U
    CŒÑh'2  ã                   @   sn  d dl mZmZmZmZmZ d dlZd dlmZ ddgZ	eee
 ejeeej  ef Zde_eee
 dœdd	„Zeejdœd
d„Zeeeej  dœdd„Zeedœdd„Zeedœdd„Zee eeej  dœdd„Zeeej  e
ejeeej  dœdd„Zeedœdd„Zee eje
eejejejf dœdd„Zeee ddœdd„ZG d d„ dejjƒZdS )!é    )ÚCallableÚDictÚListÚOptionalÚTupleN)ÚRNNTÚ
HypothesisÚRNNTBeamSearchz™Hypothesis generated by RNN-T beam search decoder,
    represented as tuple of (tokens, prediction network output, prediction network state, score).
    )ÚhypoÚreturnc                 C   s   | d S ©Nr   © ©r
   r   r   úL/tmp/user/2202/pip-unpacked-wheel-p89lpdx5/torchaudio/models/rnnt_decoder.pyÚ_get_hypo_tokens   s    r   c                 C   s   | d S ©Né   r   r   r   r   r   Ú_get_hypo_predictor_out   s    r   c                 C   s   | d S )Né   r   r   r   r   r   Ú_get_hypo_state   s    r   c                 C   s   | d S )Né   r   r   r   r   r   Ú_get_hypo_score   s    r   c                 C   s   t | d ƒS r   )Ústrr   r   r   r   Ú_get_hypo_key    s    r   )Úhyposr   c              	      sn   g }t tt| d ƒƒƒD ]P‰ g }t tt| d ƒˆ  ƒƒD ]$‰| t ‡ ‡fdd„| D ƒ¡¡ q8| |¡ q|S )Nr   c                    s   g | ]}t |ƒˆ  ˆ ‘qS r   )r   ©Ú.0r
   ©ÚiÚjr   r   Ú
<listcomp>)   s     z _batch_state.<locals>.<listcomp>)ÚrangeÚlenr   ÚappendÚtorchÚcat)r   ÚstatesZbatched_state_componentsr   r   r   Ú_batch_state$   s    "r'   )r&   ÚidxÚdevicer   c                    s"   t j|g|d�‰ ‡ fdd„| D ƒS )N©r)   c                    s   g | ]}‡ fd d„|D ƒ‘qS )c                    s   g | ]}|  d ˆ ¡‘qS )r   )Zindex_select)r   Ústate©Z
idx_tensorr   r   r    0   s     z+_slice_state.<locals>.<listcomp>.<listcomp>r   )r   Zstate_tupler,   r   r   r    0   s     z _slice_state.<locals>.<listcomp>)r$   Útensor)r&   r(   r)   r   r,   r   Ú_slice_state.   s    r.   c                 C   s   t | ƒtt| ƒƒd  S r   )r   r"   r   r   r   r   r   Ú_default_hypo_sort_key3   s    r/   )r   Únext_token_probsÚ
beam_widthr   c           	      C   sr   t  dd„ | D ƒ¡ d¡}||d d …d d…f  }| d¡ |¡\}}|j|jd dd�}||jd  }|||fS )Nc                 S   s   g | ]}t |ƒ‘qS r   ©r   ©r   Úhr   r   r   r    <   s     z+_compute_updated_scores.<locals>.<listcomp>r   éÿÿÿÿÚtrunc)Zrounding_mode)r$   r-   Ú	unsqueezeZreshapeÚtopkÚdivÚshape)	r   r0   r1   Zhypo_scoresZnonblank_scoresÚnonblank_nbest_scoresZnonblank_nbest_idxÚnonblank_nbest_hypo_idxÚnonblank_nbest_tokenr   r   r   Ú_compute_updated_scores7   s    r>   )r
   Ú	hypo_listr   c                 C   s0   t |ƒD ]"\}}t| ƒt|ƒkr||=  q,qd S ©N)Ú	enumerater   )r
   r?   r   Úelemr   r   r   Ú_remove_hypoD   s    rC   c                       s„  e Zd ZdZd eeeeee	gef  eddœ‡ fdd„Z
ejee	 dœd	d
„Zejee	 ejejdœdd„Zee	 ee	 ejeee	f ee	 dœdd„Zee	 ee	 ejeeejee	 dœdd„Zee	 ee ee eejee	 dœdd„Zejeee	  eee	 dœdd„Zejejeee	 dœdd„Zejjd!ejejeeeeej   eee	  eee	 eeej  f dœdd„ƒZ‡  ZS )"r	   a)  Beam search decoder for RNN-T model.

    See Also:
        * :class:`torchaudio.pipelines.RNNTBundle`: ASR pipeline with pretrained model.

