Spaces:
Sleeping
Sleeping
File size: 5,771 Bytes
374c1c3 f9bbbb3 374c1c3 f9bbbb3 374c1c3 |
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 |
import os
import torch
import torch.nn as nn
import torch.optim as optim
import soundfile as sf
import torchaudio
import numpy as np
from datasets import Dataset
from transformers import Wav2Vec2Model, Wav2Vec2Processor
from dotenv import load_dotenv
from sklearn.metrics import accuracy_score
# Charger .env pour Hugging Face API Key
load_dotenv()
HF_API_KEY = os.getenv("HF_API_KEY")
if not HF_API_KEY:
raise ValueError("Le token Hugging Face n'a pas été trouvé dans .env")
# Définition des labels pour la classification des émotions
LABELS = {"colere": 0, "neutre": 1, "joie": 2}
NUM_LABELS = len(LABELS)
# Charger le processeur et le modèle pour l'extraction de features
model_name = "facebook/wav2vec2-large-xlsr-53-french"
device = "cuda" if torch.cuda.is_available() else "cpu"
processor = Wav2Vec2Processor.from_pretrained(model_name)
feature_extractor = Wav2Vec2Model.from_pretrained(model_name).to(device)
# Resampleur pour convertir en 16 kHz
resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
# Définition du classifieur amélioré
class EmotionClassifier(nn.Module):
def __init__(self, feature_dim, num_labels):
super(EmotionClassifier, self).__init__()
self.fc1 = nn.Linear(feature_dim, 512)
self.relu = nn.ReLU()
self.dropout = nn.Dropout(0.3)
self.fc2 = nn.Linear(512, num_labels)
def forward(self, x):
x = self.fc1(x)
x = self.relu(x)
x = self.dropout(x)
return self.fc2(x)
# Instancier le classifieur
classifier = EmotionClassifier(feature_extractor.config.hidden_size, NUM_LABELS).to(device)
# Charger les fichiers audio et leurs labels
def load_audio_data(data_dir):
data = []
for label_name, label_id in LABELS.items():
label_dir = os.path.join(data_dir, label_name)
for file in os.listdir(label_dir):
if file.endswith(".wav"):
file_path = os.path.join(label_dir, file)
data.append({"path": file_path, "label": label_id})
return Dataset.from_list(data)
# Chargement du dataset
data_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "data"))
ds = load_audio_data(data_dir)
# Charger les fichiers audio avec SoundFile et rééchantillonner à 16 kHz
def preprocess_audio(batch):
speech, sample_rate = sf.read(batch["path"], dtype="float32")
if sample_rate != 16000:
speech = torch.tensor(speech).unsqueeze(0)
speech = resampler(speech).squeeze(0).numpy()
batch["speech"] = speech.tolist() # Convertir en liste pour éviter les erreurs de PyArrow
batch["sampling_rate"] = 16000
return batch
ds = ds.map(preprocess_audio)
# Vérifier la distribution des longueurs des fichiers audio
lengths = [len(sample["speech"]) for sample in ds]
max_length = int(np.percentile(lengths, 95))
# Transformer l'audio en features utilisables par le modèle
def prepare_features(batch):
features = processor(
batch["speech"],
sampling_rate=16000,
padding=True,
truncation=True,
max_length=max_length,
return_tensors="pt"
)
batch["input_values"] = features.input_values.squeeze(0)
batch["label"] = torch.tensor(batch["label"], dtype=torch.long)
return batch
ds = ds.map(prepare_features)
# Diviser les données en train et test
ds = ds.train_test_split(test_size=0.2)
train_ds = ds["train"]
test_ds = ds["test"]
# Fonction d'entraînement avec sauvegarde du meilleur modèle
def train_classifier(feature_extractor, classifier, train_ds, test_ds, epochs=20, batch_size=8):
optimizer = optim.AdamW(classifier.parameters(), lr=2e-5, weight_decay=0.01)
loss_fn = nn.CrossEntropyLoss()
best_accuracy = 0.0 # Variable pour stocker la meilleure accuracy
for epoch in range(epochs):
classifier.train()
total_loss, correct = 0, 0
batch_count = 0
for i in range(0, len(train_ds), batch_size):
batch = train_ds[i: i + batch_size]
optimizer.zero_grad()
input_values = processor(
batch["speech"],
sampling_rate=16000,
return_tensors="pt",
padding=True,
truncation=True,
max_length=max_length
).input_values.to(device)
with torch.no_grad():
features = feature_extractor(input_values).last_hidden_state.mean(dim=1)
logits = classifier(features)
labels = torch.tensor(batch["label"], dtype=torch.long, device=device)
if labels.numel() == 0:
continue
loss = loss_fn(logits, labels)
loss.backward()
optimizer.step()
total_loss += loss.item()
correct += (logits.argmax(dim=-1) == labels).sum().item()
batch_count += 1
train_acc = correct / len(train_ds)
# Sauvegarde du modèle seulement si la précision s'améliore
if train_acc > best_accuracy:
best_accuracy = train_acc
torch.save({
"classifier_state_dict": classifier.state_dict(),
"feature_extractor_state_dict": feature_extractor.state_dict(),
"processor": processor
}, "acc_model.pth")
print(f"✅ Nouveau meilleur modèle sauvegardé ! Accuracy: {best_accuracy:.4f}")
print(f"Epoch {epoch+1}/{epochs} - Loss: {total_loss/batch_count:.4f} - Accuracy: {train_acc:.4f}")
return classifier
# Entraînement
trained_classifier = train_classifier(feature_extractor, classifier, train_ds, test_ds, epochs=20, batch_size=8)
print("✅ Entraînement terminé, le meilleur modèle a été sauvegardé !")
|