o @TiT9@sdZddlZddlmZmZmZddlZddlmm Z ddlm Z ddl m Z ddlmZGdd d e Zd efd d Z d2dedeeeeeefdededededededeedeedeedededeededededed eed!ed"ed#ed$ed%ed&ed ef4d'd(Z d2dedeeeeeefdededededededeedeed)eed*eeed+ed,ededeededededed eed!ed"ed#ed$ed%ed&ed ef8d-d.Z / / / / / d3dededededed eed efd0d1ZdS)4zSpeech SSL models supporting pruning. Originally from: https://github.com/pytorch/audio/blob/main/torchaudio/models/wav2vec2/model.py N)ListOptionalTuple)Tensor)Module) componentsc seZdZdZ ddedededeeffdd Zej j  dd e d ee d ee d e ee ee ffd dZddZddZ dd e d ee d e e ee ffddZZS) Wav2Vec2ModelaAcoustic model used in *wav2vec 2.0* :cite:`baevski2020wav2vec`. Note: To build the model, please use one of the factory functions. :py:func:`wav2vec2_model`, :py:func:`wav2vec2_base`, :py:func:`wav2vec2_large`, :py:func:`wav2vec2_large_lv60k`, :py:func:`hubert_base`, :py:func:`hubert_large`, and :py:func:`hubert_xlarge`. See Also: * :class:`torchaudio.pipelines.Wav2Vec2Bundle`: Pretrained models (without fine-tuning) * :class:`torchaudio.pipelines.Wav2Vec2ASRBundle`: ASR pipelines with pretrained models. Args: feature_extractor (torch.nn.Module): Feature extractor that extracts feature vectors from raw audio Tensor. encoder (torch.nn.Module): Encoder that converts the audio features into the sequence of probability distribution (in negative log-likelihood) over labels. aux (torch.nn.Module or None, optional): Auxiliary module. If provided, the output from encoder is passed to this module. Nnormalize_waveformfeature_extractorencoderauxcs&t||_||_||_||_dSN)super__init__r r r r )selfr r r r  __class__R/home/ubuntu/.local/lib/python3.10/site-packages/linacodec/module/distill_wavlm.pyr,s  zWav2Vec2Model.__init__ waveformslengths num_layersreturncCsv|jr'|durddt||D}tjjjj|dd}n t||j dd}| ||\}}|j |||}||fS)a%Extract feature vectors from raw waveforms This returns the list of outputs from the intermediate layers of transformer block in encoder. Args: waveforms (Tensor): Audio tensor of shape `(batch, frames)`. lengths (Tensor or None, optional): Indicates the valid length of each audio in the batch. Shape: `(batch, )`. When the ``waveforms`` contains audios with different durations, by providing ``lengths`` argument, the model will compute the corresponding valid output lengths and apply proper mask in transformer attention layer. If ``None``, it is assumed that the entire audio waveform length is valid. num_layers (int or None, optional): If given, limit the number of intermediate layers to go through. Providing `1` will stop the computation after going through one intermediate layers. If not given, the outputs from all the intermediate layers are returned. Returns: (List[Tensor], Optional[Tensor]): List of Tensors Features from requested layers. Each Tensor is of shape: `(batch, time frame, feature dimension)` Tensor or None If ``lengths`` argument was provided, a Tensor of shape `(batch, )` is returned. It indicates the valid length in time axis of each feature Tensor. NcS&g|]\}}t|d||fqSrF layer_norm.0wavelengthrrr bz2Wav2Vec2Model.extract_features..T batch_first) r ziptorchnnutilsrnn pad_sequencerrshaper r extract_features)rrrrxrrrr.9s'zWav2Vec2Model.extract_featurescCs"|j\}}|j|}||S)zCalculate the current size.)r %get_num_params_and_final_out_channelsr get_num_params)rfeature_extractor_sizeencoder_in_features encoder_sizerrrr1ms zWav2Vec2Model.get_num_paramsc CsZ||j\}}|j|}|d}|d}|d}|d}|d}||||||fS)N use_attentionuse_feed_forward num_headsremaining_headsff_interm_features)evalr pruner ) r conv_configconv_out_indextransformer_configr5r6r7r8r9rrrr;ss zWav2Vec2Model.prunecCs|jr'|durddt||D}tjjjj|dd}n t||j dd}| ||\}}| ||}|j dur?