scripts/05_multiscale_cnn.py

from pathlib import Path
import numpy as np
import pandas as pd
from tqdm import tqdm
import librosa
import warnings
warnings.filterwarnings("ignore")

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder
from sklearn.metrics import f1_score, classification_report, confusion_matrix
from joblib import Parallel, delayed

DATA_DIR = Path("output/linguawave")
SAMPLE_RATE = 16_000
DURATION = 10
N_SAMPLES = SAMPLE_RATE * DURATION
CLASSES = ["id", "ms", "vi", "th", "en", "zh", "ar", "fr"]
N_CLASSES = len(CLASSES)

HARD_NEG_LANGS = {"id", "ms"}
OVERSAMPLE_FACTOR = 2

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Device:", DEVICE)

BRANCH_CONFIGS = [
    {"n_fft": 512,  "hop_length": 256, "n_mels": 128},
    {"n_fft": 1024, "hop_length": 256, "n_mels": 128},
    {"n_fft": 2048, "hop_length": 256, "n_mels": 128},
]
TARGET_T = N_SAMPLES // 256 + 1   # same for all branches (hop=256)


def _compute_mels_one(row_id):
    """Compute 3-branch mels for one audio file. Returns (3, 128, TARGET_T) float32."""
    y, _ = librosa.load(str(DATA_DIR / row_id), sr=SAMPLE_RATE, duration=DURATION)
    if len(y) < N_SAMPLES:
        y = np.pad(y, (0, N_SAMPLES - len(y)))
    else:
        y = y[:N_SAMPLES]
    out = []
    for cfg in BRANCH_CONFIGS:
        mel = librosa.feature.melspectrogram(
            y=y, sr=SAMPLE_RATE,
            n_mels=cfg["n_mels"], n_fft=cfg["n_fft"], hop_length=cfg["hop_length"]
        )
        log_mel = librosa.power_to_db(mel, ref=np.max).astype(np.float32)
        log_mel = (log_mel - log_mel.mean()) / (log_mel.std() + 1e-8)
        if log_mel.shape[1] < TARGET_T:
            log_mel = np.pad(log_mel, ((0, 0), (0, TARGET_T - log_mel.shape[1])))
        else:
            log_mel = log_mel[:, :TARGET_T]
        out.append(log_mel)
    return np.stack(out, axis=0)   # (3, 128, TARGET_T)


def precompute_mels(df, cache_path):
    """Compute mels for all rows in df (parallel), save to cache_path as (N,3,128,T)."""
    cp = Path(cache_path)
    if cp.exists():
        print(f"[cache] Loading {cp.name}")
        return np.load(cp, mmap_mode='r')
    print(f"Precomputing mels → {cp.name}  ({len(df)} samples, parallel) ...")
    results = Parallel(n_jobs=-1, prefer="threads")(
        delayed(_compute_mels_one)(r["id"])
        for _, r in tqdm(df.iterrows(), total=len(df))
    )
    arr = np.stack(results, axis=0)   # (N, 3, 128, TARGET_T)
    np.save(cp, arr)
    print(f"[cache] Saved {cp.name}  shape={arr.shape}")
    return np.load(cp, mmap_mode='r')


class CachedMultiScaleDataset(Dataset):
    """Loads precomputed (3, 128, T) mel arrays; augmentation via time-roll on mel."""
    def __init__(self, df, mels_cache, label_encoder, augment=False):
        self.df = df.reset_index(drop=True)
        self.mels = mels_cache       # numpy mmap array (N, 3, 128, T)
        self.le = label_encoder
        self.augment = augment
        self.has_labels = "label" in df.columns

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

    def __getitem__(self, idx):
        row = self.df.iloc[idx]
        # mels shape: (3, 128, T)
        mels = self.mels[idx].copy()   # copy slice from mmap
        if self.augment:
            shift = np.random.randint(-SAMPLE_RATE // 2 // 256, SAMPLE_RATE // 2 // 256)
            mels = np.roll(mels, shift, axis=-1)
        mels_tensors = [torch.tensor(mels[b][np.newaxis, :, :]) for b in range(3)]
        if self.has_labels:
            label = self.le.transform([row["label"]])[0]
            return mels_tensors, label
        return mels_tensors


