest data/lora_7class/train_manifest.json \
        --val-manifest data/lora_7class/val_manifest.json \
        --output-dir data/models/lora_emotion2vec_7class \
        --epochs 3 --device cpu

    # Full training (GPU, RTX 5050 8GB)
    PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
    python scripts/train_lora_emotion2vec.py \
        --train-manifest data/lora_7class/train_manifest.json \
        --val-manifest data/lora_7class/val_manifest.json \
        --output-dir data/models/lora_emotion2vec_7class \
        --epochs 20 --batch-size 4 --accumulate-steps 8 --device cuda
"""

from __future__ import annotations

import argparse
import json
import logging
import random
import sys
import time
from pathlib import Path

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
logger = logging.getLogger(__name__)

# 7-class label taxonomy (matches prepare_lora_dataset.py)
LABELS_7CLASS = ["happiness", "anger", "disgust", "fear", "neutral", "sadness", "surprise"]
LABEL2IDX = {label: i for i, label in enumerate(LABELS_7CLASS)}
NUM_CLASSES = len(LABELS_7CLASS)
DISGUST_IDX = LABEL2IDX["disgust"]  # 2


# ──────────────────────────────────────────────
# LoRA Components
# ──────────────────────────────────────────────

class LoRALinear(nn.Module):
    """Low-rank adapter wrapping a frozen nn.Linear.

    At init, B is zero so LoRA contribution is zero (original behavior preserved).
    scaling = alpha / r controls the magnitude of the LoRA update.
    """

    def __init__(self, original: nn.Linear, r: int = 16, alpha: int = 32, dropout: float = 0.1):
        super().__init__()
        self.original = original
        self.r = r
        self.scaling = alpha / r

        # Freeze original weights
        self.original.weight.requires_grad = False
        if self.original.bias is not None:
            self.original.bias.requires_grad = False

        in_features = original.in_features
        out_features = original.out_features

        # Low-rank matrices
        self.lora_A = nn.Linear(in_features, r, bias=False)
        self.lora_B = nn.Linear(r, out_features, bias=False)
        self.dropout = nn.Dropout(dropout)

        # Init: A = kaiming, B = zero (so initial LoRA output = 0)
        nn.init.kaiming_uniform_(self.lora_A.weight)
        nn.init.zeros_(self.lora_B.weight)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        base_out = self.original(x)
        lora_out = self.lora_B(self.dropout(self.lora_A(x))) * self.scaling
        return base_out + lora_out


def merge_lora_linear(lora: LoRALinear) -> nn.Linear:
    """Merge LoRA weights into a plain nn.Linear for inference.

    W_merged = W_original + scaling * B.weight @ A.weight
    """
    with torch.no_grad():
        merged_weight = (
            lora.original.weight
            + lora.scaling * lora.lora_B.weight @ lora.lora_A.weight
        )
        bias = lora.original.bias

    merged = nn.Linear(
        lora.original.in_features,
        lora.original.out_features,
        bias=bias is not None,
    )
    merged.weight = nn.Parameter(merged_weight)
    if bias is not None:
        merged.bias = nn.Parameter(bias.clone())
    return merged


def inject_lora(
    encoder: nn.Module,
    r: int = 16,
    alpha: int = 32,
    dropout: float = 0.1,
) -> None:
    """Replace attn.qkv and attn.proj in each block with LoRALinear.

    emotion2vec uses FUSED QKV: attn.qkv: Linear(768, 2304)
    and attn.proj: Linear(768, 768). 8 blocks total.
    MLP layers are NOT wrapped.
    """
    for block in encoder.blocks:
        block.attn.qkv = LoRALinear(block.attn.qkv, r=r, alpha=alpha, dropout=dropout)
        block.attn.proj = LoRALinear(block.attn.proj, r=r, alpha=alpha, dropout=dropout)


