^jq#ddlmZddlmZddlmZedeGddeZedeGd d eZed eGd d eZedeGddeZ gdZ y))strict)PreTrainedConfig)auto_docstringznvidia/parakeet-ctc-1.1b) checkpointcreZdZUdZdZdgZdZeed<dZ eed<dZ eed <d Z eed <d Z e ed <dZeed<dZeed<dZeed<dZeed<dZeed<dZeed<dZeed<dZeed<dZeezed<dZeezed<dZeezed <dZeezed!<dZeezed"<d#Zeed$<dZeed%<d&Zeed'<fd(Z xZ!S))ParakeetEncoderConfiga convolution_bias (`bool`, *optional*, defaults to `True`): Whether to use bias in convolutions of the conformer's convolution module. conv_kernel_size (`int`, *optional*, defaults to 9): The kernel size of the convolution layers in the Conformer block. subsampling_factor (`int`, *optional*, defaults to 8): The factor by which the input sequence is subsampled. subsampling_conv_channels (`int`, *optional*, defaults to 256): The number of channels in the subsampling convolution layers. num_mel_bins (`int`, *optional*, defaults to 80): Number of mel features. subsampling_conv_kernel_size (`int`, *optional*, defaults to 3): The kernel size of the subsampling convolution layers. subsampling_conv_stride (`int`, *optional*, defaults to 2): The stride of the subsampling convolution layers. dropout_positions (`float`, *optional*, defaults to 0.0): The dropout ratio for the positions in the input sequence. scale_input (`bool`, *optional*, defaults to `True`): Whether to scale the input embeddings. Example: ```python >>> from transformers import ParakeetEncoderModel, ParakeetEncoderConfig >>> # Initializing a `ParakeetEncoder` configuration >>> configuration = ParakeetEncoderConfig() >>> # Initializing a model from the configuration >>> model = ParakeetEncoderModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` parakeet_encoderpast_key_values hidden_sizenum_hidden_layersnum_attention_headsiintermediate_sizesilu hidden_actTattention_biasconvolution_bias conv_kernel_sizesubsampling_factorsubsampling_conv_channelsP num_mel_binsrsubsampling_conv_kernel_sizesubsampling_conv_strideg?dropoutgdropout_positions layerdropactivation_dropoutattention_dropoutimax_position_embeddings scale_inputg{Gz?initializer_rangec F|j|_t| di|yN)rnum_key_value_headssuper __post_init__selfkwargs __class__s ~/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/parakeet/configuration_parakeet.pyr.z#ParakeetEncoderConfig.__post_init__Ts!#'#;#;  '')"__name__ __module__ __qualname____doc__ model_typekeys_to_ignore_at_inferencer int__annotations__rrrrstrrboolrrrrrrr r!floatr"r#r$r%r&r'r(r. __classcell__r2s@r3r r s!F$J#4"5Ks  !s!JND!d!c%(s(L#() #)#$S$GUS[%(us{( Ius{ &) )%(us{(#'S'K#u#((r4r ceZdZUdZdZdeiZdZee d<dZ e e d<dZ e e d <d Zeezd ze d<d Zed ze d <fd ZxZS)ParakeetCTCConfiga ctc_loss_reduction (`str`, *optional*, defaults to `"mean"`): Specifies the reduction to apply to the output of `torch.nn.CTCLoss`. Only relevant when training an instance of [`ParakeetForCTC`]. ctc_zero_infinity (`bool`, *optional*, defaults to `True`): Whether to zero infinite losses and the associated gradients of `torch.nn.CTCLoss`. Infinite losses mainly occur when the inputs are too short to be aligned to the targets. Only relevant when training an instance of [`ParakeetForCTC`]. encoder_config (`Union[dict, ParakeetEncoderConfig]`, *optional*): The config object or dictionary of the encoder. Example: ```python >>> from transformers import ParakeetForCTC, ParakeetCTCConfig >>> # Initializing a Parakeet configuration >>> configuration = ParakeetCTCConfig() >>> # Initializing a model from the configuration >>> model = ParakeetForCTC(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` parakeet_ctcencoder_configi vocab_sizemeanctc_loss_reductionTctc_zero_infinityNr pad_token_idc t|jtrtdi|j|_n|jt|_|jj|_t |di|yr* isinstancerEdictr r(r-r.r/s r3r.zParakeetCTCConfig.