~Vji'2 ddlmZmZmZmZmZddlZddlmZddgZ eee ej eeej e fZ de _de dee fd Zde dej fd Zde deeej fd Zde de fd Zde defd Zdee deeej fdZdeeej de dejdeeej fdZde de fdZdee dej de deej ej ej ffdZde dee ddfdZGddejjZdS))CallableDictListOptionalTupleN)RNNT HypothesisRNNTBeamSearchzHypothesis generated by RNN-T beam search decoder, represented as tuple of (tokens, prediction network output, prediction network state, score). hyporeturnc|dSNrr s X/root/voice-cloning/.venv/lib/python3.11/site-packages/torchaudio/models/rnnt_decoder.py_get_hypo_tokensr 7Nc|dSNrrs r_get_hypo_predictor_outrrrc|dS)Nrrs r_get_hypo_staterrrc|dS)Nrrs r_get_hypo_scorerrrc,t|dSr)strrs r _get_hypo_keyr! s tAw<<rhyposc xg}ttt|dD]g}ttt|dD]6|t jfd|D7|||S)NrcFg|]}t|Sr)r).0r ijs r z _batch_state..)s-6e6e6eW[t7L7LQ7OPQ7R6e6e6er)rangelenrappendtorchcat)r"statesbatched_state_componentsr&r's @@r _batch_stater0$s')F 3uQx0011 2 20079 s?5844Q78899 h hA $ + +EI6e6e6e6e6e_d6e6e6e,f,f g g g g .//// Mrr.idxdevicecNtj|g|fd|DS)Nr2c,g|]}fd|DS)c<g|]}|dS)r) index_select)r%state idx_tensors rr(z+_slice_state...0s) H H H5U  : . . H H Hrr)r% state_tupler9s rr(z _slice_state..0s. c c c[ H H H HK H H H c c cr)r,tensor)r.r1r2r9s @r _slice_stater<.s4seF333J c c c c\b c c ccrc`t|tt|dzz Sr)rr*rrs r_default_hypo_sort_keyr>3s+ 4 C(8(>(>$?$?!$C DDrnext_token_probs beam_widthcJtjd|Dd}||ddddfz}|d|\}}||jdd}||jdz}|||fS)Nc,g|]}t|Srrr%hs rr(z+_compute_updated_scores..<s BBBq 2 2BBBrrtrunc) rounding_mode)r,r; unsqueezereshapetopkdivshape) r"r?r@ hypo_scoresnonblank_scoresnonblank_nbest_scoresnonblank_nbest_idxnonblank_nbest_hypo_idxnonblank_nbest_tokens r_compute_updated_scoresrT7s ,BBEBBBCCMMaPPK!$4QQQV$<rarb)selfr^r_r`rarb __class__s rrezRNNTBeamSearch.__init__\sX   &  !7D  !.D .rr2c|j}d}tjdg|}|jtj|gg|||\}}}|g|d|df}|gS)Nrr4rg)r_r,r;r^predictdetach) rfr2tokenr8 one_tensorpred_out_ pred_state init_hypos r _init_b_hyposzRNNTBeamSearch._init_b_hyposps \1#f555 "&*"4"4U\E7)TZ5[5[5[]gin"o"o!Z G QK      {renc_outr"c rtjdg|}tjd|Dd}|j|||tjdgt |z|\}}}tjj||j z d}|ddddfS)Nrr4c,g|]}t|Sr)rrDs rr(z8RNNTBeamSearch._gen_next_token_probs..s!$O$O$OA%.s %P%P%Pod&;&;%P%P%Prc g|] }| Srrr%r1r~s rr(z/RNNTBeamSearch._gen_b_hypos..s333 333r)r)r*rr!rZfloatr,r; logaddexprrrr+sort) rfr~rr?rr&h_aappend_blank_scoreh_bscorern sorted_idxs ` r _gen_b_hyposzRNNTBeamSearch._gen_b_hypossWs7||$$ 4 4A!*C!0!5!58HB8O!O S!!]22#M#$6$67S'***el?3+?+?