# ──────────────────────────────────────────────
# Model Components
# ──────────────────────────────────────────────

class MLPHead(nn.Module):
    """Multi-layer classification head: 768 → 512 → 256 → num_classes."""

    def __init__(self, in_dim: int = 768, num_classes: int = NUM_CLASSES, dropout: float = 0.3):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(in_dim, 512),
            nn.BatchNorm1d(512),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(512, 256),
            nn.BatchNorm1d(256),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(256, num_classes),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.net(x)


class FocalLoss(nn.Module):
    """Focal Loss with optional label smoothing and class weights."""

    def __init__(self, weight=None, gamma: float = 2.0, label_smoothing: float = 0.05):
        super().__init__()
        self.gamma = gamma
        self.weight = weight
        self.label_smoothing = label_smoothing

    def forward(self, logits: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
        ce_loss = F.cross_entropy(
            logits, targets, weight=self.weight,
            label_smoothing=self.label_smoothing, reduction="none",
        )
        pt = torch.exp(-ce_loss)
        focal_loss = ((1 - pt) ** self.gamma) * ce_loss
        return focal_loss.mean()


# ──────────────────────────────────────────────
# Dataset
# ──────────────────────────────────────────────

class EmotionDataset(Dataset):
    """Load audio from unified manifest JSON for 7-class LoRA training.

    Manifest format: list of {"audio_path": str, "label": str, ...}
    OR {"path": str, "label": str, ...} (for backward compat).
    """

    def __init__(
        self,
        manifest_path: str,
        max_duration_sec: float = 8.0,
        phone_augment_prob: float = 0.0,
        noise_augment_prob: float = 0.0,
    ):
        import torchaudio  # lazy import

        with open(manifest_path, encoding="utf-8") as f:
            self.samples = json.load(f)

        self.max_samples = int(max_duration_sec * 16000)
        self.phone_augment_prob = phone_augment_prob
        self.noise_augment_prob = noise_augment_prob
        self._phone_sim = None

    def _get_phone_sim(self):
        if self._phone_sim is None:
            sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "src"))
            from common.phone_simulator import PhoneSimulator, CompandingType
            self._phone_sim = PhoneSimulator(companding=CompandingType.ALAW)
        return self._phone_sim

    def __len__(self):
        return len(self.samples)

    def __getitem__(self, idx):
        import torchaudio

        sample = self.samples[idx]
        # Support both "audio_path" and "path" keys
        audio_path = sample.get("audio_path") or sample.get("path", "")

        waveform, sr = torchaudio.load(audio_path)
        # Mono
        if waveform.shape[0] > 1:
            waveform = waveform.mean(dim=0, keepdim=True)
        waveform = waveform.squeeze(0)  # (T,)

        # Resample to 16kHz
        if sr != 16000:
            waveform = torchaudio.functional.resample(waveform, sr, 16000)

        # Truncate
        if waveform.shape[0] > self.max_samples:
            waveform = waveform[:self.max_samples]

        audio = waveform.numpy()

        # Augmentation
        r = random.random()
        if r < self.phone_augment_prob:
            sim = self._get_phone_sim()
            audio, _ = sim.process(audio, 16000)
            import librosa
            audio = librosa.resample(audio, orig_sr=8000, target_sr=16000)
        elif r < self.phone_augment_prob + self.noise_augment_prob:
            snr_db = random.uniform(10, 20)
            signal_power = np.mean(audio ** 2)
            noise_power = signal_power / (10 ** (snr_db / 10))
            noise = np.random.normal(0, np.sqrt(max(noise_power, 1e-10)), len(audio)).astype(np.float32)
            audio = audio + noise

        label = LABEL2IDX[sample["label"]]
        return torch.tensor(audio, dtype=torch.float32), label


def collate_fn(batch):
    """Pad waveforms to same length in batch."""
    waveforms, labels = zip(*batch)
    max_len = max(w.shape[0] for w in waveforms)
    padded = torch.zeros(len(waveforms), max_len)
    for i, w in enumerate(waveforms):
        padded[i, :w.shape[0]] = w
    return padded, torch.tensor(labels, dtype=torch.long)


