model.py: probar cargando datasets hf

model.py

@@ -1,104 +1,171 @@
-import torch
-import numpy as np
 import os
+import json
+import random
+import torch
+import torchaudio
+from torch.utils.data import Dataset, DataLoader
 from huggingface_hub import login, upload_folder
-from datasets import load_dataset, Audio
 from transformers.integrations import TensorBoardCallback
+from sklearn.metrics import accuracy_score, precision_recall_fscore_support
 from transformers import (
-    Wav2Vec2FeatureExtractor, AutoModelForAudioClassification, 
+    Wav2Vec2FeatureExtractor, HubertConfig, HubertForSequenceClassification,
     Trainer, TrainingArguments,
     EarlyStoppingCallback
 )
-import json
-# SE USA FLOAT32 EN EL MODELO ORIGINAL
+
 MODEL = "ntu-spml/distilhubert" # modelo base utilizado, para usar otro basta con cambiar esto
 FEATURE_EXTRACTOR = Wav2Vec2FeatureExtractor.from_pretrained(MODEL)
 seed = 123
 MAX_DURATION = 1.00
-SAMPLING_RATE = FEATURE_EXTRACTOR.sampling_rate # 16000 # antes estaba float16
-token = os.getenv('MODEL_REPO_ID')
+SAMPLING_RATE = FEATURE_EXTRACTOR.sampling_rate # 16000
+token = os.getenv("HF_TOKEN")
 config_file = "models_config.json"
 clasificador = "class"
 monitor = "mon"
+batch_size = 16
+
+class AudioDataset(Dataset):
+    def __init__(self, dataset_path, label2id):
+        self.dataset_path = dataset_path
+        self.label2id = label2id
+        self.file_paths = []
+        self.labels = []
+        for label_dir, label_id in self.label2id.items():
+            label_path = os.path.join(self.dataset_path, label_dir)
+            if os.path.isdir(label_path):
+                for file_name in os.listdir(label_path):
+                    audio_path = os.path.join(label_path, file_name)
+                    self.file_paths.append(audio_path)
+                    self.labels.append(label_id)
+
+    def __len__(self):
+        return len(self.file_paths)
+
+    def __getitem__(self, idx):
+        audio_path = self.file_paths[idx]
+        label = self.labels[idx]
+        input_values = self.preprocess_audio(audio_path)
+        return {
+            "input_values": input_values,
+            "labels": torch.tensor(label)
+        }
+
+    def preprocess_audio(self, audio_path):
+        waveform, sample_rate = torchaudio.load(
+            audio_path,
+            normalize=True, # Convierte a float32
+            # num_frames= # TODO: Probar para que no haga falta recortar los audios
+            )
+        if sample_rate != SAMPLING_RATE: # Resamplear si no es 16kHz
+            resampler = torchaudio.transforms.Resample(sample_rate, SAMPLING_RATE)
+            waveform = resampler(waveform)
+        if waveform.shape[0] > 1: # Si es stereo, convertir a mono
+            waveform = waveform.mean(dim=0)
+        waveform = waveform / (torch.max(torch.abs(waveform)) + 1e-6) # Sin 1e-6 el accuracy es pésimo!!
+        max_length = int(SAMPLING_RATE * MAX_DURATION)
+        if waveform.shape[0] > max_length:
+            waveform = waveform[:max_length]
+        else:
+            waveform = torch.nn.functional.pad(waveform, (0, max_length - waveform.shape[0]))
+        inputs = FEATURE_EXTRACTOR(
+            waveform,
+            sampling_rate=SAMPLING_RATE,
+            return_tensors="pt",
+            # max_length=int(SAMPLING_RATE * MAX_DURATION),
+            # truncation=True,
+            padding=True,
+        )
+        return inputs.input_values.squeeze()
 
 def seed_everything():
-    np.random.seed(seed)
     torch.manual_seed(seed)
     torch.cuda.manual_seed(seed)
     torch.backends.cudnn.deterministic = True
     torch.backends.cudnn.benchmark = False
     os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':16384:8'
 
-def preprocess_audio(audio_arrays, batch=True):
-    if batch:
-        audios = [x["array"] for x in audio_arrays["audio"]] # para usar aquí
-    else:
-        audios = [audio_arrays] # para usar en realtime.py
-    inputs = Wav2Vec2FeatureExtractor.from_pretrained(MODEL)(
-        raw_speech=audios,
-        sampling_rate=SAMPLING_RATE,
-        return_tensors="pt", # Devolver tensores de PyTorch
-        max_length=int(SAMPLING_RATE * MAX_DURATION), # Necesario para truncation
-        truncation=True, # Muchísimo más rápido. 
-        padding=True, # Vectores igual longitud
-        do_normalize=True, # No afecta 1ª época, no sé si necesario
-        # return_attention_mask=True, # Empeora 1ª época. No sé si necesario
-        padding_value=0.0, # No afecta 1ª época, no sé si necesario
-        float=32 # No afecta 1ª época, no sé si necesario
-        )
-    return inputs
+def build_label_mappings(dataset_path):
+    label2id = {}
+    id2label = {}
+    label_id = 0
+    for label_dir in os.listdir(dataset_path):
+        if os.path.isdir(os.path.join(dataset_path, label_dir)):
+            label2id[label_dir] = label_id
+            id2label[label_id] = label_dir
+            label_id += 1
+    return label2id, id2label
 
