class EncDecSpeakerLabelModel(ModelPT, ExportableEncDecModel): """Encoder decoder class for speaker label models. Model class creates training, validation methods for setting up data performing model forward pass. Expects config dict for * preprocessor * Jasper/Quartznet Encoder * Speaker Decoder """ @classmethod def list_available_models(cls) -> List[PretrainedModelInfo]: """ This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. Returns: List of available pre-trained models. """ result = [] model = PretrainedModelInfo( pretrained_model_name="speakerverification_speakernet", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/speakerverification_speakernet/versions/1.0.0rc1/files/speakerverification_speakernet.nemo", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:speakerverification_speakernet", ) result.append(model) model = PretrainedModelInfo( pretrained_model_name="ecapa_tdnn", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/ecapa_tdnn/versions/v1/files/ecapa_tdnn.nemo", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:ecapa_tdnn", ) result.append(model) model = PretrainedModelInfo( pretrained_model_name="titanet_large", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/titanet_large/versions/v0/files/titanet-l.nemo", description= "For details about this model, please visit https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_large", ) result.append(model) return result def __init__(self, cfg: DictConfig, trainer: Trainer = None): self.world_size = 1 if trainer is not None: self.world_size = trainer.num_nodes * trainer.num_gpus super().__init__(cfg=cfg, trainer=trainer) self.preprocessor = EncDecSpeakerLabelModel.from_config_dict( cfg.preprocessor) self.encoder = EncDecSpeakerLabelModel.from_config_dict(cfg.encoder) self.decoder = EncDecSpeakerLabelModel.from_config_dict(cfg.decoder) if 'angular' in cfg.decoder and cfg.decoder['angular']: logging.info("Training with Angular Softmax Loss") scale = cfg.loss.scale margin = cfg.loss.margin self.loss = AngularSoftmaxLoss(scale=scale, margin=margin) else: logging.info("Training with Softmax-CrossEntropy loss") self.loss = CELoss() self.task = None self._accuracy = TopKClassificationAccuracy(top_k=[1]) self.labels = None @staticmethod def extract_labels(data_layer_config): labels = set() manifest_filepath = data_layer_config.get('manifest_filepath', None) if manifest_filepath is None: logging.warning( "No manifest_filepath was provided, no labels got extracted!") return None manifest_filepaths = convert_to_config_list( data_layer_config['manifest_filepath']) for manifest_filepath in itertools.chain.from_iterable( manifest_filepaths): collection = ASRSpeechLabel( manifests_files=manifest_filepath, min_duration=data_layer_config.get("min_duration", None), max_duration=data_layer_config.get("max_duration", None), index_by_file_id=False, ) labels.update(collection.uniq_labels) labels = list(sorted(labels)) logging.warning( f"Total number of {len(labels)} found in all the manifest files.") return labels def __setup_dataloader_from_config(self, config: Optional[Dict]): if 'augmentor' in config: augmentor = process_augmentations(config['augmentor']) else: augmentor = None featurizer = WaveformFeaturizer(sample_rate=config['sample_rate'], int_values=config.get( 'int_values', False), augmentor=augmentor) shuffle = config.get('shuffle', False) if config.get('is_tarred', False): if ('tarred_audio_filepaths' in config and config['tarred_audio_filepaths'] is None) or ( 'manifest_filepath' in config and config['manifest_filepath'] is None): logging.warning( "Could not load dataset as `manifest_filepath` was None or " f"`tarred_audio_filepaths` is None. Provided config : {config}" ) return None shuffle_n = config.get('shuffle_n', 4 * config['batch_size']) if shuffle else 0 dataset = get_tarred_speech_label_dataset( featurizer=featurizer, config=config, shuffle_n=shuffle_n, global_rank=self.global_rank, world_size=self.world_size, ) shuffle = False else: if 'manifest_filepath' in config and config[ 'manifest_filepath'] is None: logging.warning( f"Could not load dataset as `manifest_filepath` was None. Provided config : {config}" ) return None dataset = AudioToSpeechLabelDataset( manifest_filepath=config['manifest_filepath'], labels=config['labels'], featurizer=featurizer, max_duration=config.get('max_duration', None), min_duration=config.get('min_duration', None), trim=config.get('trim_silence', False), time_length=config.get('time_length', 8), shift_length=config.get('shift_length', 0.75), normalize_audio=config.get('normalize_audio', False), ) if type(dataset) is ChainDataset: collate_ds = dataset.datasets[0] else: collate_ds = dataset # self.labels = collate_ds.labels if self.task == 'diarization': logging.info("Setting up diarization parameters") collate_fn = collate_ds.sliced_seq_collate_fn shuffle = False else: logging.info("Setting up identification parameters") collate_fn = collate_ds.fixed_seq_collate_fn batch_size = config['batch_size'] return torch.utils.data.DataLoader( dataset=dataset, batch_size=batch_size, collate_fn=collate_fn, drop_last=config.get('drop_last', False), shuffle=shuffle, num_workers=config.get('num_workers', 0), pin_memory=config.get('pin_memory', False), ) def setup_training_data(self, train_data_layer_config: Optional[Union[DictConfig, Dict]]): self.labels = self.extract_labels(train_data_layer_config) train_data_layer_config['labels'] = self.labels if 'shuffle' not in train_data_layer_config: train_data_layer_config['shuffle'] = True self.task = 'identification' self._train_dl = self.