    Args:
        model (RNNT): RNN-T model to use.
        blank (int): index of blank token in vocabulary.
        temperature (float, optional): temperature to apply to joint network output.
            Larger values yield more uniform samples. (Default: 1.0)
        hypo_sort_key (Callable[[Hypothesis], float] or None, optional): callable that computes a score
            for a given hypothesis to rank hypotheses by. If ``None``, defaults to callable that returns
            hypothesis score normalized by token sequence length. (Default: None)
        step_max_tokens (int, optional): maximum number of tokens to emit per input time step. (Default: 100)
    ç      ð?Néd   )ÚmodelÚblankÚtemperatureÚhypo_sort_keyÚstep_max_tokensr   c                    s<   t ƒ  ¡  || _|| _|| _|d kr,t| _n|| _|| _d S r@   )ÚsuperÚ__init__rF   rG   rH   r/   rI   rJ   )ÚselfrF   rG   rH   rI   rJ   ©Ú	__class__r   r   rL   \   s    
zRNNTBeamSearch.__init__)r)   r   c           	      C   sZ   | j }d }tjdg|d�}| j tj|gg|d�||¡\}}}|g|d  ¡ |df}|gS )Nr   r*   r   g        )rG   r$   r-   rF   ÚpredictÚdetach)	rM   r)   Útokenr+   Ú
one_tensorÚpred_outÚ_Z
pred_stateZ	init_hypor   r   r   Ú_init_b_hyposp   s    $
üzRNNTBeamSearch._init_b_hypos)Úenc_outr   r)   r   c              	   C   s~   t jdg|d�}t jdd„ |D ƒdd�}| j |||t jdgt|ƒ |d�¡\}}}t jjj|| j	 dd�}|d d …ddf S )Nr   r*   c                 S   s   g | ]}t |ƒ‘qS r   )r   r3   r   r   r   r    ‚   s     z8RNNTBeamSearch._gen_next_token_probs.<locals>.<listcomp>r   )Údimr   )
r$   r-   ÚstackrF   Újoinr"   ÚnnZ
functionalZlog_softmaxrH   )rM   rW   r   r)   rS   Zpredictor_outZ
joined_outrU   r   r   r   Ú_gen_next_token_probs~   s    ü
z$RNNTBeamSearch._gen_next_token_probs)Úb_hyposÚa_hyposr0   Úkey_to_b_hypor   c                    sÎ   t t|ƒƒD ]’}|| }t|ƒ||df  }t|ƒ|krh|t|ƒ }t|ˆ ƒ tt t|ƒ¡ |¡ƒ}	nt|ƒ}	t	|ƒt
|ƒt|ƒ|	f}ˆ  |¡ ||t|ƒ< qt dd„ ˆ D ƒ¡ ¡ \}
}‡ fdd„|D ƒS )Nr5   c                 S   s   g | ]}t |ƒ‘qS r   r2   r   r   r   r   r    ¤   s     z/RNNTBeamSearch._gen_b_hypos.<locals>.<listcomp>c                    s   g | ]}ˆ | ‘qS r   r   ©r   r(   ©r]   r   r   r    ¥   s     )r!   r"   r   r   rC   Úfloatr$   r-   Z	logaddexpr   r   r   r#   Úsort)rM   r]   r^   r0   r_   r   Úh_aZappend_blank_scoreZh_bÚscorerU   Ú
sorted_idxr   ra   r   Ú_gen_b_hyposŒ   s"    
ü
zRNNTBeamSearch._gen_b_hypos)r^   r]   r0   Útr1   r)   r   c                 C   s¼   t |||ƒ\}}}	t|ƒ|k r*tdƒ }
nt||  ƒ}
g }g }g }t|ƒD ]N}t|| ƒ}||
krLt|| ƒ}| || ¡ | t|	| ƒ¡ | |¡ qL|r´|  |||||¡}ng }|S )NÚinf)r>   r"   rb   r   r!   Úintr#   Ú_gen_new_hypos)rM   r^   r]   r0   rh   r1   r)   r;   r<   r=   Zb_nbest_scoreÚ
base_hyposÚ
new_tokensZ
new_scoresr   re   Z
a_hypo_idxÚ	new_hyposr   r   r   Ú_gen_a_hypos§   s,    
üzRNNTBeamSearch._gen_a_hypos)rl   ÚtokensÚscoresrh   r)   r   c              	   C   sœ   t jdd„ |D ƒ|d�}t|ƒ}| j |t jdgt|ƒ |d�|¡\}}	}
g }t|ƒD ]@\}}t|ƒ|| g }| |||  	¡ t
|
||ƒ|| f¡ qV|S )Nc                 S   s   g | ]
}|g‘qS r   r   )r   rR   r   r   r   r    Õ   s     z1RNNTBeamSearch._gen_new_hypos.<locals>.<listcomp>r*   r   )r$   r-   r'   rF   rP   r"   rA   r   r#   rQ   r.   )rM   rl   rp   rq   rh   r)   Z
tgt_tokensr&   rT   rU   Zpred_statesrn   r   rd   rm   r   r   r   rk   Í   s    ý
(zRNNTBeamSearch._gen_new_hypos)rW   r
   r1   r   c              	      sú   |j d }|j}g }|d kr&ˆ |¡n|‰ t|ƒD ]Â}ˆ }tj tt g ¡‰ i }d}	|rÀˆ 	|d d …||d …f ||¡}
|
 