-def load_and_prepare_dataset(dataset_path):
-    dataset = load_dataset(dataset_path, split="train") # Split para que no ponga train de primeras
-    # dataset = dataset.cast_column("audio", Audio(sampling_rate=SAMPLING_RATE)) # Da mejor accuracy pero creo que cambia el preprocesado.
-    encoded_dataset = dataset.map(preprocess_audio, remove_columns=["audio"], batched=True) # num_proc hace q no vaya realtime
-    labels = encoded_dataset.features["label"].names
-    label2id = {label: str(i) for i, label in enumerate(labels)}
-    id2label = {str(i): label for i, label in enumerate(labels)}
-    encoded_dataset = encoded_dataset.train_test_split(test_size=0.2, seed=seed, stratify_by_column="label")    
-    return encoded_dataset, label2id, id2label
+def create_dataloader(dataset_path, test_size=0.2, num_workers=12, shuffle=True, pin_memory=True):
+    label2id, id2label = build_label_mappings(dataset_path)
+    dataset = AudioDataset(dataset_path, label2id)
+    dataset_size = len(dataset)
+    indices = list(range(dataset_size))
+    random.shuffle(indices)
+    split_idx = int(dataset_size * (1 - test_size))
+    train_indices = indices[:split_idx]
+    test_indices = indices[split_idx:]
+    train_dataset = torch.utils.data.Subset(dataset, train_indices)
+    test_dataset = torch.utils.data.Subset(dataset, test_indices)
+    train_dataloader = DataLoader(
+        train_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers, pin_memory=pin_memory
+    )
+    test_dataloader = DataLoader(
+        test_dataset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers, pin_memory=pin_memory
+    )
+    return train_dataloader, test_dataloader, label2id, id2label
 
 def load_model(num_labels, label2id, id2label):
-    model = AutoModelForAudioClassification.from_pretrained(
+    config = HubertConfig.from_pretrained(
         MODEL,
         num_labels=num_labels,
         label2id=label2id,
-        id2label=id2label
+        id2label=id2label,
+        finetuning_task="audio-classification"
+    )
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model = HubertForSequenceClassification.from_pretrained( # TODO: mirar parámetros. Posibles optimizaciones
+        MODEL,
+        config=config,
+        torch_dtype=torch.float32, # No afecta 1ª época, mejor ponerlo
     )
+    model.to(device)
     return model
 
 def model_params(dataset_path):
-    login(token, add_to_git_credential=True)
-    seed_everything()
-    encoded_dataset, label2id, id2label = load_and_prepare_dataset(dataset_path)
-    model = load_model(len(id2label), label2id, id2label)
-    return model, encoded_dataset, id2label
+    train_dataloader, test_dataloader, label2id, id2label = create_dataloader(dataset_path)
+    model = load_model(num_labels=len(id2label), label2id=label2id, id2label=id2label)    
+    return model, train_dataloader, test_dataloader, id2label
 
 def compute_metrics(eval_pred):
-    predictions = np.argmax(eval_pred.predictions, axis=1)
-    references = eval_pred.label_ids
+    predictions = torch.argmax(torch.tensor(eval_pred.predictions), dim=-1)
+    references = torch.tensor(eval_pred.label_ids)
+    accuracy = accuracy_score(references, predictions)
+    precision, recall, f1, _ = precision_recall_fscore_support(references, predictions, average='weighted')
     return {
-        "accuracy": np.mean(predictions == references),
+        "accuracy": accuracy,
+        "precision": precision,
+        "recall": recall,
+        "f1": f1,
     }
 
-def model_training(training_args, output_dir, dataset_path):
-    model, encoded_dataset, _ = model_params(dataset_path)
-    tensorboard_callback = TensorBoardCallback()
-    early_stopping_callback = EarlyStoppingCallback(early_stopping_patience=3)
+def main(training_args, output_dir, dataset_path):
+    seed_everything()
+    model, train_dataloader, test_dataloader, _ = model_params(dataset_path)
     trainer = Trainer(
         model=model,
         args=training_args,
         compute_metrics=compute_metrics,
-        train_dataset=encoded_dataset["train"],
-        eval_dataset=encoded_dataset["test"],
-        callbacks=[tensorboard_callback, early_stopping_callback]
+        train_dataset=train_dataloader.dataset,
+        eval_dataset=test_dataloader.dataset,
+        callbacks=[TensorBoardCallback(), EarlyStoppingCallback(early_stopping_patience=3)]
     )
     torch.cuda.empty_cache() # liberar memoria de la GPU
     trainer.train() # se pueden modificar los parámetros para continuar el train
+    login(token, add_to_git_credential=True)
     trainer.push_to_hub(token=token) # Subir modelo a mi cuenta. Necesario para hacer la predicción, no sé por qué.
     trainer.save_model(output_dir) # para subir el modelo a Hugging Face. Necesario para hacer la predicción, no sé por qué.
     os.makedirs(output_dir, exist_ok=True) # Crear carpeta con el modelo si no existe
-    upload_folder(repo_id=f"A-POR-LOS-8000/{output_dir}",folder_path=output_dir, token=token) # subir modelo a organización
+    # upload_folder(repo_id=f"A-POR-LOS-8000/{output_dir}",folder_path=output_dir, token=token) # subir modelo a organización
 
 def load_config(model_name):
     with open(config_file, 'r') as f:
@@ -109,9 +176,9 @@ def load_config(model_name):
     return model_config
 
 if __name__ == "__main__":
-    config = load_config(clasificador) # PARA CAMBIAR MODELOS
-    # config = load_config(monitor) # PARA CAMBI
+    # config = load_config(clasificador) # PARA CAMBIAR MODELOS
+    config = load_config(monitor) # PARA CAMBIAR MODELOS
     training_args = config["training_args"]
     output_dir = config["output_dir"]
     dataset_path = config["dataset_path"]
-    model_training(training_args, output_dir, dataset_path)
+    main(training_args, output_dir, dataset_path)