extract_features.py

import torch
import pickle
import numpy as np
from datasets import load_dataset
from transformers import AutoProcessor, AutoModel
from huggingface_hub import HfApi
from tqdm import tqdm

api = HfApi()

print("Cargando dataset desde HuggingFace...")
dataset = load_dataset(
    "MrlolDev/voxtral-emotion-speech",
    split="train+validation+test",
)
print(f"Total clips: {len(dataset)}")

print("Cargando Voxtral encoder...")
MODEL_ID = "mistralai/Voxtral-Mini-4B-Realtime-2602"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModel.from_pretrained(MODEL_ID, torch_dtype=torch.float16)
model.eval()
device = torch.device("cuda")
model = model.to(device)
print(f"Modelo en GPU: {next(model.parameters()).device}")

EMOTIONS = ["neutral", "happy", "sad", "angry", "fear", "surprise"]


def extract_features(audio_array, sampling_rate):
    inputs = processor(audio_array, sampling_rate=sampling_rate, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        encoder_output = model.encoder(**inputs)
        hidden = encoder_output.last_hidden_state  # (1, T, 1280)
        features = hidden.mean(dim=1).squeeze(0)  # (1280,)

    return features.cpu().float().numpy()


records = []
errors = 0

print("Extrayendo features...")
for sample in tqdm(dataset):
    try:
        audio_array = np.array(sample["audio"]["array"], dtype=np.float32)
        sr = sample["audio"]["sampling_rate"]

        features = extract_features(audio_array, sr)

        records.append(
            {
                "features": features,
                "label": EMOTIONS.index(sample["emotion"]),
                "emotion": sample["emotion"],
                "split": sample.get("split", "train"),
                "sensevoice_score": sample.get("sensevoice_score", 0.0),
            }
        )
    except Exception as e:
        errors += 1
        print(f"Error en sample: {e}")

print(f"\nExtracted: {len(records)} features | Errors: {errors}")

with open("features.pkl", "wb") as f:
    pickle.dump(records, f)

print("Saved to features.pkl")

from collections import Counter

dist = Counter(r["emotion"] for r in records)
for emotion, count in sorted(dist.items()):
    print(f"  {emotion}: {count}")

print("\nUploading features.pkl to HuggingFace...")
api.upload_file(
    path_or_fileobj="features.pkl",
    path_in_repo="features.pkl",
    repo_id="MrlolDev/voxtral-emotion-speech",
    repo_type="dataset",
)
print("✅ features.pkl uploaded to HuggingFace")

print("\nUploading README.md to HuggingFace...")
api.upload_file(
    path_or_fileobj="README.md",
    path_in_repo="README.md",
    repo_id="MrlolDev/voxtral-emotion-speech",
    repo_type="dataset",
)
print("✅ README.md uploaded to HuggingFace")