File size: 9,811 Bytes
f35a6e2
 
 
4fd0c50
 
f35a6e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4fd0c50
f35a6e2
 
 
 
 
 
 
 
 
 
 
 
4fd0c50
 
 
 
 
f35a6e2
 
 
 
 
 
 
 
 
4fd0c50
 
 
 
 
 
 
 
 
 
 
f35a6e2
 
 
 
 
 
4fd0c50
f35a6e2
4fd0c50
f35a6e2
4fd0c50
 
f35a6e2
 
4fd0c50
 
f35a6e2
 
 
 
4fd0c50
 
 
f35a6e2
 
 
4fd0c50
f35a6e2
 
 
 
 
 
 
4fd0c50
f35a6e2
 
 
 
 
 
 
 
 
 
 
 
 
4fd0c50
f35a6e2
 
 
 
 
 
 
 
 
 
 
 
 
4fd0c50
f35a6e2
 
 
 
 
 
 
 
 
 
 
4fd0c50
 
f35a6e2
 
 
 
 
 
 
 
 
 
 
 
 
4fd0c50
 
 
 
 
 
 
 
 
 
 
 
f35a6e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4fd0c50
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f35a6e2
 
 
 
 
4fd0c50
f35a6e2
 
4fd0c50
 
 
f35a6e2
4fd0c50
 
f35a6e2
 
4fd0c50
 
 
 
 
f35a6e2
 
 
 
 
 
 
 
4fd0c50
 
 
f35a6e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4fd0c50
f35a6e2
 
 
 
 
 
 
 
 
4fd0c50
f35a6e2
 
 
 
 
 
 
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
"""
Training loop for Wav2Vec-based deepfake audio detection.

Supports both Stage 1 (frozen backbone, simple) and Stage 2 (fine-tuning,
with mixed precision + warmup scheduler).
"""

import os
import time
from dataclasses import dataclass, field
from typing import Optional, Callable, List

import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from tqdm import tqdm

from src.evaluation.metrics import (
    compute_eer,
    compute_auc,
    aggregate_window_scores_to_utterance,
)


@dataclass
class TrainConfig:
    """Hyperparameters for one training stage."""
    learning_rate: float = 1e-3
    batch_size: int = 32
    epochs: int = 5
    weight_decay: float = 0.01
    grad_clip: float = 1.0
    early_stopping_patience: int = 3
    checkpoint_dir: str = "/content/drive/MyDrive/deepfake_audio/checkpoints"
    checkpoint_name: str = "stage1_best.pt"
    log_every_n_steps: int = 20
    use_wandb: bool = True
    wandb_project: str = "deepfake-audio-detection"
    wandb_run_name: Optional[str] = None
    class_weights: Optional[List[float]] = None
    # Stage 2 additions
    use_mixed_precision: bool = False
    warmup_ratio: float = 0.0  # fraction of total steps used for LR warmup
    use_lr_scheduler: bool = False  # set True with warmup_ratio > 0


def make_loss_fn(class_weights: Optional[List[float]], device: str) -> Callable:
    if class_weights is not None:
        weights = torch.tensor(class_weights, dtype=torch.float32, device=device)
        return nn.CrossEntropyLoss(weight=weights)
    return nn.CrossEntropyLoss()


def make_lr_scheduler(optimizer, total_steps: int, warmup_ratio: float):
    """Linear warmup followed by linear decay to zero."""
    warmup_steps = int(total_steps * warmup_ratio)
    def lr_lambda(step):
        if step < warmup_steps:
            return step / max(1, warmup_steps)
        progress = (step - warmup_steps) / max(1, total_steps - warmup_steps)
        return max(0.0, 1.0 - progress)
    return torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)


@torch.no_grad()
def evaluate(
    model: nn.Module,
    dev_loader: DataLoader,
    device: str,
    desc: str = "Eval",
    use_mixed_precision: bool = False,
) -> dict:
    """Run inference over the dev set and compute per-utterance metrics."""
    model.eval()
    all_window_scores, all_window_labels, all_window_utts = [], [], []
    total_loss, n_batches = 0.0, 0
    loss_fn = nn.CrossEntropyLoss()

    autocast_ctx = torch.amp.autocast(device_type="cuda", enabled=use_mixed_precision)

    for waveforms, labels, utt_ids in tqdm(dev_loader, desc=desc, leave=False):
        waveforms = waveforms.to(device, non_blocking=True)
        labels_gpu = labels.to(device, non_blocking=True)

        with autocast_ctx:
            logits = model(waveforms)
            loss = loss_fn(logits, labels_gpu)
        total_loss += loss.item()
        n_batches += 1

        probs = torch.softmax(logits.float(), dim=-1)
        spoof_probs = probs[:, 1].detach().cpu().numpy()

        all_window_scores.extend(spoof_probs.tolist())
        all_window_labels.extend(labels.tolist())
        all_window_utts.extend(list(utt_ids))

    utt_scores, utt_ids_sorted = aggregate_window_scores_to_utterance(
        np.array(all_window_scores), all_window_utts, method="mean",
    )
    label_map = {}
    for s, l, u in zip(all_window_scores, all_window_labels, all_window_utts):
        if u not in label_map:
            label_map[u] = l
    utt_labels = np.array([label_map[u] for u in utt_ids_sorted])

    eer, threshold = compute_eer(utt_scores, utt_labels)
    auc = compute_auc(utt_scores, utt_labels)
    preds = (utt_scores > threshold).astype(int)
    accuracy = float((preds == utt_labels).mean())

    return {
        "eer": eer, "auc": auc, "accuracy": accuracy,
        "threshold": float(threshold),
        "loss": total_loss / max(n_batches, 1),
        "n_utterances": len(utt_ids_sorted),
    }


