o Gi)@szddlmZWneyddlmZYnwddlmZmZddlZddlZGdddejj Z    dd ed d eej ej ffddZ dS))Literal)ListUnionNc sxeZdZdZ   ddedededffd d Zd ej d e ej e e e fd ede de ejej ff ddZZS)MWERLosszMinimum Word Error Rate Loss compuration in k2. See equation 2 of https://arxiv.org/pdf/2106.02302.pdf about its definition. ?Tsum temperatureuse_double_scores reductionnonemeanrcs0|dvsJ|t||_||_||_dS)a Args: temperature: For long utterances, the dynamic range of scores will be too large and the posteriors will be mostly 0 or 1. To prevent this it might be a good idea to have an extra argument that functions like a temperature. We scale the logprobs by before doing the normalization. use_double_scores: True to use double precision floating point. False to use single precision. reduction: Specifies the reduction to apply to the output: 'none' | 'sum' | 'mean'. 'none': no reduction will be applied. The returned 'loss' is a k2.RaggedTensor, with loss.tot_size(0) == batch_size. loss.tot_size(1) == total_num_paths_of_current_batch If you want the MWER loss for each utterance, just do: `loss_per_utt = loss.sum()` Then loss_per_utt.shape[0] should be batch_size. See more example usages in 'k2/python/tests/mwer_test.py' 'sum': sum loss of each path over the whole batch together. 'mean': divide above 'sum' by total num paths over the whole batch. r N)super__init__rr r )selfrr r  __class__@/home/ubuntu/.local/lib/python3.10/site-packages/k2/mwer_loss.pyrs   zMWERLoss.__init__lattice ref_texts nbest_scale num_pathsreturncCstjj|||j|d}|jj}|jj|}|j|}| } tj || jd} tj | | |j ddd} | j |jdd} | | } t||jj}||j}t||}|jddj}|| }|jd krp|}|S|jd kr{|}|St||}|S) aiCompute the Minimum Word Error loss given a lattice and corresponding ref_texts. Args: lattice: An FsaVec with axes [utt][state][arc]. ref_texts: It can be one of the following types: - A list of list-of-integers, e..g, `[ [1, 2], [1, 2, 3] ]` - An instance of :class:`k2.RaggedTensor`. Must have `num_axes == 2` and with dtype `torch.int32`. nbest_scale: Scale `lattice.score` before passing it to :func:`k2.random_paths`. A smaller value leads to more unique paths at the risk of being not to sample the path with the best score. num_paths: Number of paths to **sample** from the lattice using :func:`k2.random_paths`. Returns: Minimum Word Error Rate loss. )rrr r)deviceT)refshypshyp_to_ref_mapsorted_match_refF)r log_semiring)use_logrr )k2Nbest from_latticer scoresr kept_pathshapetobuild_levenshtein_graphslevenshtein_graphlevenshtein_alignmentrow_idsget_tot_scores RaggedTensorfsarr normalizevaluesexpr r )rrrrrnbestrpath_arc_shapestream_path_shaperrr+ tot_scoreswersragged_nbest_logp path_logpragged_path_logpprob_normalizedlossrrrforward8sD       zMWERLoss.forward)rTr)__name__ __module__ __qualname____doc__floatboolrrr"Fsarr.rinttorchTensorr= __classcell__rrrrrs.$r?rTrr r rcCs*|dvsJ|t|||}|||||S)aCompute the Minimum loss given a lattice and corresponding ref_texts. Args: lattice: An FsaVec with axes [utt][state][arc]. ref_texts: It can be one of the following types: - A list of list-of-integers, e..g, `[ [1, 2], [1, 2, 3] ]` - An instance of :class:`k2.RaggedTensor`. Must have `num_axes == 2` and with dtype `torch.int32`. nbest_scale: Scale `lattice.score` before passing it to :func:`k2.random_paths`. A smaller value leads to more unique paths at the risk of being not to sample the path with the best score.'' num_paths: Number of paths to **sample** from the lattice using :func:`k2.random_paths`. temperature: For long utterances, the dynamic range of scores will be too large and the posteriors will be mostly 0 or 1. To prevent this it might be a good idea to have an extra argument that functions like a temperature. We scale the logprobs by before doing the normalization. use_double_scores: True to use double precision floating point. False to use single precision. reduction: Specifies the reduction to apply to the output: 'none' | 'sum' | 'mean'. 'none': no reduction will be applied. The returned 'loss' is a k2.RaggedTensor, with loss.tot_size(0) == batch_size. loss.tot_size(1) == total_num_paths_of_current_batch If you want the MWER loss for each utterance, just do: `loss_per_utt = loss.sum()` Then loss_per_utt.shape[0] should be batch_size. See more example usages in 'k2/python/tests/mwer_test.py' 'sum': sum loss of each path over the whole batch together. 'mean': divide above 'sum' by total num paths over the whole batch. Returns: Minimum Word Error Rate loss. r )r)rrrrrr r mrrr mwer_losss2 rL)rIrJrTr)typingr ImportErrortyping_extensionsrrrFr"nnModulerrGr.rLrrrrs& s