__setup_dataloader_from_config( config=train_data_layer_config) def setup_validation_data(self, val_data_layer_config: Optional[Union[DictConfig, Dict]]): val_data_layer_config['labels'] = self.labels self.task = 'identification' self._validation_dl = self.__setup_dataloader_from_config( config=val_data_layer_config) def setup_test_data(self, test_data_layer_params: Optional[Union[DictConfig, Dict]]): if hasattr(self, 'dataset'): test_data_layer_params['labels'] = self.labels if 'task' in test_data_layer_params and test_data_layer_params['task']: self.task = test_data_layer_params['task'].lower() self.time_length = test_data_layer_params.get('time_length', 1.5) self.shift_length = test_data_layer_params.get( 'shift_length', 0.75) else: self.task = 'identification' self.embedding_dir = test_data_layer_params.get('embedding_dir', './') self._test_dl = self.__setup_dataloader_from_config( config=test_data_layer_params) self.test_manifest = test_data_layer_params.get( 'manifest_filepath', None) def test_dataloader(self): if self._test_dl is not None: return self._test_dl @property def input_types(self) -> Optional[Dict[str, NeuralType]]: if hasattr(self.preprocessor, '_sample_rate'): audio_eltype = AudioSignal(freq=self.preprocessor._sample_rate) else: audio_eltype = AudioSignal() return { "input_signal": NeuralType(('B', 'T'), audio_eltype), "input_signal_length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self) -> Optional[Dict[str, NeuralType]]: return { "logits": NeuralType(('B', 'D'), LogitsType()), "embs": NeuralType(('B', 'D'), AcousticEncodedRepresentation()), } @typecheck() def forward(self, input_signal, input_signal_length): processed_signal, processed_signal_len = self.preprocessor( input_signal=input_signal, length=input_signal_length, ) encoded, length = self.encoder(audio_signal=processed_signal, length=processed_signal_len) logits, embs = self.decoder(encoder_output=encoded, length=length) return logits, embs # PTL-specific methods def training_step(self, batch, batch_idx): audio_signal, audio_signal_len, labels, _ = batch logits, _ = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) loss = self.loss(logits=logits, labels=labels) self.log('loss', loss) self.log('learning_rate', self._optimizer.param_groups[0]['lr']) self._accuracy(logits=logits, labels=labels) top_k = self._accuracy.compute() self._accuracy.reset() for i, top_i in enumerate(top_k): self.log(f'training_batch_accuracy_top@{i}', top_i) return {'loss': loss} def validation_step(self, batch, batch_idx, dataloader_idx: int = 0): audio_signal, audio_signal_len, labels, _ = batch logits, _ = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) loss_value = self.loss(logits=logits, labels=labels) acc_top_k = self._accuracy(logits=logits, labels=labels) correct_counts, total_counts = self._accuracy.correct_counts_k, self._accuracy.total_counts_k return { 'val_loss': loss_value, 'val_correct_counts': correct_counts, 'val_total_counts': total_counts, 'val_acc_top_k': acc_top_k, } def multi_validation_epoch_end(self, outputs, dataloader_idx: int = 0): val_loss_mean = torch.stack([x['val_loss'] for x in outputs]).mean() correct_counts = torch.stack( [x['val_correct_counts'] for x in outputs]).sum(axis=0) total_counts = torch.stack([x['val_total_counts'] for x in outputs]).sum(axis=0) self._accuracy.correct_counts_k = correct_counts self._accuracy.total_counts_k = total_counts topk_scores = self._accuracy.compute() self._accuracy.reset() logging.info("val_loss: {:.3f}".format(val_loss_mean)) self.log('val_loss', val_loss_mean) for top_k, score in zip(self._accuracy.top_k, topk_scores): self.log('val_epoch_accuracy_top@{}'.format(top_k), score) return { 'val_loss': val_loss_mean, 'val_acc_top_k': topk_scores, } def test_step(self, batch, batch_idx, dataloader_idx: int = 0): audio_signal, audio_signal_len, labels, _ = batch logits, _ = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) loss_value = self.loss(logits=logits, labels=labels) acc_top_k = self._accuracy(logits=logits, labels=labels) correct_counts, total_counts = self._accuracy.correct_counts_k, self._accuracy.total_counts_k return { 'test_loss': loss_value, 'test_correct_counts': correct_counts, 'test_total_counts': total_counts, 'test_acc_top_k': acc_top_k, } def multi_test_epoch_end(self, outputs, dataloader_idx: int = 0): test_loss_mean = torch.stack([x['test_loss'] for x in outputs]).mean() correct_counts = torch.stack( [x['test_correct_counts'] for x in outputs]).sum(axis=0) total_counts = torch.stack([x['test_total_counts'] for x in outputs]).sum(axis=0) self._accuracy.correct_counts_k = correct_counts self._accuracy.total_counts_k = total_counts topk_scores = self._accuracy.compute() self._accuracy.reset() logging.info("test_loss: {:.3f}".format(test_loss_mean)) self.log('test_loss', test_loss_mean) for top_k, score in zip(self._accuracy.top_k, topk_scores): self.log('test_epoch_accuracy_top@{}'.format(top_k), score) return { 'test_loss': test_loss_mean, 'test_acc_top_k': topk_scores, } def setup_finetune_model(self, model_config: DictConfig): """ setup_finetune_model method sets up training data, validation data and test data with new provided config, this checks for the previous labels set up during training from scratch, if None, it sets up labels for provided finetune data from manifest files Args: model_config: cfg which has train_ds, optional validation_ds, optional test_ds and mandatory encoder and decoder model params make sure you set num_classes correctly for finetune data Returns: None """ logging.info( "Setting up data loaders with manifests provided from model_config" ) if 'train_ds' in model_config and model_config.train_ds is not None: self.setup_training_data(model_config.train_ds) else: raise KeyError( "train_ds is not found in model_config but you need it for fine tuning" ) if self.labels is None or len(self.labels) == 0: raise ValueError( f'New labels must be non-empty list of labels. But I got: {self.labels}' ) if 'validation_ds' in model_config and model_config.validation_ds is not None: self.setup_multiple_validation_data(model_config.validation_ds) if 'test_ds' in model_config and model_config.test_ds is not None: self.setup_multiple_test_data(model_config.test_ds) if self.labels is not None: # checking for new finetune dataset labels logging.warning( "Trained dataset labels are same as finetune dataset labels -- continuing change of decoder parameters" ) else: logging.warning( "Either you provided a dummy manifest file during training from scratch or you restored from a pretrained nemo file" ) decoder_config = model_config.decoder new_decoder_config = copy.deepcopy(decoder_config) if new_decoder_config['num_classes'] != len(self.labels): raise ValueError( "number of classes provided {} is not same as number of different labels in finetuning data: {}" .format(new_decoder_config['num_classes'], len(self.labels))) del self.decoder self.decoder = EncDecSpeakerLabelModel.from_config_dict( new_decoder_config) with open_dict(self._cfg.decoder): self._cfg.decoder = new_decoder_config logging.info( f"Changed decoder output to # {self.decoder._num_classes} classes." ) @torch.no_grad() def get_embedding(self, path2audio_file): audio, sr = librosa.load(path2audio_file, sr=None) target_sr = self._cfg.train_ds.get('sample_rate', 16000) if sr != target_sr: audio = librosa.core.resample(audio, sr, target_sr) audio_length = audio.shape[0] device = self.device audio_signal, audio_signal_len = ( torch.tensor([audio], device=device), torch.tensor([audio_length], device=device), ) mode = self.training self.freeze() _, embs = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) self.train(mode=mode) if mode is True: self.unfreeze() del audio_signal, audio_signal_len return embs
class EncDecClassificationModel(_EncDecBaseModel): """Encoder decoder Classification models.""" def __init__(self, cfg: DictConfig, trainer: Trainer = None): if cfg.get("is_regression_task", False): raise ValueError( f"EndDecClassificationModel requires the flag is_regression_task to be set as false" ) super().__init__(cfg=cfg, trainer=trainer) def _setup_preprocessor(self): return EncDecClassificationModel.from_config_dict( self._cfg.preprocessor) def _setup_encoder(self): return EncDecClassificationModel.from_config_dict(self._cfg.encoder) def _setup_decoder(self): return EncDecClassificationModel.from_config_dict(self._cfg.decoder) def _setup_loss(self): return CrossEntropyLoss() def _setup_metrics(self): self._accuracy = TopKClassificationAccuracy(dist_sync_on_step=True) @classmethod def list_available_models(cls) -> Optional[List[PretrainedModelInfo]]: """ This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. Returns: List of available pre-trained models. """ results = [] model = PretrainedModelInfo( pretrained_model_name="vad_telephony_marblenet", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:vad_telephony_marblenet", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/vad_telephony_marblenet/versions/1.0.0rc1/files/vad_telephony_marblenet.nemo", ) results.append(model) model = PretrainedModelInfo( pretrained_model_name="vad_marblenet", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:vad_marblenet", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/vad_marblenet/versions/1.0.0rc1/files/vad_marblenet.nemo", ) results.append(model) model = PretrainedModelInfo( pretrained_model_name="commandrecognition_en_matchboxnet3x1x64_v1", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:commandrecognition_en_matchboxnet3x1x64_v1", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/commandrecognition_en_matchboxnet3x1x64_v1/versions/1.0.0rc1/files/commandrecognition_en_matchboxnet3x1x64_v1.nemo", ) results.append(model) model = PretrainedModelInfo( pretrained_model_name="commandrecognition_en_matchboxnet3x2x64_v1", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:commandrecognition_en_matchboxnet3x2x64_v1", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/commandrecognition_en_matchboxnet3x2x64_v1/versions/1.0.0rc1/files/commandrecognition_en_matchboxnet3x2x64_v1.nemo", ) results.append(model) model = PretrainedModelInfo( pretrained_model_name="commandrecognition_en_matchboxnet3x1x64_v2", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:commandrecognition_en_matchboxnet3x1x64_v2", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/commandrecognition_en_matchboxnet3x1x64_v2/versions/1.0.0rc1/files/commandrecognition_en_matchboxnet3x1x64_v2.nemo", ) results.append(model) model = PretrainedModelInfo( pretrained_model_name="commandrecognition_en_matchboxnet3x2x64_v2", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:commandrecognition_en_matchboxnet3x2x64_v2", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/commandrecognition_en_matchboxnet3x2x64_v2/versions/1.0.0rc1/files/commandrecognition_en_matchboxnet3x2x64_v2.nemo", ) results.append(model) model = PretrainedModelInfo( pretrained_model_name= "commandrecognition_en_matchboxnet3x1x64_v2_subset_task", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:commandrecognition_en_matchboxnet3x1x64_v2_subset_task", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/commandrecognition_en_matchboxnet3x1x64_v2_subset_task/versions/1.0.0rc1/files/commandrecognition_en_matchboxnet3x1x64_v2_subset_task.nemo", ) results.append(model) model = PretrainedModelInfo( pretrained_model_name= "commandrecognition_en_matchboxnet3x2x64_v2_subset_task", description= "For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:commandrecognition_en_matchboxnet3x2x64_v2_subset_task", location= "https://api.ngc.nvidia.com/v2/models/nvidia/nemo/commandrecognition_en_matchboxnet3x2x64_v2_subset_task/versions/1.0.0rc1/files/commandrecognition_en_matchboxnet3x2x64_v2_subset_task.nemo", ) results.append(model) return results @property def output_types(self) -> Optional[Dict[str, NeuralType]]: return {"outputs": NeuralType(('B', 'D'), LogitsType())} # PTL-specific methods def training_step(self, batch, batch_nb): self.training_step_end() audio_signal, audio_signal_len, labels, labels_len = batch logits = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) loss_value = self.loss(logits=logits, labels=labels) self.log('train_loss', loss_value) self.log('learning_rate', self._optimizer.param_groups[0]['lr']) self._accuracy(logits=logits, labels=labels) topk_scores = self._accuracy.compute() self._accuracy.reset() for top_k, score in zip(self._accuracy.top_k, topk_scores): self.log('training_batch_accuracy_top@{}'.format(top_k), score) return { 'loss': loss_value, } def validation_step(self, batch, batch_idx, dataloader_idx=0): audio_signal, audio_signal_len, labels, labels_len = batch logits = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) loss_value = self.loss(logits=logits, labels=labels) acc = self._accuracy(logits=logits, labels=labels) correct_counts, total_counts = self._accuracy.correct_counts_k, self._accuracy.total_counts_k return { 'val_loss': loss_value, 'val_correct_counts': correct_counts, 'val_total_counts': total_counts, 'val_acc': acc, } def test_step(self, batch, batch_idx, dataloader_idx=0): audio_signal, audio_signal_len, labels, labels_len = batch logits = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) loss_value = self.loss(logits=logits, labels=labels) acc = self._accuracy(logits=logits, labels=labels) correct_counts, total_counts = self._accuracy.correct_counts_k, self._accuracy.total_counts_k return { 'test_loss': loss_value, 'test_correct_counts': correct_counts, 'test_total_counts': total_counts, 'test_acc': acc, } def multi_validation_epoch_end(self, outputs, dataloader_idx: int = 0): val_loss_mean = torch.stack([x['val_loss'] for x in outputs]).mean() correct_counts = torch.stack( [x['val_correct_counts'] for x in outputs]).sum(axis=0) total_counts = torch.stack([x['val_total_counts'] for x in outputs]).sum(axis=0) self._accuracy.correct_counts_k = correct_counts self._accuracy.total_counts_k = total_counts topk_scores = self._accuracy.compute() self._accuracy.reset() tensorboard_log = {'val_loss': val_loss_mean} for top_k, score in zip(self._accuracy.top_k, topk_scores): tensorboard_log['val_epoch_top@{}'.format(top_k)] = score