__post_init__}d d))4 0"7"N$:M:M"ND   ("7"9D !%!4!4!F!F ''r4)r5r6r7r8r9r sub_configsrFr;r<rHr=rIr>rErNrrJr.r@rAs@r3rCrCYsg0 J#%:;KJ$$"t"59ND++d29#L#*#((r4rCznvidia/parakeet-rnnt-0.6bceZdZUdZdZdeiZdZee d<dZ ee d<dZ ee d <d Z e e d <d Zee d <dZeezdze d<dZee d<dZee d<dZee d<fdZxZS)ParakeetRNNTConfiga decoder_hidden_size (`int`, *optional*, defaults to 640): Hidden size of the LSTM prediction network and joint network. num_decoder_layers (`int`, *optional*, defaults to 2): Number of LSTM layers in the prediction network. max_symbols_per_step (`int`, *optional*, defaults to 10): Maximum number of symbols to emit per encoder time step during greedy decoding. encoder_config (`Union[dict, ParakeetEncoderConfig]`, *optional*): The config object or dictionary of the encoder. blank_token_id (`int`, *optional*, defaults to 8192): Blank token id. Different from `pad_token_id` for RNN-T. Example: ```python >>> from transformers import ParakeetForRNNT, ParakeetRNNTConfig >>> # Initializing a Parakeet RNN-T configuration >>> configuration = ParakeetRNNTConfig() >>> # Initializing a model from the configuration >>> model = ParakeetForRNNT(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` parakeet_rnntrEi rFidecoder_hidden_sizernum_decoder_layersrelur max_symbols_per_stepNrJi blank_token_idTis_encoder_decoderc t|jtrtdi|j|_n|jt|_|jj|_t |di|yr*rLr/s r3r.z ParakeetRNNTConfig.__post_init__rOr4)r5r6r7r8r9r rPrFr;r<rTrUrr=rXrErNrrJrYrZr>r.r@rAs@r3rRrRs6!J#%:;KJ""J "#"59ND++d29L#NC##((r4rRznvidia/parakeet-tdt-0.6b-v3c:eZdZUdZdZdZeeeedfze d<y)ParakeetTDTConfigaR A TDT (Token-and-Duration Transducer) extends the base RNN-T configuration [`ParakeetRNNTConfig`] with a `durations` field: the joint network gains a duration head (its output width grows from `vocab_size` to `vocab_size + len(durations)`), and during greedy decoding the encoder frame pointer advances by the predicted duration rather than a fixed single frame. decoder_hidden_size (`int`, *optional*, defaults to 640): Hidden size of the LSTM prediction network and joint network. num_decoder_layers (`int`, *optional*, defaults to 2): Number of LSTM layers in the prediction network. max_symbols_per_step (`int`, *optional*, defaults to 10): Maximum number of symbols to emit per encoder time step during greedy decoding. durations (`list[int]`, *optional*, defaults to `[0, 1, 2, 3, 4]`): Token duration values that can be predicted. Each value represents how many frames a token or blank emission spans. encoder_config (`Union[dict, ParakeetEncoderConfig]`, *optional*): The config object or dictionary of the encoder. blank_token_id (`int`, *optional*, defaults to 8192): Blank token id. Different from `pad_token_id` for TDT. Example: ```python >>> from transformers import ParakeetForTDT, ParakeetTDTConfig >>> # Initializing a Parakeet TDT configuration >>> configuration = ParakeetTDTConfig() >>> # Initializing a model from the configuration >>> model = ParakeetForTDT(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` parakeet_tdt)rrr. durationsN) r5r6r7r8r9ralistr;tupler<r+r4r3r]r]s*!F J-rhs/3#56?(,?(7?(D56(((((7((V67/()/(8/(d89%=*%=:%=P dr4