class SingleBranchCNN(nn.Module):
    def __init__(self, out_dim=256):
        super().__init__()
        def block(ic, oc):
            return nn.Sequential(
                nn.Conv2d(ic, oc, 3, padding=1), nn.BatchNorm2d(oc), nn.ReLU(True),
                nn.Conv2d(oc, oc, 3, padding=1), nn.BatchNorm2d(oc), nn.ReLU(True),
                nn.MaxPool2d(2, 2),
            )
        self.net = nn.Sequential(
            block(1, 32), block(32, 64), block(64, 128), block(128, out_dim),
            nn.AdaptiveAvgPool2d(1),
            nn.Flatten(),
        )

    def forward(self, x):
        return self.net(x)


class MultiScaleCNN(nn.Module):
    def __init__(self, n_classes=N_CLASSES, branch_dim=256):
        super().__init__()
        self.branch1 = SingleBranchCNN(branch_dim)
        self.branch2 = SingleBranchCNN(branch_dim)
        self.branch3 = SingleBranchCNN(branch_dim)
        self.head = nn.Sequential(
            nn.Linear(branch_dim * 3, 256),
            nn.ReLU(True),
            nn.Dropout(0.4),
            nn.Linear(256, n_classes),
        )

    def forward(self, mels):
        m1, m2, m3 = mels
        return self.head(torch.cat([self.branch1(m1), self.branch2(m2), self.branch3(m3)], dim=1))


# ── data ──────────────────────────────────────────────────────────────
train_df = pd.read_csv(DATA_DIR / "train.csv")
test_df  = pd.read_csv(DATA_DIR / "test.csv")

le = LabelEncoder()
le.fit(CLASSES)

tr_df, val_df = train_test_split(
    train_df, test_size=0.15, random_state=42, stratify=train_df["label"]
)

hard_rows = tr_df[tr_df["label"].isin(HARD_NEG_LANGS)]
tr_df_aug = pd.concat(
    [tr_df] + [hard_rows] * (OVERSAMPLE_FACTOR - 1),
    ignore_index=True
).sample(frac=1, random_state=42).reset_index(drop=True)

print(f"Train (augmented): {len(tr_df_aug)}  Val: {len(val_df)}  Test: {len(test_df)}")

# Precompute mels for unique train+val samples and test
Path("cache").mkdir(exist_ok=True)
full_train_mels = precompute_mels(train_df.reset_index(drop=True), "cache/lw_05_train_mels.npy")
test_mels       = precompute_mels(test_df.reset_index(drop=True),  "cache/lw_05_test_mels.npy")

# Build index maps: tr_df_aug / val_df rows → full_train_mels row index
train_id_to_idx = {row["id"]: i for i, row in train_df.reset_index(drop=True).iterrows()}

tr_aug_mels = np.stack([full_train_mels[train_id_to_idx[r["id"]]] for _, r in tr_df_aug.iterrows()])
val_mels    = np.stack([full_train_mels[train_id_to_idx[r["id"]]] for _, r in val_df.iterrows()])

print("Building augmented train mel array ...")
print(f"  tr_aug shape: {tr_aug_mels.shape}")

# Reset df indices to match stacked arrays
tr_df_aug_reset = tr_df_aug.reset_index(drop=True)
val_df_reset    = val_df.reset_index(drop=True)


def collate_fn(batch):
    mels_list = [item[0] for item in batch]
    labels    = torch.tensor([item[1] for item in batch])
    stacked = [torch.stack([m[b] for m in mels_list]) for b in range(3)]
    return stacked, labels

def collate_fn_test(batch):
    stacked = [torch.stack([item[b] for item in batch]) for b in range(3)]
    return stacked

BATCH = 64
train_ds = CachedMultiScaleDataset(tr_df_aug_reset, tr_aug_mels, le, augment=True)
val_ds   = CachedMultiScaleDataset(val_df_reset,    val_mels,    le, augment=False)
test_ds  = CachedMultiScaleDataset(test_df.reset_index(drop=True), test_mels, le, augment=False)

train_loader = DataLoader(train_ds, batch_size=BATCH, shuffle=True,
                          num_workers=4, pin_memory=True, collate_fn=collate_fn)
val_loader   = DataLoader(val_ds,   batch_size=BATCH, shuffle=False,
                          num_workers=4, pin_memory=True, collate_fn=collate_fn)
test_loader  = DataLoader(test_ds,  batch_size=BATCH, shuffle=False,
                          num_workers=4, pin_memory=True, collate_fn=collate_fn_test)

print(f"Train batches: {len(train_loader)}  Val batches: {len(val_loader)}")

