_i8SSKrSSKrSSKrSSKJrJrJr SSKrSSKJ r SSK J r SSK J r SSKJr SSKJr SSKJr SSKJr S S KJrJr S S KJr \R6"\R8\ R:S 9"S S\ 55r"SS\ R>5r g)N)DictOptionalTuple)Models) TorchModel)Tensor)MODELS) update_conf)Tasks) GlobalCMVNload_kaldi_cmvn)FSMN) module_namec^\rSrSrSrSS\S\S\S\S\S\\ 4 U4S jjjr S r S \ \\ 44S jrS rSrU=r$) FSMNDecoratorz>A decorator of FSMN for integrating into modelscope framework model_dir cmvn_filebackbone input_dim output_dimtrainingc>[T U]"U/UQ70UD6 SUlSUlU(aUR X#UU5UlgU[ R RUS5S.Ulg)aBinitialize the fsmn model from the `model_dir` path. Args: model_dir (str): the model path. cmvn_file (str): cmvn file backbone (dict): params related to backbone input_dim (int): input dimension of network output_dim (int): output dimension of network training (bool): training or inference mode Nz config.yaml)model_workspace config_path)super__init__model model_cfg init_modelospathjoin) selfrrrrrrargskwargs __class__s g/var/www/html/land-ocr/venv/lib/python3.13/site-packages/modelscope/models/audio/kws/nearfield/model.pyrFSMNDecorator.__init__sd& 4T4V4  i)35DJ$-!ww||I}EDNc\[US5(aURR5 gg)Ntmp_dir)hasattrr-cleanupr%s r)__del__FSMNDecorator.__del__<s$ 4 # # LL " $r+returncpURbUbURRU5$UR$)z8 Args: input (torch.Tensor): Input tensor (B, T, D) )rforwardr )r%inputs r)r5FSMNDecorator.forward@s2 :: !e&7::%%e, ,>> !r+c xUb^[U5upV[[R"U5R 5[R"U5R 55nOSnSnSn USn USn USn USn USnUSnUSnUS nUS n[ X:XXUUUUU5 nSnSn[ X4XXUU5nU$) Ninput_affine_dim num_layers linear_dimproj_dim left_order right_order left_stride right_strideoutput_affine_dim)rr torch from_numpyfloatrKWSModel)r%rrrrmeanistd global_cmvn hidden_dim preprocessingr:r;r<r=r>r?r@rArB classifier activation kws_models r)r!FSMNDecorator.init_modelJs  (3JD$  &,,.  &,,.K K  #$67l+ l+ J'l+ }- }- / $%89 Z k;$&7E  YJ*j*N r+)rr )NNii' F)__name__ __module__ __qualname____firstlineno____doc__strdictintrboolrr1rrr5r!__static_attributes__ __classcell__r(s@r)rrsK#'"&"%#',1    !  $D> B#"S&[ 1"r+rc~^\rSrSrSrS\S\S\S\\RS\\RS\RS \RS \R4U4S jjr S r S r \ R"SSS\ RS94S\ RS\ RS\\ R\ R44SjjrSrSrU=r$)rFlarOur model consists of four parts: 1. global_cmvn: Optional, (idim, idim) 2. preprocessing: feature dimension projection, (idim, hdim) 3. backbone: backbone or feature extractor of the whole network, (hdim, hdim) 4. classifier: output layer or classifier of KWS model, (hdim, odim) 5. activation: nn.Sigmoid for wakeup word nn.Identity for speech command dataset idimodimhdimrIrKrrLrMc >[T U]5 XlX lX0lX@lXPlX`lXplXl g)a Args: idim (int): input dimension of network odim (int): output dimension of network hdim (int): hidden dimension of network global_cmvn (nn.Module): cmvn for input feature, (idim, idim) preprocessing (nn.Module): feature dimension projection, (idim, hdim) backbone (nn.Module): backbone or feature extractor of the whole network, (hdim, hdim) classifier (nn.Module): output layer or classifier of KWS model, (hdim, odim) activation (nn.Module): nn.Identity for training, nn.Sigmoid for inference N) rrr^r_r`rIrKrrLrM) r%r^r_r`rIrKrrLrMr(s r)rKWSModel.__init__ws:,    &* $$r+c6URR5$N)r to_kaldi_netr0s r)reKWSModel.to_kaldi_nets}}))++r+c8URRU5$rd)rto_pytorch_net)r% kaldi_files r)rhKWSModel.to_pytorch_nets}}++J77r+r)dtypexin_cacher3cURbURU5nURbURU5nURX5upURbURU5nURbUR U5nX4$rd)rIrKrrLrM)r%rlrm out_caches r)r5KWSModel.forwards    '  #A    )""1%A}}Q1  ?? &"A ?? &"A|r+cURbURR5 URR5 grd)rK fuse_modulesrr0s r)rrKWSModel.fuse_moduless0    )    + + - ""$r+)rMrrLrIr`r^r_rK)rPrQrRrSrTrWrnnModulerrerhrCzerosrErrr5rrrYrZr[s@r)rFrFls%%% % bii( %   * %))%II%II%@,8 "'Q1EKK!H <<,, u||U\\) * $%%r+rF)!r"systempfiletypingrrrrCtorch.nnrtmodelscope.metainformodelscope.modelsrmodelscope.models.basermodelscope.models.builderr "modelscope.utils.audio.audio_utilsr modelscope.utils.constantr cmvnr rfsmnrregister_modulekeyword_spotting"speech_kws_fsmn_char_ctc_nearfieldrrurFr+r)rs (( &(),:+- 99;QJQ;QhF%ryyF%r+