¡ }
ˆ ˆ ||
|¡‰ |	ˆjkržqÀˆ |ˆ |
|||¡}|rT|	d7 }	qTt ‡fdd„ˆ D ƒ¡ |¡\}}‡ fdd„|D ƒ‰ q2ˆ S )Nr   r   c                    s   g | ]}ˆ   |¡‘qS r   )rI   )r   Zhyp)rM   r   r   r      s     z*RNNTBeamSearch._search.<locals>.<listcomp>c                    s   g | ]}ˆ | ‘qS r   r   r`   ra   r   r   r      s     )r:   r)   rV   r!   r$   ÚjitZannotater   r   r\   Úcpurg   rJ   ro   r-   r8   )rM   rW   r
   r1   Zn_time_stepsr)   r^   rh   r_   Zsymbols_current_tr0   rU   rf   r   )r]   rM   r   Ú_searchâ   s8    
"
ú
"zRNNTBeamSearch._search)ÚinputÚlengthr1   r   c                 C   s–   |  ¡ dkr.|  ¡ dkr&|jd dks.tdƒ‚|  ¡ dkrD| d¡}|jdkr`|jdkr`tdƒ‚|  ¡ dkrv| d¡}| j ||¡\}}|  |d	|¡S )
a  Performs beam search for the given input sequence.

        T: number of frames;
        D: feature dimension of each frame.

        Args:
            input (torch.Tensor): sequence of input frames, with shape (T, D) or (1, T, D).
            length (torch.Tensor): number of valid frames in input
                sequence, with shape () or (1,).
            beam_width (int): beam size to use during search.

        Returns:
            List[Hypothesis]: top-``beam_width`` hypotheses found by beam search.
        r   r   r   r   ú*input must be of shape (T, D) or (1, T, D)r   ©r   ú"length must be of shape () or (1,)N)rX   r:   Ú
ValueErrorr7   rF   Z
transcribert   )rM   ru   rv   r1   rW   rU   r   r   r   Úforward  s    &

zRNNTBeamSearch.forward)ru   rv   r1   r+   Ú
hypothesisr   c                 C   sž   |  ¡ dkr.|  ¡ dkr&|jd dks.tdƒ‚|  ¡ dkrD| d¡}|jdkr`|jdkr`tdƒ‚|  ¡ dkrv| d¡}| j |||¡\}}}|  |||¡|fS )	a™  Performs beam search for the given input sequence in streaming mode.

        T: number of frames;
        D: feature dimension of each frame.

        Args:
            input (torch.Tensor): sequence of input frames, with shape (T, D) or (1, T, D).
            length (torch.Tensor): number of valid frames in input
                sequence, with shape () or (1,).
            beam_width (int): beam size to use during search.
            state (List[List[torch.Tensor]] or None, optional): list of lists of tensors
                representing transcription network internal state generated in preceding
                invocation. (Default: ``None``)
            hypothesis (List[Hypothesis] or None): hypotheses from preceding invocation to seed
                search with. (Default: ``None``)

        Returns:
            (List[Hypothesis], List[List[torch.Tensor]]):
                List[Hypothesis]
                    top-``beam_width`` hypotheses found by beam search.
                List[List[torch.Tensor]]
                    list of lists of tensors representing transcription network
                    internal state generated in current invocation.
        r   r   r   r   rw   r   rx   ry   )rX   r:   rz   r7   rF   Ztranscribe_streamingrt   )rM   ru   rv   r1   r+   r|   rW   rU   r   r   r   Úinfer'  s    !&

zRNNTBeamSearch.infer)rD   NrE   )NN)Ú__name__Ú
__module__Ú__qualname__Ú__doc__r   rj   rb   r   r   r   rL   r$   r)   r   rV   ÚTensorr\   r   r   rg   ro   rk   rt   r{   rr   Úexportr   r}   Ú__classcell__r   r   rN   r   r	   K   sp      úù  þ
úø(ù
û)  ú
ù)Útypingr   r   r   r   r   r$   Ztorchaudio.modelsr   Ú__all__rj   r‚   rb   r   r�   r   r   r   r   r   r   r'   r)   r.   r/   r>   rC   r[   ÚModuler	   r   r   r   r   Ú<module>   s(    
*ü