# ──────────────────────────────────────────────
# Model Loading & Forward
# ──────────────────────────────────────────────

def load_model(device: str, r: int = 16, alpha: int = 32, dropout: float = 0.1):
    """Load emotion2vec_plus_base, freeze all, inject LoRA, replace proj."""
    from funasr import AutoModel

    logger.info("Loading emotion2vec_plus_base...")
    fmodel = AutoModel(model="iic/emotion2vec_plus_base", device=device, hub="hf")
    encoder = fmodel.model

    # Freeze everything
    for param in encoder.parameters():
        param.requires_grad = False

    # Inject LoRA adapters into attention layers
    inject_lora(encoder, r=r, alpha=alpha, dropout=dropout)

    # Replace 9-class proj with 7-class MLPHead
    old_proj = encoder.proj
    encoder.proj = MLPHead(768, NUM_CLASSES)
    logger.info("Replaced proj: Linear(768, %d) -> MLPHead(768->512->256->%d)",
                old_proj.out_features, NUM_CLASSES)

    # Move entire model (including new LoRA params + MLPHead) to device
    encoder = encoder.to(device)
    return encoder


def forward_pass(encoder, waveforms: torch.Tensor, device: str) -> torch.Tensor:
    """Differentiable forward pass through emotion2vec with LoRA.

    Args:
        encoder: emotion2vec model with LoRA injected
        waveforms: (B, T) float32, 16kHz
        device: compute device

    Returns:
        logits: (B, 7)
    """
    waveforms = waveforms.to(device)

    # Layer norm (per emotion2vec inference)
    if encoder.cfg.normalize:
        normed = []
        for i in range(waveforms.shape[0]):
            normed.append(F.layer_norm(waveforms[i], waveforms[i].shape))
        waveforms = torch.stack(normed)

    # Extract features
    feats = encoder.extract_features(waveforms, padding_mask=None)
    x = feats["x"]  # (B, T', 768)

    # Mean pool + classify
    pooled = x.mean(dim=1)  # (B, 768)
    logits = encoder.proj(pooled)  # (B, 7)
    return logits


# ──────────────────────────────────────────────
# Validation
# ──────────────────────────────────────────────

@torch.no_grad()
def validate(encoder, val_loader, device, criterion):
    """Run validation, return metrics including disgust_f1."""
    encoder.eval()
    total_loss = 0
    y_true, y_pred = [], []

    for waveforms, labels in val_loader:
        labels = labels.to(device)
        logits = forward_pass(encoder, waveforms, device)
        loss = criterion(logits, labels)
        total_loss += loss.item() * labels.size(0)

        preds = logits.argmax(dim=-1)
        y_true.extend(labels.cpu().tolist())
        y_pred.extend(preds.cpu().tolist())

    from sklearn.metrics import precision_recall_fscore_support, accuracy_score, confusion_matrix

    accuracy = accuracy_score(y_true, y_pred)
    _, _, f1_per_class, _ = precision_recall_fscore_support(
        y_true, y_pred, labels=list(range(NUM_CLASSES)), average=None, zero_division=0,
    )
    macro_f1 = float(np.mean(f1_per_class))
    disgust_f1 = float(f1_per_class[DISGUST_IDX])

    per_class = {LABELS_7CLASS[i]: round(float(f1_per_class[i]), 4) for i in range(NUM_CLASSES)}
    avg_loss = total_loss / max(len(y_true), 1)
    cm = confusion_matrix(y_true, y_pred, labels=list(range(NUM_CLASSES)))

    return {
        "loss": round(avg_loss, 4),
        "accuracy": round(accuracy, 4),
        "macro_f1": round(macro_f1, 4),
        "disgust_f1": round(disgust_f1, 4),
        "per_class_f1": per_class,
        "confusion_matrix": cm.tolist(),
        "y_true": y_true,
        "y_pred": y_pred,
    }