return {'log': tensorboard_log} def multi_test_epoch_end(self, outputs, dataloader_idx: int = 0): test_loss_mean = torch.stack([x['test_loss'] for x in outputs]).mean() correct_counts = torch.stack([ x['test_correct_counts'].unsqueeze(0) for x in outputs ]).sum(axis=0) total_counts = torch.stack( [x['test_total_counts'].unsqueeze(0) for x in outputs]).sum(axis=0) self._accuracy.correct_counts_k = correct_counts self._accuracy.total_counts_k = total_counts topk_scores = self._accuracy.compute() self._accuracy.reset() tensorboard_log = {'test_loss': test_loss_mean} for top_k, score in zip(self._accuracy.top_k, topk_scores): tensorboard_log['test_epoch_top@{}'.format(top_k)] = score return {'log': tensorboard_log} @typecheck() def forward(self, input_signal, input_signal_length): logits = super().forward(input_signal=input_signal, input_signal_length=input_signal_length) return logits def change_labels(self, new_labels: List[str]): """ Changes labels used by the decoder model. Use this method when fine-tuning on from pre-trained model. This method changes only decoder and leaves encoder and pre-processing modules unchanged. For example, you would use it if you want to use pretrained encoder when fine-tuning on a data in another dataset. If new_labels == self.decoder.vocabulary then nothing will be changed. Args: new_labels: list with new labels. Must contain at least 2 elements. Typically, \ this is set of labels for the dataset. Returns: None """ if new_labels is not None and not isinstance(new_labels, ListConfig): new_labels = ListConfig(new_labels) if self._cfg.labels == new_labels: logging.warning( f"Old labels ({self._cfg.labels}) and new labels ({new_labels}) match. Not changing anything" ) else: if new_labels is None or len(new_labels) == 0: raise ValueError( f'New labels must be non-empty list of labels. But I got: {new_labels}' ) # Update config self._cfg.labels = new_labels decoder_config = self.decoder.to_config_dict() new_decoder_config = copy.deepcopy(decoder_config) self._update_decoder_config(new_labels, new_decoder_config) del self.decoder self.decoder = EncDecClassificationModel.from_config_dict( new_decoder_config) OmegaConf.set_struct(self._cfg.decoder, False) self._cfg.decoder = new_decoder_config OmegaConf.set_struct(self._cfg.decoder, True) if 'train_ds' in self._cfg and self._cfg.train_ds is not None: self._cfg.train_ds.labels = new_labels if 'validation_ds' in self._cfg and self._cfg.validation_ds is not None: self._cfg.validation_ds.labels = new_labels if 'test_ds' in self._cfg and self._cfg.test_ds is not None: self._cfg.test_ds.labels = new_labels logging.info( f"Changed decoder output to {self.decoder.num_classes} labels." ) def _update_decoder_config(self, labels, cfg): """ Update the number of classes in the decoder based on labels provided. Args: labels: The current labels of the model cfg: The config of the decoder which will be updated. """ OmegaConf.set_struct(cfg, False) if 'params' in cfg: cfg.params.num_classes = len(labels) else: cfg.num_classes = len(labels) OmegaConf.set_struct(cfg, True)
class EncDecSpeakerLabelModel(ModelPT, ExportableEncDecModel): """ Encoder decoder class for speaker label models. Model class creates training, validation methods for setting up data performing model forward pass. Expects config dict for * preprocessor * Jasper/Quartznet Encoder * Speaker Decoder """ @classmethod def list_available_models(cls) -> List[PretrainedModelInfo]: """ This method returns a list of pre-trained model which can be instantiated directly from NVIDIA's NGC cloud. Returns: List of available pre-trained models. """ result = [] model = PretrainedModelInfo( pretrained_model_name="speakerverification_speakernet", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/speakerverification_speakernet/versions/1.0.0rc1/files/speakerverification_speakernet.nemo", description="For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:speakerverification_speakernet", ) result.append(model) model = PretrainedModelInfo( pretrained_model_name="ecapa_tdnn", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/ecapa_tdnn/versions/v1/files/ecapa_tdnn.nemo", description="For details about this model, please visit https://ngc.nvidia.com/catalog/models/nvidia:nemo:ecapa_tdnn", ) result.append(model) model = PretrainedModelInfo( pretrained_model_name="titanet_large", location="https://api.ngc.nvidia.com/v2/models/nvidia/nemo/titanet_large/versions/v0/files/titanet-l.nemo", description="For details about this model, please visit https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_large", ) result.append(model) return result def __init__(self, cfg: DictConfig, trainer: Trainer = None): self.world_size = 1 if trainer is not None: self.world_size = trainer.num_nodes * trainer.num_devices super().