@@JJK]^^__011 %%',,$$ C NN3   03M-,, - - %P%P%P%P%PQQVVXX :3333 3333rtr@c t|||\}}} t||krtd } nt|| } g} g} g} t |D]}t||}|| krmt ||}| ||| t | || || r|| | | ||}ng}|S)Ninf)rTr*rrr)intr+_gen_new_hypos)rfrr~r?rr@r2rPrRrS b_nbest_score base_hypos new_tokens new_scoresr&r a_hypo_idx new_hyposs r _gen_a_hyposzRNNTBeamSearch._gen_a_hyposs1 $G-=z J J  ! #  w<<* $ $"5\\MMM+GZK,@AAM') " "$ z"" ) )A/233E}$$ !8!;<< !!'*"5666!!#&:1&=">">???!!%(((  -++J JPQSYZZII*,Irrtokensscoresc tjd|D|}t|}|j|tjdgt |z||\}} } g} t |D]d\} } t| || gz}| |||  t| | ||| fe| S)Ncg|]}|gSrr)r%rks rr(z1RNNTBeamSearch._gen_new_hypos..s"?"?"?uE7"?"?"?rr4r) r,r;r0r^rir*rXrr+rjr<)rfrrrrr2 tgt_tokensr.rmrn pred_statesrr&rrs rrzRNNTBeamSearch._gen_new_hyposs\"?"?"?"?"?OOO j))#':#5#5  L!s:.v > > > $ $  ![ ')  ++ r rFAs)#..&)*> [Z[]c@d@dflmnfop q q q qrr c  |jd}|j}g}||n| t|D]} }tjttg i}d} |r |dd||dzf||} | } || | | j krn# | | |||}|r| dz } |t jfd D|\} } fd| D  S)Nrrc:g|]}|Sr)ra)r%hyprfs rr(z*RNNTBeamSearch._search..s')U)U)Uc$*<*.s:::ws|:::r)rMr2rqr)r,jitannotaterr r}cpurrbrr;rK)rfrrr r@ n_time_stepsr2rrrsymbols_current_tr?rnrr~s` @r_searchzRNNTBeamSearch._searchs }Q' $&04 $$$V,,,$|$$ ; ;AGi((j)92>>G35M !  +#'#=#=gaaaQQRUl>SU\^d#e#e #3#7#7#9#9 ++GW>NP]^^$(<<<++$ +%*%# +&"L)U)U)U)UW)U)U)UVV[[\fggMAz::::z:::GGrinputlengthc |dkr8|dkr|jddkstd|dkr|d}|jdkr|jdkrtd|dkr|d}|j||\}}||d |S) aPerforms 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. rrrr*input must be of shape (T, D) or (1, T, D)rr"length must be of shape () or (1,)N)rurM ValueErrorrIr^ transcriber)rfrrr@rrrns rforwardzRNNTBeamSearch.forward s 99;;!  UYY[[A%5%5%+a.A:M:MIJJ J 99;;!  OOA&&E <2  &,$"6"6ABB B ::<<1  %%a((FZ**5&99 ||GT:666rr8 hypothesisc|dkr8|dkr|jddkstd|dkr|d}|jdkr|jdkrtd|dkr|d}|j|||\}}}|||||fS) aPerforms 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. rrrrrrrr)rurMrrIr^transcribe_streamingr)rfrrr@r8rrrrns rinferzRNNTBeamSearch.infer'sB 99;;!  UYY[[A%5%5%+a.A:M:MIJJ J 99;;!  OOA&&E <2  &,$"6"6ABB B ::<<1  %%a((F J;;E65QQE||GZ< /  / //////( EL T*5E     #| #,0,< #FKl #  # # # #4j!4j!4 , 4 CO, 4 j  44446$j!$j!$ , $  $  $ $ j $$$$L$S U      j *''tJ'(' ' j  ''''R7U\75<7S7UYZdUe77778 Y 5915 +D+D|+D +D +D T%,/01 +D T*-. +D tJd5<&8!99 :+D+D+D+D+D+D+D+Dr)typingrrrrrr,torchaudio.modelsr__all__rrrr rrrrrr r!r0r2r<r>rTrZrxModuler rrrrs88888888888888 """""" ) *49elDel1C,DeK L   :$s)**d5<.@)A* sZ(T$u|2D-Edd5<01ddU\dVZ[_`e`l[mVndddd EEEEEE P   Pl P P 5<u| 34 P P P Pzd:.>4HDHDHDHDHDUX_HDHDHDHDHDr