o i@s>ddlZddlmmZddlmZGdddejjZdS)N)check_argument_typescszeZdZdZ     ddededed ed ed ed effd d Zde j fddZ ddZ ddZ ddZddZZS)CTCaCTC module. Args: odim: dimension of outputs encoder_output_size: number of encoder projection units dropout_rate: dropout rate (0.0 ~ 1.0) ctc_type: builtin or gtnctc reduce: reduce the CTC loss into a scalar ignore_nan_grad: Same as zero_infinity (keeping for backward compatiblity) zero_infinity: Whether to zero infinite losses and the associated gradients. builtinTNodimencoder_output_size dropout_ratectc_typereduceignore_nan_grad zero_infinityc stsJt|}||_tj|||_||_|dur |}|jdkr/tjj d|d|_ n|jdkr?ddl m } | j |_ ntd|j||_dS)Nrnone) reductionr gtnctcr)GTNCTCLossFunctionz(ctc_type must be "builtin" or "gtnctc": )rsuper__init__rtorchnnLinearctc_lor CTCLossctc_loss#espnet.nets.pytorch_backend.gtn_ctcrapply ValueErrorr ) selfrrrr r r r eprojsr __class__C/home/ubuntu/.local/lib/python3.10/site-packages/espnet2/asr/ctc.pyrs"      z CTC.__init__returncCs|jdkr(|d}|||||}|d}|jr"||}|S||}|S|jdkr?tjjj|dd}||||ddSt )Nrrdimrr ) r log_softmaxrsizer sumrr functionalNotImplementedError)rth_pred th_targetth_ilenth_olenlossr( log_probsr r r!loss_fn4s     z CTC.loss_fncsz|tj||jd}|jdkrddD}n|dd}tfddt|D}| ||||j |j |j d}|S) a@Calculate CTC loss. Args: hs_pad: batch of padded hidden state sequences (B, Tmax, D) hlens: batch of lengths of hidden state sequences (B) ys_pad: batch of padded character id sequence tensor (B, Lmax) ys_lens: batch of lengths of character sequence (B) )prcSsg|]}||dkqS)r ).0yr r r! VszCTC.forward..rr$cs g|] \}}|d|fqS)Nr )r5ilys_padr r!r7[s )devicedtype) rFdropoutrr transposercat enumerater2tor<r=)rhs_padhlensr;ys_lensys_hatys_truer0r r:r!forwardHs  z CTC.forwardcCtj||ddS)zsoftmax of frame activations Args: Tensor hs_pad: 3d tensor (B, Tmax, eprojs) Returns: torch.Tensor: softmax applied 3d tensor (B, Tmax, odim) r#r%)r>softmaxrrrDr r r!rKcz CTC.softmaxcCrJ)zlog_softmax of frame activations Args: Tensor hs_pad: 3d tensor (B, Tmax, eprojs) Returns: torch.Tensor: log softmax applied 3d tensor (B, Tmax, odim) r#r%)r>r'rrLr r r!r'mrMzCTC.log_softmaxcCrJ)zargmax of frame activations Args: torch.Tensor hs_pad: 3d tensor (B, Tmax, eprojs) Returns: torch.Tensor: argmax applied 2d tensor (B, Tmax) r#r%)rargmaxrrLr r r!rNwrMz CTC.argmax)rrTNT)__name__ __module__ __qualname____doc__intfloatstrboolrrTensorr2rIrKr'rN __classcell__r r rr!rs6!  r) rtorch.nn.functionalrr*r> typeguardrModulerr r r r!s