__init__(cfg=cfg, trainer=trainer) self.preprocessor = EncDecSpeakerLabelModel.from_config_dict(cfg.preprocessor) self.encoder = EncDecSpeakerLabelModel.from_config_dict(cfg.encoder) self.decoder = EncDecSpeakerLabelModel.from_config_dict(cfg.decoder) if 'angular' in cfg.decoder and cfg.decoder['angular']: logging.info("loss is Angular Softmax") scale = cfg.loss.scale margin = cfg.loss.margin self.loss = AngularSoftmaxLoss(scale=scale, margin=margin) else: logging.info("loss is Softmax-CrossEntropy") self.loss = CELoss() self.task = None self._accuracy = TopKClassificationAccuracy(top_k=[1]) self.labels = None if hasattr(self._cfg, 'spec_augment') and self._cfg.spec_augment is not None: self.spec_augmentation = EncDecSpeakerLabelModel.from_config_dict(self._cfg.spec_augment) else: self.spec_augmentation = None @staticmethod def extract_labels(data_layer_config): labels = set() manifest_filepath = data_layer_config.get('manifest_filepath', None) if manifest_filepath is None: logging.warning("No manifest_filepath was provided, no labels got extracted!") return None manifest_filepaths = convert_to_config_list(data_layer_config['manifest_filepath']) for manifest_filepath in itertools.chain.from_iterable(manifest_filepaths): collection = ASRSpeechLabel( manifests_files=manifest_filepath, min_duration=data_layer_config.get("min_duration", None), max_duration=data_layer_config.get("max_duration", None), index_by_file_id=True, ) labels.update(collection.uniq_labels) labels = list(sorted(labels)) logging.warning(f"Total number of {len(labels)} found in all the manifest files.") return labels def __setup_dataloader_from_config(self, config: Optional[Dict]): if 'augmentor' in config: augmentor = process_augmentations(config['augmentor']) else: augmentor = None featurizer = WaveformFeaturizer( sample_rate=config['sample_rate'], int_values=config.get('int_values', False), augmentor=augmentor ) shuffle = config.get('shuffle', False) if config.get('is_tarred', False): if ('tarred_audio_filepaths' in config and config['tarred_audio_filepaths'] is None) or ( 'manifest_filepath' in config and config['manifest_filepath'] is None ): logging.warning( "Could not load dataset as `manifest_filepath` was None or " f"`tarred_audio_filepaths` is None. Provided config : {config}" ) return None shuffle_n = config.get('shuffle_n', 4 * config['batch_size']) if shuffle else 0 dataset = get_tarred_speech_label_dataset( featurizer=featurizer, config=config, shuffle_n=shuffle_n, global_rank=self.global_rank, world_size=self.world_size, ) shuffle = False else: if 'manifest_filepath' in config and config['manifest_filepath'] is None: logging.warning(f"Could not load dataset as `manifest_filepath` was None. Provided config : {config}") return None dataset = AudioToSpeechLabelDataset( manifest_filepath=config['manifest_filepath'], labels=config['labels'], featurizer=featurizer, max_duration=config.get('max_duration', None), min_duration=config.get('min_duration', None), trim=config.get('trim_silence', False), normalize_audio=config.get('normalize_audio', False), ) if hasattr(dataset, 'fixed_seq_collate_fn'): collate_fn = dataset.fixed_seq_collate_fn else: collate_fn = dataset.datasets[0].fixed_seq_collate_fn batch_size = config['batch_size'] return torch.utils.data.DataLoader( dataset=dataset, batch_size=batch_size, collate_fn=collate_fn, drop_last=config.get('drop_last', False), shuffle=shuffle, num_workers=config.get('num_workers', 0), pin_memory=config.get('pin_memory', False), ) def setup_training_data(self, train_data_layer_config: Optional[Union[DictConfig, Dict]]): self.labels = self.extract_labels(train_data_layer_config) train_data_layer_config['labels'] = self.labels if 'shuffle' not in train_data_layer_config: train_data_layer_config['shuffle'] = True self._train_dl = self.__setup_dataloader_from_config(config=train_data_layer_config) # Need to set this because if using an IterableDataset, the length of the dataloader is the total number # of samples rather than the number of batches, and this messes up the tqdm progress bar. # So we set the number of steps manually (to the correct number) to fix this. if 'is_tarred' in train_data_layer_config and train_data_layer_config['is_tarred']: # We also need to check if limit_train_batches is already set. # If it's an int, we assume that the user has set it to something sane, i.e. <= # training batches, # and don't change it. Otherwise, adjust batches accordingly if it's a float (including 1.0). if self._trainer is not None and isinstance(self._trainer.limit_train_batches, float): self._trainer.limit_train_batches = int( self._trainer.limit_train_batches * ceil((len(self._train_dl.dataset) / self.world_size) / train_data_layer_config['batch_size']) ) elif self._trainer is None: logging.warning( "Model Trainer was not set before constructing the dataset, incorrect number of " "training batches will be used. Please set the trainer and rebuild the dataset." ) def setup_validation_data(self, val_data_layer_config: Optional[Union[DictConfig, Dict]]): val_data_layer_config['labels'] = self.labels self._validation_dl = self.