# ── model ──────────────────────────────────────────────────────────────
model = MultiScaleCNN().to(DEVICE)
print(f"Trainable params: {sum(p.numel() for p in model.parameters() if p.requires_grad):,}")

from sklearn.utils.class_weight import compute_class_weight
class_weights = compute_class_weight("balanced", classes=le.classes_, y=tr_df["label"].to_numpy())
for lang in HARD_NEG_LANGS:
    class_weights[le.transform([lang])[0]] /= OVERSAMPLE_FACTOR
class_weights_tensor = torch.tensor(class_weights, dtype=torch.float).to(DEVICE)

criterion = nn.CrossEntropyLoss(weight=class_weights_tensor)
EPOCHS = 5
optimizer = optim.AdamW(model.parameters(), lr=5e-4, weight_decay=1e-4)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=EPOCHS)

best_f1 = 0.0
best_weights = None

for epoch in range(1, EPOCHS + 1):
    model.train()
    running_loss = 0.0
    for mels, y_batch in tqdm(train_loader, desc=f"Epoch {epoch:02d} train", leave=False):
        mels = [m.to(DEVICE) for m in mels]
        y_batch = y_batch.to(DEVICE)
        optimizer.zero_grad()
        loss = criterion(model(mels), y_batch)
        loss.backward()
        nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()
        running_loss += loss.item() * y_batch.size(0)

    model.eval()
    all_preds, all_labels = [], []
    with torch.no_grad():
        for mels, y_batch in val_loader:
            mels = [m.to(DEVICE) for m in mels]
            all_preds.extend(model(mels).argmax(dim=1).cpu().numpy())
            all_labels.extend(y_batch.numpy())

    val_f1 = f1_score(all_labels, all_preds, average="macro")
    scheduler.step()
    if val_f1 > best_f1:
        best_f1 = val_f1
        best_weights = {k: v.clone() for k, v in model.state_dict().items()}
    print(f"Epoch {epoch:02d}/{EPOCHS}  val_F1={val_f1:.4f}  best={best_f1:.4f}")

print(f"\nBest validation Macro F1: {best_f1:.4f}")

import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import seaborn as sns

model.load_state_dict(best_weights)
model.eval()
all_preds, all_labels = [], []
with torch.no_grad():
    for mels, y_batch in val_loader:
        mels = [m.to(DEVICE) for m in mels]
        all_preds.extend(model(mels).argmax(dim=1).cpu().numpy())
        all_labels.extend(y_batch.numpy())

print(classification_report(all_labels, all_preds, target_names=le.classes_))

cm = confusion_matrix(all_labels, all_preds)
id_idx = le.transform(["id"])[0]; ms_idx = le.transform(["ms"])[0]
print(f"id→ms confusions: {cm[id_idx, ms_idx]}  |  ms→id confusions: {cm[ms_idx, id_idx]}")

Path("submissions").mkdir(exist_ok=True)
torch.save(best_weights, "submissions/model_approach5_multiscale.pt")

all_probs = []
with torch.no_grad():
    for mels in tqdm(test_loader, desc="Test inference"):
        mels = [m.to(DEVICE) for m in mels]
        all_probs.append(torch.softmax(model(mels), dim=1).cpu().numpy())

test_probs = np.vstack(all_probs)
np.save("submissions/probs_approach5_multiscale.npy", test_probs)
test_preds = le.inverse_transform(test_probs.argmax(axis=1))
sub = pd.DataFrame({"id": test_df["id"], "label": test_preds})
sub.to_csv("submissions/sub_approach5_multiscale.csv", index=False)
print("Saved submissions/sub_approach5_multiscale.csv")

probs4_path = Path("submissions/probs_approach4_cnn_mel.npy")
probs5_path = Path("submissions/probs_approach5_multiscale.npy")
if probs4_path.exists() and probs5_path.exists():
    ensemble_probs = (np.load(probs4_path) + np.load(probs5_path)) / 2
    ensemble_preds = le.inverse_transform(ensemble_probs.argmax(axis=1))
    pd.DataFrame({"id": test_df["id"], "label": ensemble_preds}).to_csv(
        "submissions/sub_ensemble_4_5.csv", index=False)
    print("Ensemble submission saved: submissions/sub_ensemble_4_5.csv")
else:
    print("Run notebook 04 first to generate probs_approach4_cnn_mel.npy")