| |}||fS)aCompute the sequence of probability distribution over labels. Args: waveforms (Tensor): Audio tensor of shape `(batch, frames)`. lengths (Tensor or None, optional): Indicates the valid length of each audio in the batch. Shape: `(batch, )`. When the ``waveforms`` contains audios with different durations, by providing ``lengths`` argument, the model will compute the corresponding valid output lengths and apply proper mask in transformer attention layer. If ``None``, it is assumed that all the audio in ``waveforms`` have valid length. Default: ``None``. Returns: (Tensor, Optional[Tensor]): Tensor The sequences of probability distribution (in logit) over labels. Shape: `(batch, frames, num labels)`. Tensor or None If ``lengths`` argument was provided, a Tensor of shape `(batch, )` is returned. It indicates the valid length in time axis of the output Tensor. NcSrrrrrrrr"r#z)Wav2Vec2Model.forward..Tr$r&) r r'r(r)r*r+r,rrr-r r r )rrrr/rrrforwards   zWav2Vec2Model.forwardr)NN)__name__ __module__ __qualname____doc__boolrrrr(jitexportrintrrr.r1r;r? __classcell__rrrrr sD  3r rcKs$d|vr tdi|Stdi|S)z2Wraps the original wav2vec2_model and wavlm_model.encoder_remaining_headsNr) wavlm_modelwav2vec2_model_original)configsrrrwav2vec2_modelsrMFextractor_modeextractor_conv_layer_configextractor_conv_biasencoder_embed_dimencoder_projection_dropoutencoder_pos_conv_kernelencoder_pos_conv_groupsencoder_num_layersencoder_use_attentionencoder_use_feed_forwardencoder_num_headsencoder_head_dimencoder_attention_dropoutencoder_ff_interm_featuresencoder_ff_interm_dropoutencoder_dropoutencoder_layer_norm_firstencoder_layer_drop aux_num_outr extractor_prune_conv_channelsencoder_prune_attention_headsencoder_prune_attention_layer'encoder_prune_feed_forward_intermediate encoder_prune_feed_forward_layercCs|durdgdgddgd}tj||||d}tjdid|d d d |d |d |d|d|d|d| d| d| d| d| d|d|d|d|d|d|d|d|}d}|durntjj||d}t||||S) asBuilds custom :class:`~torchaudio.models.Wav2Vec2Model`. Note: The "feature extractor" below corresponds to `ConvFeatureExtractionModel `__ in the original ``fairseq`` implementation. This is referred as "(convolutional) feature encoder" in the *wav2vec 2.0* :cite:`baevski2020wav2vec` paper. The "encoder" below corresponds to `TransformerEncoder `__, and this is referred as "Transformer" in the paper. Args: extractor_mode (str): Operation mode of feature extractor. Valid values are ``"group_norm"`` or ``"layer_norm"``. If ``"group_norm"``, then a single normalization is applied in the first convolution block. Otherwise, all the convolution blocks will have layer normalization. This option corresponds to ``extractor_mode`` from ``fairseq``. extractor_conv_layer_config (list of integer tuples or None): Configuration of convolution layers in feature extractor. List of convolution configuration, i.e. ``[(output_channel, kernel_size, stride), ...]`` If ``None`` is provided, then the following default value is used. .. code-block:: python [ (512, 10, 5), (512, 3, 2), (512, 3, 2), (512, 3, 2), (512, 3, 2), (512, 2, 2), (512, 2, 2), ] This option corresponds to ``conv_feature_layers`` from ``fairseq``. extractor_conv_bias (bool): Whether to include bias term to each convolution operation. This option corresponds to ``conv_bias`` from ``fairseq``. encoder_embed_dim (int): The dimension of embedding in encoder. This option corresponds to ``encoder_embed_dim`` from ``fairseq``. encoder_projection_dropout (float): The dropout probability applied after the input feature is projected to ``encoder_embed_dim``. This option corresponds to ``dropout_input`` from ``fairseq``. encoder_pos_conv_kernel (int): The kernel size of convolutional positional embeddings. This option corresponds to ``conv_pos`` from ``fairseq``. encoder_pos_conv_groups (int): The number of groups of convolutional positional embeddings. This option corresponds to ``conv_pos_groups`` from ``fairseq``. encoder_num_layers (int): The number of self attention layers in transformer block. This option corresponds to ``encoder_layers`` from ``fairseq``. encoder_num_heads (int): The number of heads in self attention layers. This option corresponds to ``encoder_attention_heads`` from ``fairseq``. encoder_attention_dropout (float): The dropout probability applied after softmax in self-attention layer. This option corresponds to ``attention_dropout`` from ``fairseq``. encoder_ff_interm_features (int): The dimension of hidden features in feed forward layer. This option corresponds to ``encoder_ffn_embed_dim`` from ``fairseq``. encoder_ff_interm_dropout (float): The dropout probability applied in feedforward layer. This option correspinds to ``activation_dropout`` from ``fairseq``. encoder_dropout (float): The dropout probability