# ──────────────────────────────────────────────
# Checkpoint
# ──────────────────────────────────────────────

def save_lora_checkpoint(encoder, path: Path, epoch: int, metrics: dict,
                         best_f1: float = 0.0, patience_counter: int = 0,
                         optimizer=None, scheduler=None, scaler=None,
                         training_log=None):
    """Save LoRA weights + MLPHead + optimizer state for resume."""
    lora_state = {}
    for name, module in encoder.named_modules():
        if isinstance(module, LoRALinear):
            lora_state[f"{name}.lora_A.weight"] = module.lora_A.weight.data.cpu()
            lora_state[f"{name}.lora_B.weight"] = module.lora_B.weight.data.cpu()

    state = {
        "lora_weights": lora_state,
        "proj": encoder.proj.state_dict(),
        "epoch": epoch,
        "metrics": metrics,
        "best_f1": best_f1,
        "patience_counter": patience_counter,
        "num_classes": NUM_CLASSES,
        "labels": LABELS_7CLASS,
    }
    if optimizer is not None:
        state["optimizer"] = optimizer.state_dict()
    if scheduler is not None:
        state["scheduler"] = scheduler.state_dict()
    if scaler is not None:
        state["scaler"] = scaler.state_dict()
    if training_log is not None:
        state["training_log"] = training_log
    torch.save(state, str(path))


def load_lora_checkpoint(encoder, path: Path, device: str,
                         optimizer=None, scheduler=None, scaler=None):
    """Load LoRA checkpoint and restore training state."""
    logger.info("Resuming from checkpoint: %s", path)
    ckpt = torch.load(str(path), map_location=device, weights_only=False)

    # Restore LoRA weights
    for name, module in encoder.named_modules():
        if isinstance(module, LoRALinear):
            a_key = f"{name}.lora_A.weight"
            b_key = f"{name}.lora_B.weight"
            if a_key in ckpt["lora_weights"]:
                module.lora_A.weight.data.copy_(ckpt["lora_weights"][a_key].to(device))
                module.lora_B.weight.data.copy_(ckpt["lora_weights"][b_key].to(device))

    # Restore MLPHead
    encoder.proj.load_state_dict(ckpt["proj"])

    # Restore optimizer/scheduler/scaler if available
    if optimizer is not None and "optimizer" in ckpt:
        optimizer.load_state_dict(ckpt["optimizer"])
    if scheduler is not None and "scheduler" in ckpt:
        scheduler.load_state_dict(ckpt["scheduler"])
    if scaler is not None and "scaler" in ckpt:
        scaler.load_state_dict(ckpt["scaler"])

    logger.info("Resumed from epoch %d (best_f1=%.4f, patience=%d)",
                ckpt["epoch"], ckpt["best_f1"], ckpt["patience_counter"])
    return {
        "epoch": ckpt["epoch"],
        "best_f1": ckpt["best_f1"],
        "patience_counter": ckpt["patience_counter"],
        "training_log": ckpt.get("training_log", []),
    }


# ──────────────────────────────────────────────
# Confusion Matrix Plot
# ──────────────────────────────────────────────

def plot_confusion_matrix(cm, output_path: Path, epoch: int):
    """Save confusion matrix as PNG."""
    try:
        import matplotlib
        matplotlib.use("Agg")
        import matplotlib.pyplot as plt
        import seaborn as sns

        fig, ax = plt.subplots(figsize=(9, 7))
        cm_norm = cm / np.maximum(cm.sum(axis=1, keepdims=True), 1)
        sns.heatmap(
            cm_norm, annot=True, fmt=".2f", cmap="Blues",
            xticklabels=LABELS_7CLASS, yticklabels=LABELS_7CLASS, ax=ax,
        )
        for i in range(NUM_CLASSES):
            for j in range(NUM_CLASSES):
                ax.text(j + 0.5, i + 0.7, f"({cm[i][j]})",
                        ha="center", va="center", fontsize=6, color="gray")

