scripts/00_starter.py

from pathlib import Path
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import librosa
import librosa.display
import warnings
warnings.filterwarnings("ignore")

DATA_DIR = Path("output/linguawave")
SAMPLE_RATE = 16_000
DURATION = 10          # seconds
N_SAMPLES = SAMPLE_RATE * DURATION   # 160_000
CLASSES = ["id", "ms", "vi", "th", "en", "zh", "ar", "fr"]
TONAL = {"vi", "th", "zh"}


train_df = pd.read_csv(DATA_DIR / "train.csv")
test_df  = pd.read_csv(DATA_DIR / "test.csv")
print("Train shape:", train_df.shape)
print("Test  shape:", test_df.shape)
train_df.head()


counts = train_df["language"].value_counts().reindex(CLASSES)
colors = ["#e15759" if c in TONAL else "#4e79a7" for c in CLASSES]

fig, ax = plt.subplots(figsize=(9, 4))
bars = ax.bar(CLASSES, counts.values, color=colors)
ax.set_title("Training samples per language")
ax.set_xlabel("Language")
ax.set_ylabel("Count")

# legend
from matplotlib.patches import Patch
legend_elements = [
    Patch(facecolor="#e15759", label="Tonal (vi, th, zh)"),
    Patch(facecolor="#4e79a7", label="Non-tonal"),
]
ax.legend(handles=legend_elements)

for bar, count in zip(bars, counts.values):
    ax.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 5,
            str(count), ha="center", va="bottom", fontsize=9)
plt.tight_layout()
plt.show()
print(counts)


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

for idx, lang in enumerate(CLASSES):
    row = train_df[train_df["language"] == lang].iloc[0]
    fpath = DATA_DIR / row["id"]
    y, sr = librosa.load(fpath, sr=SAMPLE_RATE, duration=DURATION)

    mel = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128, n_fft=1024, hop_length=256)
    mel_db = librosa.power_to_db(mel, ref=np.max)

    ax = axes[idx]
    librosa.display.specshow(mel_db, sr=sr, hop_length=256,
                             x_axis="time", y_axis="mel", ax=ax)
    marker = " ♪" if lang in TONAL else ""
    ax.set_title(f"{lang}{marker}", fontsize=12,
                 color="#e15759" if lang in TONAL else "#4e79a7")
    ax.set_xlabel("")
    ax.set_ylabel("")

fig.suptitle("Mel Spectrograms – one sample per language\n(red = tonal language)", y=1.02)
plt.tight_layout()
plt.show()


def extract_features_skeleton(fpath, sr=SAMPLE_RATE, n_mfcc=20):
    """Skeleton – returns None; replace with real implementation."""
    y, _ = librosa.load(fpath, sr=sr, duration=DURATION)
    # TODO: compute features
    mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=n_mfcc)
    feats = np.concatenate([mfcc.mean(axis=1), mfcc.std(axis=1)])  # 40-dim
    return feats

# Quick sanity check
sample_path = DATA_DIR / train_df.iloc[0]["id"]
feats = extract_features_skeleton(sample_path)
print("Feature vector shape:", feats.shape)


from sklearn.dummy import DummyClassifier
from sklearn.metrics import f1_score
from sklearn.preprocessing import LabelEncoder

# Encode labels
le = LabelEncoder()
le.fit(CLASSES)

# Dummy model (replace with real model)
dummy = DummyClassifier(strategy="most_frequent", random_state=42)
dummy.fit([[0]] * len(train_df), le.transform(train_df["language"]))

# Predict on test set
test_preds = dummy.predict([[0]] * len(test_df))
test_labels = le.inverse_transform(test_preds)

# Build submission CSV
sub = pd.DataFrame({"id": test_df["id"], "language": test_labels})
sub_dir = Path("submissions")
sub_dir.mkdir(exist_ok=True)
sub.to_csv(sub_dir / "sub_approach0_starter.csv", index=False)
print("Submission saved!")
print(sub.head())
print()
print("Macro F1 (on dummy val) – replace with real validation:")
# Example validation snippet
# val_preds = model.predict(X_val)
# f1 = f1_score(y_val, val_preds, average="macro")
# print(f"Macro F1: {f1:.4f}")