applied at the end of feed forward layer. This option corresponds to ``dropout`` from ``fairseq``. encoder_layer_norm_first (bool): Control the order of layer norm in transformer layer and each encoder layer. If True, in transformer layer, layer norm is applied before features are fed to encoder layers. In encoder layer, two layer norms are applied before and after self attention. If False, in transformer layer, layer norm is applied after features are fed to encoder layers. In encoder layer, two layer norms are applied after self attention, before and after feed forward. This option corresponds to ``layer_norm_first`` from ``fairseq``. encoder_layer_drop (float): Probability to drop each encoder layer during training. This option corresponds to ``layerdrop`` from ``fairseq``. aux_num_out (int or None): When provided, attach an extra linear layer on top of encoder, which can be used for fine-tuning. Returns: Wav2Vec2Model: The resulting model. N rgrgrlrlrlprune_conv_channels in_featuresr&r embed_dim dropout_inputpos_conv_kernelpos_conv_groupsrr5r6r7head_dimattention_dropoutr9ff_interm_dropoutdropoutlayer_norm_first layer_dropprune_attention_headsprune_attention_layerprune_feed_forward_intermediateprune_feed_forward_layerrq out_featuresr)r_get_feature_extractor _get_encoderr(r)Linearr )rNrOrPrQrRrSrTrUrVrWrXrYrZr[r\r]r^r_r`r rarbrcrdrer r r rrrrKsh       rKencoder_total_num_headsrIencoder_num_bucketsencoder_max_distancecCs|durdgdgddgd}tj||||d}tjd!id|d d d |d |d |d|d|d|d| d| d| d| d| d|d|d|d|d|d|d|d|d|d|}d}|durttjj||d }t||||S)"aBuilds custom WaveLM model :cite:`chen2022wavlm`. The architecture is compatible with Wav2Vec2 model :cite:`baevski2020wav2vec`, and so the output object is :class:`~torchaudio.models.Wav2Vec2Model`. Most of the arguments have the same meaning as in :py:func:`wav2vec2_model` so please refer there for documentation. Args: extractor_mode (str): Operation mode of feature extractor. See :py:func:`wav2vec2_model`. extractor_conv_layer_config (list of integer tuples or None): See :py:func:`wav2vec2_model`. extractor_conv_bias (bool): See :py:func:`wav2vec2_model`. encoder_embed_dim (int): See :py:func:`wav2vec2_model`. encoder_projection_dropout (float): See :py:func:`wav2vec2_model`. encoder_pos_conv_kernel (int): See :py:func:`wav2vec2_model`. encoder_pos_conv_groups (int): See :py:func:`wav2vec2_model`. encoder_num_layers (int): See :py:func:`wav2vec2_model`. encoder_num_heads (int): See :py:func:`wav2vec2_model`. encoder_num_buckets (int): Number of buckets for relative position embedding. encoder_max_distance (int): Maximum distance for relative position embedding. encoder_attention_dropout (float): See :py:func:`wav2vec2_model`. encoder_ff_interm_features (int): See :py:func:`wav2vec2_model`. encoder_ff_interm_dropout (float): See :py:func:`wav2vec2_model`. encoder_dropout (float): See :py:func:`wav2vec2_model`. encoder_layer_norm_first (bool): See :py:func:`wav2vec2_model`. encoder_layer_drop (float): See :py:func:`wav2vec2_model`. aux_num_out (int or None): See :py:func:`wav2vec2_model`. Returns: Wav2Vec2Model: The resulting model. Nrfrjrmrnrlrorqr&rrrrsrtrurr5r6total_num_headsr8 num_buckets max_distancerwr9rxryrzr{r|r}r~rrr)rr_get_wavlm_encoderr(r)rr )rNrOrPrQrRrSrTrUrVrWrrIrrrZr[r\r]r^r_r`r rarbrcrdrer r r rrrrJksn\       rJ皙?cCsvtdiddddddddd |d d d d ddddddddd|ddd|d|ddd|d|S)aBuilds "base" WaveLM model :cite:`chen2022wavlm`. The architecture is compatible with Wav2Vec2 model :cite:`baevski2020wav2vec`, and so the output class is :class:`~torchaudio.models.Wav2Vec2Model`. Args: encoder_projection_dropout (float): See :py:func:`wav2vec2_model`. encoder_attention_dropout (float): See :py:func:`wav2vec2_model`. encoder_ff_interm_dropout (float): See :py:func:`wav2vec2_model`. encoder_dropout (float): See :py:func:`wav2vec2_model`. encoder_layer_drop (float): See :py:func:`wav2vec2_model`. aux_num_out (int, optional): See :py:func:`wav2vec2_model`. Returns: Wav2Vec2Model: The resulting model. rN group_normrONrPFrQirRrSrTrU rXri@ri rZr[i r\r]r^r_r`r)rJ)rRrZr\r]r_r`rrr wavlm_basesJ     r)FFFFF)rrrrrN)rCmathtypingrrrr(torch.nn.functionalr) functionalrrtorch.nnrrr rMstrrGrDfloatrKrJrrrrrs4         N