        ax.set_xlabel("Predicted")
        ax.set_ylabel("True")
        ax.set_title(f"LoRA 7-Class Confusion Matrix (Epoch {epoch})")
        plt.tight_layout()
        plt.savefig(str(output_path), dpi=150)
        plt.close()
        logger.info("Confusion matrix saved to %s", output_path)
    except ImportError:
        logger.warning("matplotlib/seaborn not available — skipping confusion matrix plot")


# ──────────────────────────────────────────────
# Training
# ──────────────────────────────────────────────

def train(args):
    device = args.device
    use_amp = (device == "cuda")
    accumulate_steps = args.accumulate_steps

    # Load model
    encoder = load_model(device, r=args.lora_r, alpha=args.lora_alpha, dropout=args.lora_dropout)

    # Log trainable vs total
    trainable = sum(p.numel() for p in encoder.parameters() if p.requires_grad)
    total = sum(p.numel() for p in encoder.parameters())
    logger.info("Trainable: %dK / Total: %dK (%.1f%%)",
                trainable // 1000, total // 1000, 100 * trainable / total)

    # Datasets
    train_dataset = EmotionDataset(
        args.train_manifest,
        max_duration_sec=8.0,
        phone_augment_prob=args.phone_augment_prob,
        noise_augment_prob=args.noise_augment_prob,
    )
    val_dataset = EmotionDataset(args.val_manifest, max_duration_sec=8.0)

    logger.info("Train: %d samples, Val: %d samples", len(train_dataset), len(val_dataset))
    logger.info("AMP: %s, Accumulate: %d, Effective batch: %d",
                use_amp, accumulate_steps, args.batch_size * accumulate_steps)

    n_workers = 2
    use_pin = (device == "cuda")
    train_loader = DataLoader(
        train_dataset, batch_size=args.batch_size, shuffle=True,
        collate_fn=collate_fn, num_workers=n_workers, drop_last=True,
        pin_memory=use_pin, persistent_workers=(n_workers > 0),
    )
    val_loader = DataLoader(
        val_dataset, batch_size=args.batch_size, shuffle=False,
        collate_fn=collate_fn, num_workers=n_workers,
        pin_memory=use_pin, persistent_workers=(n_workers > 0),
    )

    # Class weights: inverse frequency + disgust 2.5x boost
    class_counts = np.zeros(NUM_CLASSES)
    for sample in train_dataset.samples:
        class_counts[LABEL2IDX[sample["label"]]] += 1
    class_weights = 1.0 / np.maximum(class_counts, 1)
    class_weights = class_weights / class_weights.sum() * NUM_CLASSES
    class_weights[DISGUST_IDX] *= 2.5  # Disgust boost
    logger.info("Class weights: %s",
                {LABELS_7CLASS[i]: round(float(class_weights[i]), 3) for i in range(NUM_CLASSES)})

    criterion = FocalLoss(
        weight=torch.tensor(class_weights, dtype=torch.float32).to(device),
        gamma=2.0,
        label_smoothing=0.05,
    )

    # Optimizer: differential LR
    lora_params = []
    proj_params = list(encoder.proj.parameters())
    proj_ids = {id(p) for p in proj_params}
    for name, param in encoder.named_parameters():
        if param.requires_grad and id(param) not in proj_ids:
            lora_params.append(param)

    optimizer = torch.optim.AdamW([
        {"params": lora_params, "lr": args.lora_lr},
        {"params": proj_params, "lr": args.proj_lr},
    ], weight_decay=args.weight_decay)

    # OneCycleLR scheduler
    steps_per_epoch = max(len(train_loader) // accumulate_steps, 1)
    total_steps = steps_per_epoch * args.epochs
    scheduler = torch.optim.lr_scheduler.OneCycleLR(
        optimizer,
        max_lr=[args.lora_lr, args.proj_lr],
        total_steps=total_steps,
        pct_start=0.1,
        anneal_strategy="cos",
    )

