PhddlmZddlZddlmZmZmZddlZddlZejjdddl m cm ZgdZdZGddeZd ZGd d Zd d ef ddZy)) annotationsN)List NamedTupleUnionlibctc_prefix_decoder)CUCTCHypothesis CUCTCDecodercuda_ctc_decodercg}t|dd5}|D]4}|jj}|j|d6 ddd|S#1swY|SxYw)Nrzutf-8)encodingr)openstripsplitappend) vocab_filevocabflines vC:\Users\daisl\Desktop\realtime-object-detection\venv\Lib\site-packages\torchaudio/models/decoder/_cuda_ctc_decoder.py_get_vocab_listrs\ E j# 0AD::<%%'D LLa ! 1 L 1 Ls :AA c4eZdZUdZded< ded< ded<y) rzSRepresents hypothesis generated by CUCTC beam search decoder :class:`CUCTCDecoder`.z List[int]tokens List[str]wordsfloatscoreN)__name__ __module__ __qualname____doc____annotations__rrrs#^ h  L+r$rgffffff?cHeZdZdZdddedf d dZdZd dZy) r zCUDA CTC beam search decoder. .. devices:: CUDA Note: To build the decoder, please use the factory function :func:`cuda_ctc_decoder`. r Nc|r9t|tjjjs t d|r |j n'tjjj }tj||_ tjdtjtjd|_|dk7r t d||_||_d|_||_|dk\r|dks t dt'j(||_t-|t/||_y) ay Args: blank_id (int): token id corresopnding to blank, only support 0 for now. (Default: 0) vocab_list (List[str]): list of vocabulary tokens beam_size (int, optional): max number of hypos to hold after each decode step (Default: 10) nbest (int): number of best decodings to return blank_skip_threshold (float): skip frames if log_prob(blank) > log(blank_skip_threshold), to speed up decoding. (Default: 0.95). cuda_stream (torch.cuda.streams.Stream): using assigned cuda stream (Default: using default stream) z-cuda_stream must be torch.cuda.streams.Streamrcudadtypedevicezblank_id must be 0r'z,blank_skip_threshold must be between 0 and 1N) isinstancetorchr)streamsStreamAssertionError cuda_streamcurrent_streamcuctcprefixCTC_alloc internal_dataemptyint8r,memoryblank_id vocab_listspace_idnbestmathlogblank_skip_thresholdminlen beam_size)selfr;r:rCr=r@r2 cuda_stream_s r__init__zCUCTCDecoder.__init__3s* k5::+=+=+D+DE$%TUU2={..5::C\C\C^CjCj "22<@kk!5::ell6>RS q= !56 6  $  $).Ba.G !OP P$(HH-A$B!YJ8r$cPt tj|jyy)N)r4prefixCTC_freer6)rDs r__del__zCUCTCDecoder.__del__Ys!    !3!3 4 r$c|jtjk(s td|jtjk(s td|j r |j s td|j r|j s tdtj|j|jj|jjd|j|j|j|j|j|j|j |j" \}}|dkDrtj$|tj&|j(j+|_ tj|j|jj|jjd|j|j|j|j|j|j|j |j" \}}t-|}g}t/|D]x}|j1t/|j2D cgc]F} t5||| d||| dD cgc]} |j6| c} ||| dHc} } z|Scc} wcc} } w) a Args: log_prob (torch.FloatTensor): GPU tensor of shape `(batch, frame, num_tokens)` storing sequences of probability distribution over labels; log_softmax(output of acoustic model). lengths (dtype torch.int32): GPU tensor of shape `(batch, )` storing the valid length of in time axis of the output Tensor in each batch. Returns: List[List[CUCTCHypothesis]]: List of sorted best hypotheses for each audio sequence in the batch. z$encoder_out_lens must be torch.int32zlog_prob must be torch.float32zinputs must be cuda tensorsz input tensors must be contiguousrr*r')rrr)r+r.int32r1float32is_cuda is_contiguousr4$ctc_beam_search_decoder_batch_gpu_v2r6r9data_ptrsizestriderCr:r<r@r7r8r, contiguousrBrangerr=rr;) rDlog_probencoder_out_lens required_size score_hyps_ batch_sizehyposijword_ids r__call__zCUCTCDecoder.__call__]s| %%4 !GH H~~. !AB B  %5%=%= !>? ?&&(-=-K-K-M !CD D$)$N$N    KK " KK  Q       % % ' MMO OO  NN MM MM  % % % ! z 1 ++m5::hoo^iikDK!FF"" $$&   #!!# ))+ !  )) MAz_ z"A LL#4::.  / $)!}Q/2GQRS}UVGWXYGZ[GZGtw7GZ[(mA.q1 /  # \s !K< K7K< 7K< ) r;rr:intrCr`r=r`r@rr2ztorch.cuda.streams.Stream)rU torch.TensorrVra)rrr r!_DEFAULT_BLANK_SKIP_THREASHOLDrFrIr_r#r$rr r *sb&D15$9$9$9 $9  $9 $ $9/$9L5=r$r r'r&cZt|tk(r t|}t||||S)aBuilds an instance of :class:`CUCTCDecoder`. Args: tokens (str or List[str]): File or list containing valid tokens. If using a file, the expected format is for tokens mapping to the same index to be on the same line beam_size (int, optional): The maximum number of hypos to hold after each decode step (Default: 10) nbest (int): The number of best decodings to return blank_id (int): The token ID corresopnding to the blank symbol. blank_skip_threshold (float): skip frames if log_prob(blank) > log(blank_skip_threshold), to speed up decoding (Default: 0.95). Returns: CUCTCDecoder: decoder Example >>> decoder = cuda_ctc_decoder( >>> vocab_file="tokens.txt", >>> blank_skip_threshold=0.95, >>> ) >>> results = decoder(log_probs, encoder_out_lens) # List of shape (B, nbest) of Hypotheses )r;rCr=r@)typestrrr )rr=rCr@s rr r s,6 F|s ( 6Yebv wwr$) rzUnion[str, List[str]]r=r`rCr`r@rreturnr ) __future__rr>typingrrrr. torchaudio _extension _load_lib!torchaudio.lib.pybind11_prefixctclibpybind11_prefixctcr4__all__rrrbr r r#r$rrps" **   7811 B ,j ,"&ppj"@ x !x xx x  xr$