__setup_dataloader_from_config(config=val_data_layer_config) def setup_test_data(self, test_data_layer_params: Optional[Union[DictConfig, Dict]]): if hasattr(self, 'dataset'): test_data_layer_params['labels'] = self.labels self.embedding_dir = test_data_layer_params.get('embedding_dir', './') self._test_dl = self.__setup_dataloader_from_config(config=test_data_layer_params) self.test_manifest = test_data_layer_params.get('manifest_filepath', None) def test_dataloader(self): if self._test_dl is not None: return self._test_dl @property def input_types(self) -> Optional[Dict[str, NeuralType]]: if hasattr(self.preprocessor, '_sample_rate'): audio_eltype = AudioSignal(freq=self.preprocessor._sample_rate) else: audio_eltype = AudioSignal() return { "input_signal": NeuralType(('B', 'T'), audio_eltype), "input_signal_length": NeuralType(tuple('B'), LengthsType()), } @property def output_types(self) -> Optional[Dict[str, NeuralType]]: return { "logits": NeuralType(('B', 'D'), LogitsType()), "embs": NeuralType(('B', 'D'), AcousticEncodedRepresentation()), } def forward_for_export(self, processed_signal, processed_signal_len): encoded, length = self.encoder(audio_signal=processed_signal, length=processed_signal_len) logits, embs = self.decoder(encoder_output=encoded, length=length) return logits, embs @typecheck() def forward(self, input_signal, input_signal_length): processed_signal, processed_signal_len = self.preprocessor( input_signal=input_signal, length=input_signal_length, ) if self.spec_augmentation is not None and self.training: processed_signal = self.spec_augmentation(input_spec=processed_signal, length=processed_signal_len) encoded, length = self.encoder(audio_signal=processed_signal, length=processed_signal_len) logits, embs = self.decoder(encoder_output=encoded, length=length) return logits, embs # PTL-specific methods def training_step(self, batch, batch_idx): audio_signal, audio_signal_len, labels, _ = batch logits, _ = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) loss = self.loss(logits=logits, labels=labels) self.log('loss', loss) self.log('learning_rate', self._optimizer.param_groups[0]['lr']) self._accuracy(logits=logits, labels=labels) top_k = self._accuracy.compute() self._accuracy.reset() for i, top_i in enumerate(top_k): self.log(f'training_batch_accuracy_top@{i}', top_i) return {'loss': loss} def validation_step(self, batch, batch_idx, dataloader_idx: int = 0): audio_signal, audio_signal_len, labels, _ = batch logits, _ = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) loss_value = self.loss(logits=logits, labels=labels) acc_top_k = self._accuracy(logits=logits, labels=labels) correct_counts, total_counts = self._accuracy.correct_counts_k, self._accuracy.total_counts_k return { 'val_loss': loss_value, 'val_correct_counts': correct_counts, 'val_total_counts': total_counts, 'val_acc_top_k': acc_top_k, } def multi_validation_epoch_end(self, outputs, dataloader_idx: int = 0): val_loss_mean = torch.stack([x['val_loss'] for x in outputs]).mean() correct_counts = torch.stack([x['val_correct_counts'] for x in outputs]).sum(axis=0) total_counts = torch.stack([x['val_total_counts'] for x in outputs]).sum(axis=0) self._accuracy.correct_counts_k = correct_counts self._accuracy.total_counts_k = total_counts topk_scores = self._accuracy.compute() self._accuracy.reset() logging.info("val_loss: {:.3f}".format(val_loss_mean)) self.log('val_loss', val_loss_mean) for top_k, score in zip(self._accuracy.top_k, topk_scores): self.log('val_epoch_accuracy_top@{}'.format(top_k), score) return { 'val_loss': val_loss_mean, 'val_acc_top_k': topk_scores, } def test_step(self, batch, batch_idx, dataloader_idx: int = 0): audio_signal, audio_signal_len, labels, _ = batch logits, _ = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) loss_value = self.loss(logits=logits, labels=labels) acc_top_k = self._accuracy(logits=logits, labels=labels) correct_counts, total_counts = self._accuracy.correct_counts_k, self._accuracy.total_counts_k return { 'test_loss': loss_value, 'test_correct_counts': correct_counts, 'test_total_counts': total_counts, 'test_acc_top_k': acc_top_k, } def multi_test_epoch_end(self, outputs, dataloader_idx: int = 0): test_loss_mean = torch.stack([x['test_loss'] for x in outputs]).mean() correct_counts = torch.stack([x['test_correct_counts'] for x in outputs]).sum(axis=0) total_counts = torch.stack([x['test_total_counts'] for x in outputs]).sum(axis=0) self._accuracy.correct_counts_k = correct_counts self._accuracy.total_counts_k = total_counts topk_scores = self._accuracy.compute() self._accuracy.reset() logging.info("test_loss: {:.3f}".format(test_loss_mean)) self.log('test_loss', test_loss_mean) for top_k, score in zip(self._accuracy.top_k, topk_scores): self.log('test_epoch_accuracy_top@{}'.format(top_k), score) return { 'test_loss': test_loss_mean, 'test_acc_top_k': topk_scores, } def setup_finetune_model(self, model_config: DictConfig): """ setup_finetune_model method sets up training data, validation data and test data with new provided config, this checks for the previous labels set up during training from scratch, if None, it sets up labels for provided finetune data from manifest files Args: model_config: cfg which has train_ds, optional validation_ds, optional test_ds, mandatory encoder and decoder model params. Make sure you set num_classes correctly for finetune data. Returns: None """ logging.info("Setting up data loaders with manifests provided from model_config") if 'train_ds' in model_config and model_config.train_ds is not None: self.setup_training_data(model_config.train_ds) else: raise KeyError("train_ds is not found in model_config but you need it for fine tuning") if self.labels is None or len(self.labels) == 0: raise ValueError(f'New labels must be non-empty list of labels. But I got: {self.labels}') if 'validation_ds' in model_config and model_config.validation_ds is not None: self.setup_multiple_validation_data(model_config.validation_ds) if 'test_ds' in model_config and model_config.test_ds is not None: self.setup_multiple_test_data(model_config.test_ds) if self.labels is not None: # checking for new finetune dataset labels logging.warning( "Trained dataset labels are same as finetune dataset labels -- continuing change of decoder parameters" ) else: logging.warning( "Either you provided a dummy manifest file during training from scratch or you restored from a pretrained nemo file" ) decoder_config = model_config.decoder new_decoder_config = copy.deepcopy(decoder_config) if new_decoder_config['num_classes'] != len(self.labels): raise ValueError( "number of classes provided {} is not same as number of different labels in finetuning data: {}".format( new_decoder_config['num_classes'], len(self.labels) ) ) del self.decoder self.decoder = EncDecSpeakerLabelModel.from_config_dict(new_decoder_config) with open_dict(self._cfg.decoder): self._cfg.decoder = new_decoder_config logging.info(f"Changed decoder output to # {self.decoder._num_classes} classes.") @torch.no_grad() def get_embedding(self, path2audio_file): """ Returns the speaker embeddings for a provided audio file. Args: path2audio_file: path to audio wav file Returns: embs: speaker embeddings """ audio, sr = librosa.load(path2audio_file, sr=None) target_sr = self._cfg.train_ds.get('sample_rate', 16000) if sr != target_sr: audio = librosa.core.resample(audio, sr, target_sr) audio_length = audio.shape[0] device = self.device audio = np.array(audio) audio_signal, audio_signal_len = ( torch.tensor([audio], device=device), torch.tensor([audio_length], device=device), ) mode = self.training self.freeze() _, embs = self.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) self.train(mode=mode) if mode is True: self.unfreeze() del audio_signal, audio_signal_len return embs @torch.no_grad() def verify_speakers(self, path2audio_file1, path2audio_file2, threshold=0.7): """ Verify if two audio files are from the same speaker or not. Args: path2audio_file1: path to audio wav file of speaker 1 path2audio_file2: path to audio wav file of speaker 2 threshold: cosine similarity score used as a threshold to distinguish two embeddings (default = 0.7) Returns: True if both audio files are from same speaker, False otherwise """ embs1 = self.get_embedding(path2audio_file1).squeeze() embs2 = self.get_embedding(path2audio_file2).squeeze() # Length Normalize X = embs1 / torch.linalg.norm(embs1) Y = embs2 / torch.linalg.norm(embs2) # Score similarity_score = torch.dot(X, Y) / ((torch.dot(X, X) * torch.dot(Y, Y)) ** 0.5) similarity_score = (similarity_score + 1) / 2 # Decision if similarity_score >= threshold: logging.info(" two audio files are from same speaker") return True else: logging.info(" two audio files are from different speakers") return False @staticmethod @torch.no_grad() def get_batch_embeddings(speaker_model, manifest_filepath, batch_size=32, sample_rate=16000, device='cuda'): speaker_model.eval() if device == 'cuda': speaker_model.to(device) featurizer = WaveformFeaturizer(sample_rate=sample_rate) dataset = AudioToSpeechLabelDataset(manifest_filepath=manifest_filepath, labels=None, featurizer=featurizer) dataloader = torch.utils.data.DataLoader( dataset=dataset, batch_size=batch_size, collate_fn=dataset.fixed_seq_collate_fn, ) all_logits = [] all_labels = [] all_embs = [] for test_batch in tqdm(dataloader): if device == 'cuda': test_batch = [x.to(device) for x in test_batch] audio_signal, audio_signal_len, labels, _ = test_batch logits, embs = speaker_model.forward(input_signal=audio_signal, input_signal_length=audio_signal_len) all_logits.extend(logits.cpu().numpy()) all_labels.extend(labels.cpu().numpy()) all_embs.extend(embs.cpu().numpy()) all_logits, true_labels, all_embs = np.asarray(all_logits), np.asarray(all_labels), np.asarray(all_embs) return all_embs, all_logits, true_labels, dataset.id2label