    # Output dir
    output_dir = Path(args.output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)

    # AMP scaler
    scaler = torch.amp.GradScaler("cuda", enabled=use_amp)

    # Training state
    training_log = []
    best_f1 = 0.0
    patience_counter = 0
    start_epoch = 1
    disgust_gate = 0.3

    # Resume from checkpoint if requested
    if args.resume:
        resume_path = output_dir / "last_lora.pt"
        if resume_path.exists():
            resume_state = load_lora_checkpoint(
                encoder, resume_path, device,
                optimizer=optimizer, scheduler=scheduler, scaler=scaler,
            )
            start_epoch = resume_state["epoch"] + 1
            best_f1 = resume_state["best_f1"]
            patience_counter = resume_state["patience_counter"]
            training_log = resume_state["training_log"]
            logger.info("Resuming training from epoch %d (best_f1=%.4f)", start_epoch, best_f1)
        else:
            logger.warning("--resume set but no checkpoint found at %s, starting fresh", resume_path)

    for epoch in range(start_epoch, args.epochs + 1):
        epoch_start = time.time()
        encoder.train()
        total_loss = 0
        correct = 0
        total_samples = 0
        optimizer.zero_grad()

        for batch_idx, (waveforms, labels) in enumerate(train_loader):
            labels = labels.to(device)

            with torch.amp.autocast("cuda", enabled=use_amp):
                logits = forward_pass(encoder, waveforms, device)
                loss = criterion(logits, labels) / accumulate_steps

            scaler.scale(loss).backward()

            if (batch_idx + 1) % accumulate_steps == 0 or (batch_idx + 1) == len(train_loader):
                scaler.unscale_(optimizer)
                trainable_params = [p for p in encoder.parameters() if p.requires_grad]
                torch.nn.utils.clip_grad_norm_(trainable_params, 1.0)
                scaler.step(optimizer)
                scaler.update()
                optimizer.zero_grad()
                scheduler.step()

            total_loss += loss.item() * accumulate_steps * labels.size(0)
            preds = logits.argmax(dim=-1)
            correct += (preds == labels).sum().item()
            total_samples += labels.size(0)

            if (batch_idx + 1) % 10 == 0:
                cur_lr = optimizer.param_groups[0]["lr"]
                logger.info("  Epoch %d [%d/%d] loss=%.4f lr=%.2e",
                            epoch, batch_idx + 1, len(train_loader),
                            loss.item() * accumulate_steps, cur_lr)

        train_loss = total_loss / max(total_samples, 1)
        train_acc = correct / max(total_samples, 1)

        # Validate
        val_metrics = validate(encoder, val_loader, device, criterion)

        epoch_time = time.time() - epoch_start
        logger.info(
            "Epoch %d/%d (%.0fs): train_loss=%.4f train_acc=%.3f | "
            "val_loss=%.4f val_f1=%.3f val_acc=%.3f disgust_f1=%.3f",
            epoch, args.epochs, epoch_time,
            train_loss, train_acc,
            val_metrics["loss"], val_metrics["macro_f1"],
            val_metrics["accuracy"], val_metrics["disgust_f1"],
        )
        logger.info("  Per-class F1: %s", val_metrics["per_class_f1"])

        epoch_log = {
            "epoch": epoch,
            "train_loss": round(train_loss, 4),
            "train_acc": round(train_acc, 4),
            "val_loss": val_metrics["loss"],
            "val_accuracy": val_metrics["accuracy"],
            "val_macro_f1": val_metrics["macro_f1"],
            "val_disgust_f1": val_metrics["disgust_f1"],
            "val_per_class_f1": val_metrics["per_class_f1"],
            "epoch_time_sec": round(epoch_time, 1),
        }
        training_log.append(epoch_log)

