o Gi&!@szddlmZWneyddlmZYnwddlmZddlZddlmZddlm Z ddl m Z ddl m Z ddlmZGd d d ejZ  dde de dedededdedeejdejfddZdS))Literal)OptionalN)intersect_dense) DenseFsaVec)Fsa) RaggedTensorc sdeZdZdZ  ddededdeffdd Z dd ed e dede e j de j f ddZ ZS)CtcLossaCtc Loss computation in k2. It produces the same output as `torch.CtcLoss` if given the same input. One difference between `k2.CtcLoss` and `torch.CtcLoss` is that k2 accepts a general FSA while PyTorch requires a linear FSA (represented as a list). That means, `k2.CtcLoss` supports words with multiple pronunciations. See `k2/python/tests/ctc_loss_test.py `_ for usage. We assume that the blank label is always 0. The arguments `reduction` and `target_lengths` have the same meaning as their counterparts in `torch.CtcLoss`. sumT output_beam reductionnonemeanr use_double_scorescs,t|dvs J||_||_||_dS)a Args: output_beam: Beam to prune output, similar to lattice-beam in Kaldi. Relative to best path of output. reduction: Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the output losses will be **divided** by the target lengths and then the **mean** over the batch is taken. 'sum': sum the output losses over batches. use_double_scores: True to use double precision floating point in computing the total scores. False to use single precision. r N)super__init__r r r)selfr r r __class__?/home/ubuntu/.local/lib/python3.10/site-packages/k2/ctc_loss.pyr#s   zCtcLoss.__init__Ndecoding_graph dense_fsa_vec delay_penaltytarget_lengthsreturnc CsRt|||j|dkr dndd}|dkrxt|jd|j}|j|j }|j |d?dd}| |j } t |drQt |jtjsCJ|jd k|jdkB|jB} nt |d sXJt |jtjsaJ|jd k|jdkB} | | d|j| 7_|jd |jd } d| } | tj} |jd kr| S|jdkr| S|jdksJ| |} | S)abCompute the CTC loss given a decoding graph and a dense fsa vector. Args: decoding_graph: An FsaVec. It can be the composition result of a CTC topology and a transcript. dense_fsa_vec: It represents the neural network output. Refer to the help information in :class:`k2.DenseFsaVec`. delay_penalty: A constant to penalize symbol delay, which is used to make symbol emit earlier for streaming models. It is almost the same as the `delay_penalty` in our `rnnt_loss`, See https://github.com/k2-fsa/k2/issues/955 and https://arxiv.org/pdf/2211.00490.pdf for more details. target_lengths: Used only when `reduction` is `mean`. It is a 1-D tensor of batch size representing lengths of the targets, e.g., number of phones or number of word pieces in a sentence. Returns: If `reduction` is `none`, return a 1-D tensor with size equal to batch size. If `reduction` is `mean` or `sum`, return a scalar. r frame_idxN)a_fsasb_fsasr frame_idx_namer)valuealpha_is_repeat_token_r aux_labelsT) log_semiringrrr r)rr rarcsshape remove_axisrdurationtodeviceaddvalueshasattr isinstancer%torchTensorlabelsr& masked_fill_scoresget_tot_scoresrfloat32r r r) rrrrrlatticerr+offsetpenaltymask tot_scoreslossrrrforward>sL       zCtcLoss.forward)r T)rN)__name__ __module__ __qualname____doc__floatrboolrrrrr2r3r? __classcell__rrrrr s0r rr Trrr rr r rrrcCst|||d}|||||dS)aVCompute the CTC loss given a decoding graph and a dense fsa vector. Args: decoding_graph: An FsaVec. It can be the composition result of a ctc topology and a transcript. dense_fsa_vec: It represents the neural network output. Refer to the help information in :class:`k2.DenseFsaVec`. output_beam: Beam to prune output, similar to lattice-beam in Kaldi. Relative to best path of output. delay_penalty: A constant to penalize symbol delay, which is used to make symbol emit earlier for streaming models. It is almost the same as the `delay_penalty` in our `rnnt_loss`, See https://github.com/k2-fsa/k2/issues/955 and https://arxiv.org/pdf/2211.00490.pdf for more details. reduction: Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the output losses will be divided by the target lengths and then the mean over the batch is taken. 'sum': sum the output losses over batches. use_double_scores: True to use double precision floating point in computing the total scores. False to use single precision. target_lengths: Used only when `reduction` is `mean`. It is a 1-D tensor of batch size representing lengths of the targets, e.g., number of phones or number of word pieces in a sentence. Returns: If `reduction` is `none`, return a 1-D tensor with size equal to batch size. If `reduction` is `mean` or `sum`, return a scalar. )r r r)rrrr)r )rrr rr rrmrrrctc_losss+rI)rGrr TN)typingr ImportErrortyping_extensionsrr2torch.nnnnautogradrrrfsarraggedrModuler rDrEr3rIrrrrsH