--- license: openrail language: - en metrics: - f1 library_name: fairseq pipeline_tag: audio-classification --- # Model Card for Model ID We explore benefits of unsupervised pretraining of wav2vec 2.0 (W2V2) using large-scale unlabeled home recordings collected using LittleBeats and LENA (Language Environment Analysis) devices. LittleBeats (LB) is a new infant wearable multi-modal device that we developed, which simultaneously records audio, movement of the infant, as well as heart-rate variablity. We use W2V2 to advance LB audio pipeline such that it automatically provides reliable labels of speaker diarization and vocalization classifications for family members, including infants, parents, and siblings, at home. We show that W2V2 pretrained on thousands hours of large-scale unlabeled home audio outperforms oracle W2V2 pretrained on 52k-hours released by Facebook/Meta in terms of automatic family audio analysis tasks. # Model Details ## Model Description Two versions of pretrained W2V2 models are available: - **LB1100/checkpoint_best.pt** pretrained using 1100-hour of LB home recordings collected from 110 families of children under 5-year-old - **LL4300/checkpoint_best.pt** pretrained using 1100-hour of LB home recordings collected from 110 families + 3200-hour of LENA home recordings from 275 families of children under 5-year-old ## Model Sources For more information regarding this model, please checkout our paper - **Paper [optional]:** [More Information Needed] # Uses We develop fine-tuning recipe using SpeechBrain toolkit available at - **Repository:** https://github.com/jialuli3/speechbrain/tree/infant-voc-classification/recipes/wav2vec_kic ## Quick Start [optional] If you wish to use fairseq framework, the following code snippet can be used to load our pretrained model ''' """This lobe enables the integration of fairseq pretrained wav2vec models. Reference: https://arxiv.org/abs/2006.11477 Reference: https://arxiv.org/abs/1904.05862 FairSeq >= 1.0.0 needs to be installed: https://fairseq.readthedocs.io/en/latest/ Authors * Titouan Parcollet 2021 * Salima Mdhaffar 2021 """ import torch import torch.nn.functional as F from torch import nn from speechbrain.utils.data_utils import download_file import pdb # We check if fairseq is installed. try: import fairseq except ImportError: MSG = "Please install Fairseq to use pretrained wav2vec\n" MSG += "E.G. run: pip install fairseq" raise ImportError(MSG) class FairseqWav2Vec2(nn.Module): """This lobe enables the integration of fairseq pretrained wav2vec2.0 models. Source paper: https://arxiv.org/abs/2006.11477 FairSeq >= 1.0.0 needs to be installed: https://fairseq.readthedocs.io/en/latest/ The model can be used as a fixed features extractor or can be finetuned. It will download automatically the model if a url is given (e.g FairSeq repository from GitHub). Arguments --------- pretrained_path : str Path of the pretrained wav2vec2 model. It can be a url or a local path. save_path : str Path and filename of the downloaded model. input_norm : bool (default: None) If True, a layer_norm (affine) will be applied to the input waveform. By default, it is extracted from the checkpoint of the downloaded model in order to match the pretraining conditions. However, if this information is not given in the checkpoint, it has to be given manually. output_norm : bool (default: True) If True, a layer_norm (affine) will be applied to the output obtained from the wav2vec model. freeze : bool (default: True) If True, the model is frozen. If False, the model will be trained alongside with the rest of the pipeline. pretrain : bool (default: True) If True, the model is pretrained with the specified source. If False, the randomly-initialized model is instantiated. dropout : float (default: None) If different from None (0.0 to 1.0), it will override the given fairseq dropout rates. This is useful if the wav2vec2 model has been trained without dropout and one wants to reactivate it for downstream task fine-tuning (better performance observed). Example ------- >>> inputs = torch.rand([10, 600]) >>> model_url = "https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_small.pt" >>> save_path = "models_checkpoints/wav2vec2.pt" >>> model = FairseqWav2Vec2(model_url, save_path) >>> outputs = model(inputs) >>> outputs.shape torch.Size([10, 100, 768]) """ def __init__( self, pretrained_path, save_path, input_norm=None, output_norm=True, freeze=True, pretrain=True, dropout=None, encoder_dropout = None, output_all_hiddens=False, tgt_layer=None, include_CNN_layer=True, ): super().__init__() # Download the pretrained wav2vec2 model. It can be local or online. download_file(pretrained_path, save_path) # During pretraining dropout might be set to 0. However, we might want # to apply dropout when fine-tuning on a downstream task. Hence we need # to modify the fairseq cfg to activate dropout (if requested). overrides={} if encoder_dropout is not None: overrides = { "model": { "encoder_layerdrop": encoder_dropout, } } if not freeze: if dropout is not None and encoder_dropout is not None: overrides = { "model": { "dropout": dropout, "encoder_layerdrop": encoder_dropout, "dropout_input": dropout, "attention_dropout": dropout, } } elif dropout is not None: overrides = { "model": { "dropout": dropout, "dropout_input": dropout, "attention_dropout": dropout, } } ( model, cfg, task, ) = fairseq.checkpoint_utils.load_model_ensemble_and_task( [save_path], arg_overrides=overrides ) # wav2vec pretrained models may need the input waveform to be normalized # Hence, we check if the model has be trained with or without it. # If the information isn't contained in the checkpoint IT HAS TO BE GIVEN # BY THE USER. if input_norm is None: if hasattr(cfg["task"], "normalize"): self.normalize = cfg["task"].normalize elif hasattr(cfg, "normalize"): self.normalize = cfg.normalize else: self.normalize = False else: self.normalize = input_norm model = model[0] self.model = model self.freeze = freeze self.output_norm = output_norm if self.freeze: self.model.eval() # Freeze parameters for param in model.parameters(): param.requires_grad = False else: self.model.train() for param in model.parameters(): param.requires_grad = True # Randomly initialized layers if pretrain is False if not (pretrain): self.reset_layer(self.model) # Following the fairseq implementation of downstream training, # we remove some modules that are unnecessary. self.remove_pretraining_modules() self.output_all_hiddens = output_all_hiddens self.tgt_layer = tgt_layer self.include_CNN_layer = include_CNN_layer def forward(self, wav): """Takes an input waveform and return its corresponding wav2vec encoding. Arguments --------- wav : torch.Tensor (signal) A batch of audio signals to transform to features. """ # If we freeze, we simply remove all grads and features from the graph. if self.freeze: with torch.no_grad(): return self.extract_features(wav).detach() return self.extract_features(wav) def extract_features(self, wav): """Extracts the wav2vect embeddings""" # We normalize the input signal if needed. if self.normalize: wav = F.layer_norm(wav, wav.shape) # Extract wav2vec output if self.tgt_layer=="CNN": #initial embeddings from conv out = self.model.extract_features(wav, padding_mask=None, mask=False) out = self.model.post_extract_proj(out['features']) elif isinstance(self.tgt_layer, int): out = self.model.extract_features(wav, padding_mask=None, mask=False, layer=self.tgt_layer)['x'] else: # out = self.model.extract_features(wav, padding_mask=None, mask=False, layer=self.tgt_layer) if self.output_all_hiddens or isinstance(self.tgt_layer, list): out = self.aggregate_features(out, include_CNN_layer=self.include_CNN_layer) # 13, B, T, D if isinstance(self.tgt_layer, list): out = out[self.tgt_layer] else: out = out['x'] # We normalize the output if required if self.output_norm: out = F.layer_norm(out, out.shape) return out def aggregate_features(self, out, include_CNN_layer=True): features = [] if include_CNN_layer: features = [self.model.post_extract_proj(out['features'])] self.model.layerdrop = 0 for i in range(len(out['layer_results'])): curr_feature = out['layer_results'][i][0].transpose(0,1) features.append(curr_feature) features = torch.stack(features) return features def reset_layer(self, model): """Reinitializes the parameters of the network""" if hasattr(model, "reset_parameters"): model.reset_parameters() for child_layer in model.children(): if model != child_layer: self.reset_layer(child_layer) def remove_pretraining_modules(self): """ Remove uneeded modules. Inspired by the same fairseq function.""" self.model.quantizer = None self.model.project_q = None self.model.target_glu = None self.model.final_proj = None class FairseqWav2Vec1(nn.Module): """This lobes enables the integration of fairseq pretrained wav2vec1.0 models. Arguments --------- pretrained_path : str Path of the pretrained wav2vec1 model. It can be a url or a local path. save_path : str Path and filename of the downloaded model. output_norm : bool (default: True) If True, a layer_norm (affine) will be applied to the output obtained from the wav2vec model. freeze : bool (default: True) If True, the model is frozen. If False, the model will be trained alongside with the rest of the pipeline. pretrain : bool (default: True) If True, the model is pretrained with the specified source. If False, the randomly-initialized model is instantiated. Example ------- >>> inputs = torch.rand([10, 600]) >>> model_url = "" >>> save_path = "models_checkpoints/wav2vec.pt" >>> model = FairseqWav2Vec1(model_url, save_path) >>> outputs = model(inputs) >>> outputs.shape torch.Size([10, 100, 512]) """ def __init__( self, pretrained_path, save_path, output_norm=True, freeze=True, pretrain=True, ): super().__init__() self.freeze = freeze self.output_norm = output_norm # Download the pretrained wav2vec1 model. It can be local or online. download_file(pretrained_path, save_path) ( model, cfg, task, ) = fairseq.checkpoint_utils.load_model_ensemble_and_task( [pretrained_path] ) self.model = model self.model = self.model[0] if self.freeze: model.eval() # Randomly initialized layers if pretrain is False if not (pretrain): self.reset_layer(self.model) def forward(self, wav): """Takes an input waveform and return its corresponding wav2vec encoding. Arguments --------- wav : torch.Tensor (signal) A batch of audio signals to transform to features. """ # If we freeze, we simply remove all grads and features from the graph. if self.freeze: with torch.no_grad(): return self.extract_features(wav).detach() return self.extract_features(wav) def extract_features(self, wav): """Extracts the wav2vect embeddings""" out = self.model.feature_extractor(wav) out = self.model.feature_aggregator(out).squeeze(0) out = out.transpose(2, 1) # We normalize the output if required if self.output_norm: out = F.layer_norm(out, out.shape) return out def reset_layer(self, model): """Reinitializes the parameters of the network""" if hasattr(model, "reset_parameters"): model.reset_parameters() for child_layer in model.children(): if model != child_layer: self.reset_layer(child_layer) ''' # Evaluation We test 4 unlabeled datasets on unsupervised pretrained W2V2-base models: - **base (oracle version):** originally released version pretrained on ~52k-hour unlabeled audio - **Libri960h:** oracle version fine-tuned using 960h Librispeech - **LB1100h:** pretrain W2V2 using 1100h LB home recordings - **LL4300h:** pretrain W2V2 using 4300h LB+LENA home recordings We then fine-tune pretrained models on 11.7h of LB labeled home recordings, the f1 scores across three tasks are ![results](results.png) For more details of experiments and results, please refer to our paper. # Citation If you found this model helpful to you, please cite us as **BibTeX:** # Model Card Contact Jialu Li (she, her, hers) Ph.D candidate @ Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign E-mail: jialuli3@illinois.edu Homepage: https://sites.google.com/view/jialuli/