        # Disgust F1 Gate + best model save
        gate_pass = val_metrics["disgust_f1"] >= disgust_gate
        if not gate_pass:
            logger.warning("[GATE FAIL] disgust_f1=%.3f < %.1f — not saving as best",
                           val_metrics["disgust_f1"], disgust_gate)

        if val_metrics["macro_f1"] > best_f1 and gate_pass:
            best_f1 = val_metrics["macro_f1"]
            patience_counter = 0
            save_lora_checkpoint(
                encoder, output_dir / "best_lora.pt", epoch, val_metrics, best_f1,
                optimizer=optimizer, scheduler=scheduler, scaler=scaler,
                training_log=training_log,
            )
            logger.info("  New best! macro_f1=%.4f (gate passed) saved to best_lora.pt", best_f1)

            if "confusion_matrix" in val_metrics:
                cm = np.array(val_metrics["confusion_matrix"])
                plot_confusion_matrix(cm, output_dir / f"confusion_matrix_epoch{epoch}.png", epoch)
        elif gate_pass:
            patience_counter += 1
        else:
            # Gate fail does not increment patience
            pass

        # Always save last (with full state for resume)
        save_lora_checkpoint(
            encoder, output_dir / "last_lora.pt", epoch, val_metrics, best_f1, patience_counter,
            optimizer=optimizer, scheduler=scheduler, scaler=scaler,
            training_log=training_log,
        )

        # Save training log
        with open(output_dir / "training_log.json", "w") as f:
            json.dump(training_log, f, indent=2)

        # Early stopping (only on gate-passing epochs)
        if patience_counter >= args.patience:
            logger.info("Early stopping at epoch %d (patience=%d)", epoch, args.patience)
            break

        if device == "cuda":
            torch.cuda.empty_cache()

    # Save config
    config = {
        "model": "iic/emotion2vec_plus_base",
        "method": "LoRA",
        "lora_r": args.lora_r,
        "lora_alpha": args.lora_alpha,
        "num_classes": NUM_CLASSES,
        "labels": LABELS_7CLASS,
        "label2idx": LABEL2IDX,
        "best_val_f1": best_f1,
        "training_args": vars(args),
    }
    with open(output_dir / "config.json", "w") as f:
        json.dump(config, f, indent=2, ensure_ascii=False)

    logger.info("Training complete. Best F1=%.4f at %s", best_f1, output_dir / "best_lora.pt")


# ──────────────────────────────────────────────
# Main
# ──────────────────────────────────────────────

def main():
    parser = argparse.ArgumentParser(description="LoRA fine-tune emotion2vec 7-class")
    parser.add_argument("--train-manifest", required=True)
    parser.add_argument("--val-manifest", required=True)
    parser.add_argument("--output-dir", default="data/models/lora_emotion2vec_7class")
    parser.add_argument("--device", default="cpu", choices=["cpu", "cuda"])
    parser.add_argument("--epochs", type=int, default=20)
    parser.add_argument("--batch-size", type=int, default=4)
    parser.add_argument("--accumulate-steps", type=int, default=8)
    parser.add_argument("--lora-r", type=int, default=16)
    parser.add_argument("--lora-alpha", type=int, default=32)
    parser.add_argument("--lora-dropout", type=float, default=0.1)
    parser.add_argument("--lora-lr", type=float, default=2e-4)
    parser.add_argument("--proj-lr", type=float, default=2e-3)
    parser.add_argument("--weight-decay", type=float, default=0.01)
    parser.add_argument("--patience", type=int, default=7)
    parser.add_argument("--phone-augment-prob", type=float, default=0.3)
    parser.add_argument("--noise-augment-prob", type=float, default=0.15)
    parser.add_argument("--resume", action="store_true",
                        help="Resume from last_lora.pt checkpoint in output-dir")
    args = parser.parse_args()

    train(args)


if __name__ == "__main__":
    main()