scripts/03_bag_of_codewords.py

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

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

N_CLUSTERS = 64
FRAME_SUBSAMPLE = 5   # keep every 5th MFCC frame for K-means fitting
N_MFCC = 20


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

from joblib import Parallel, delayed
Path("cache").mkdir(exist_ok=True)

def _get_frames(row_id):
    y, _ = librosa.load(str(DATA_DIR / row_id), sr=SAMPLE_RATE, duration=DURATION)
    target_len = SAMPLE_RATE * DURATION
    if len(y) < target_len:
        y = np.pad(y, (0, target_len - len(y)))
    else:
        y = y[:target_len]
    mfcc = librosa.feature.mfcc(y=y, sr=SAMPLE_RATE, n_mfcc=N_MFCC)
    return mfcc.T[::FRAME_SUBSAMPLE]

print("Collecting MFCC frames from training files ...")
frame_cp = Path("cache") / "lw_03_frames.npy"
if frame_cp.exists():
    print("[cache] Loading lw_03_frames.npy"); all_frames = np.load(frame_cp)
else:
    results = Parallel(n_jobs=-1, prefer="threads")(
        delayed(_get_frames)(r["id"]) for _, r in tqdm(train_df.iterrows(), total=len(train_df)))
    all_frames = np.vstack(results)
    np.save(frame_cp, all_frames); print("[cache] Saved lw_03_frames.npy")

print(f"Total subsampled frames: {all_frames.shape}")


from sklearn.cluster import MiniBatchKMeans

print(f"Fitting KMeans with k={N_CLUSTERS} ...")
kmeans = MiniBatchKMeans(n_clusters=N_CLUSTERS, random_state=42, batch_size=4096, n_init=5)
kmeans.fit(all_frames)
print("KMeans fitting complete.")


def file_to_histogram(fpath, kmeans=kmeans):
    """Assign each MFCC frame to a cluster and return normalised frequency histogram."""
    y, _ = librosa.load(str(fpath), sr=SAMPLE_RATE, duration=DURATION)
    target_len = SAMPLE_RATE * DURATION
    if len(y) < target_len:
        y = np.pad(y, (0, target_len - len(y)))
    else:
        y = y[:target_len]
    mfcc = librosa.feature.mfcc(y=y, sr=SAMPLE_RATE, n_mfcc=N_MFCC).T  # (T, 20)
    labels = kmeans.predict(mfcc)
    hist, _ = np.histogram(labels, bins=N_CLUSTERS, range=(0, N_CLUSTERS))
    return hist.astype(np.float32) / (hist.sum() + 1e-8)   # normalise

print("Building train histograms ...")
train_hist_cp = Path("cache") / "lw_03_train_hist.npy"
test_hist_cp  = Path("cache") / "lw_03_test_hist.npy"
if train_hist_cp.exists():
    print("[cache] Loading histograms")
    X_train_hist = np.load(train_hist_cp)
    X_test_hist  = np.load(test_hist_cp)
else:
    X_train_hist = np.array(Parallel(n_jobs=-1, prefer="threads")(
        delayed(file_to_histogram)(DATA_DIR / r["id"])
        for _, r in tqdm(train_df.iterrows(), total=len(train_df))))
    print("Building test histograms ...")
    X_test_hist = np.array(Parallel(n_jobs=-1, prefer="threads")(
        delayed(file_to_histogram)(DATA_DIR / r["id"])
        for _, r in tqdm(test_df.iterrows(), total=len(test_df))))
    np.save(train_hist_cp, X_train_hist); np.save(test_hist_cp, X_test_hist)
    print("[cache] Saved histograms")
y_train = train_df["label"].to_numpy()
print("X_train_hist shape:", X_train_hist.shape)
print("X_test_hist shape:", X_test_hist.shape)


from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder

le = LabelEncoder()
le.fit(CLASSES)
y_enc = le.transform(y_train)

X_tr, X_val, y_tr, y_val = train_test_split(
    X_train_hist, y_enc, test_size=0.15, random_state=42, stratify=y_enc
)
print(f"Train: {X_tr.shape[0]}  Val: {X_val.shape[0]}")


from sklearn.linear_model import LogisticRegression
from sklearn.metrics import f1_score, classification_report

lr = LogisticRegression(max_iter=1000, C=1.0,
                        solver="lbfgs", random_state=42, n_jobs=-1)
lr.fit(X_tr, y_tr)
print("Logistic Regression training complete.")


val_preds = lr.predict(X_val)
macro_f1 = f1_score(y_val, val_preds, average="macro")
print(f"Validation Macro F1: {macro_f1:.4f}")
print()
print(classification_report(y_val, val_preds, target_names=le.classes_))


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

fig, axes = plt.subplots(2, 4, figsize=(16, 6), sharey=False)
axes = axes.flatten()

for idx, lang in enumerate(CLASSES):
    mask = y_train == lang
    avg_hist = X_train_hist[mask].mean(axis=0)
    axes[idx].bar(range(N_CLUSTERS), avg_hist)
    axes[idx].set_title(lang)
    axes[idx].set_xlabel("Codeword")
    axes[idx].set_ylabel("Freq")

plt.suptitle("Average codeword histogram per language")
plt.tight_layout()
plt.savefig("/dev/null")


test_preds_enc = lr.predict(X_test_hist)
test_preds = le.inverse_transform(test_preds_enc)

sub = pd.DataFrame({"id": test_df["id"], "label": test_preds})
Path("submissions").mkdir(exist_ok=True)
sub.to_csv("submissions/sub_approach3_bag_of_codewords.csv", index=False)
print("Saved submissions/sub_approach3_bag_of_codewords.csv")
sub.head()