def train(
    model: nn.Module,
    train_loader: DataLoader,
    dev_loader: DataLoader,
    config: TrainConfig,
    device: str = "cuda",
) -> dict:
    """Train the model for `config.epochs` epochs, evaluating each epoch."""
    wandb = None
    if config.use_wandb:
        import wandb as _wandb
        run = _wandb.init(
            project=config.wandb_project,
            name=config.wandb_run_name,
            config={
                "learning_rate": config.learning_rate,
                "batch_size": config.batch_size,
                "epochs": config.epochs,
                "weight_decay": config.weight_decay,
                "use_mixed_precision": config.use_mixed_precision,
                "warmup_ratio": config.warmup_ratio,
                "trainable_params": sum(p.numel() for p in model.parameters() if p.requires_grad),
            },
            settings=_wandb.Settings(init_timeout=180),
        )
        wandb = _wandb

    trainable = [p for p in model.parameters() if p.requires_grad]
    optimizer = torch.optim.AdamW(
        trainable,
        lr=config.learning_rate,
        weight_decay=config.weight_decay,
    )

    # Optional LR scheduler with warmup
    scheduler = None
    if config.use_lr_scheduler and config.warmup_ratio > 0:
        total_steps = len(train_loader) * config.epochs
        scheduler = make_lr_scheduler(optimizer, total_steps, config.warmup_ratio)

    # Mixed precision setup
    scaler = torch.amp.GradScaler("cuda", enabled=config.use_mixed_precision)
    autocast_ctx_factory = lambda: torch.amp.autocast(
        device_type="cuda", enabled=config.use_mixed_precision
    )

    loss_fn = make_loss_fn(config.class_weights, device)

    os.makedirs(config.checkpoint_dir, exist_ok=True)
    checkpoint_path = os.path.join(config.checkpoint_dir, config.checkpoint_name)

    history = {"train_loss": [], "dev_eer": [], "dev_auc": [], "dev_accuracy": []}
    best_eer = float("inf")
    epochs_without_improvement = 0
    global_step = 0

    for epoch in range(config.epochs):
        model.train()
        epoch_start = time.time()
        epoch_losses = []

        pbar = tqdm(train_loader, desc=f"Epoch {epoch+1}/{config.epochs}")
        for waveforms, labels, utt_ids in pbar:
            waveforms = waveforms.to(device, non_blocking=True)
            labels = labels.to(device, non_blocking=True)

            optimizer.zero_grad()
            with autocast_ctx_factory():
                logits = model(waveforms)
                loss = loss_fn(logits, labels)

            if config.use_mixed_precision:
                scaler.scale(loss).backward()
                scaler.unscale_(optimizer)
                torch.nn.utils.clip_grad_norm_(trainable, config.grad_clip)
                scaler.step(optimizer)
                scaler.update()
            else:
                loss.backward()
                torch.nn.utils.clip_grad_norm_(trainable, config.grad_clip)
                optimizer.step()

            if scheduler is not None:
                scheduler.step()

            epoch_losses.append(loss.item())
            global_step += 1

            if wandb is not None and global_step % config.log_every_n_steps == 0:
                log_data = {
                    "train/step_loss": loss.item(),
                    "train/global_step": global_step,
                    "train/lr": optimizer.param_groups[0]["lr"],
                }
                wandb.log(log_data)

            pbar.set_postfix(loss=f"{loss.item():.4f}",
                            lr=f"{optimizer.param_groups[0]['lr']:.2e}")

        train_loss = float(np.mean(epoch_losses))
        dev_metrics = evaluate(
            model, dev_loader, device,
            desc=f"Epoch {epoch+1} dev",
            use_mixed_precision=config.use_mixed_precision,
        )
        epoch_time = time.time() - epoch_start

        history["train_loss"].append(train_loss)
        history["dev_eer"].append(dev_metrics["eer"])
        history["dev_auc"].append(dev_metrics["auc"])
        history["dev_accuracy"].append(dev_metrics["accuracy"])

        print(f"\nEpoch {epoch+1}/{config.epochs} ({epoch_time:.0f}s)")
        print(f"  train_loss:   {train_loss:.4f}")
        print(f"  dev_eer:      {dev_metrics['eer']*100:.2f}%")
        print(f"  dev_auc:      {dev_metrics['auc']:.4f}")
        print(f"  dev_accuracy: {dev_metrics['accuracy']*100:.2f}%")

        if wandb is not None:
            wandb.log({
                "epoch": epoch + 1,
                "train/epoch_loss": train_loss,
                "dev/eer": dev_metrics["eer"],
                "dev/auc": dev_metrics["auc"],
                "dev/accuracy": dev_metrics["accuracy"],
                "dev/loss": dev_metrics["loss"],
            })

        if dev_metrics["eer"] < best_eer:
            best_eer = dev_metrics["eer"]
            epochs_without_improvement = 0
            torch.save({
                "epoch": epoch + 1,
                "model_state_dict": model.state_dict(),
                "optimizer_state_dict": optimizer.state_dict(),
                "scheduler_state_dict": scheduler.state_dict() if scheduler else None,
                "best_eer": best_eer,
                "config": vars(config),
            }, checkpoint_path)
            print(f"  → Saved best checkpoint (EER={best_eer*100:.2f}%)")
        else:
            epochs_without_improvement += 1
            print(f"  No improvement for {epochs_without_improvement} epoch(s)")

        if epochs_without_improvement >= config.early_stopping_patience:
            print(f"\nEarly stopping after {epoch+1} epochs.")
            break

    if wandb is not None:
        wandb.summary["best_dev_eer"] = best_eer
        wandb.finish()

    return {"history": history, "best_eer": best_eer, "checkpoint_path": checkpoint_path}