o wi@sddlmZddlmZmZddlmZmZmZddl Z ddl m Z ddl m Z eGdddZeGd d d ZGd d d eZdS) )ABC) dataclassfield)AnyOptionalTupleN) DictConfig)loggingc@sBeZdZUdZdZeed<dZeed<dZe ed<dZ e ed <d S) GraphIntersectDenseConfigz%Graph dense intersection config. g4@ search_beamg$@ output_beammin_active_statesi'max_active_statesN) __name__ __module__ __qualname____doc__r float__annotations__r rintrrrb/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/nemo/collections/asr/parts/k2/classes.pyr s    r c@seZdZUdZdZeed<dZeed<dZ e e ed<dZ eed <e d d d Zeed <dZeed<dZeed<dZeed<dS)GraphModuleConfigzQConfig for graph modules. Typically used with graph losses and decoders. default topo_typeTtopo_with_self_loopsNtoken_lmFintersect_prunedcCstS)N)r rrrr.szGraphModuleConfig.)default_factoryintersect_confg boost_coeffrpredictor_window_sizepredictor_step_size)rrrrrstrrrboolrrrrrr!r r"rr#rr%rrrrr$s      rc @sVeZdZdZddZdefddZdejdejd ejd e ejejejffd d Z d S) ASRK2Mixinack2 Mixin class that simplifies the construction of various models with k2-based losses. It does the following: - Sets up the graph loss and decoder (methods _init_k2 and update_k2_modules). - Registers external graphs, if needed. - Augments forward(...) with optional graph decoding to get accurate predictions. cCst|ds tdt|jdr|jjdurtd|jj|_|jdd}|dk|_|jrM|jjd d}|durAtd |d |d |}||jjd <||jdS) z k2-related initialization implementation. This method is expected to run after the __init__ which sets self._cfg self._cfg is expected to have the attribute graph_module_cfg _cfgz0self._cfg must be set before calling _init_k2().graph_module_cfgNz:self._cfg.graph_module_cfg must be set and cannot be None.criterion_typemlmaprzVgraph_module_cfg.backend_cfg.token_lm is empty. It must be set for criterion_type == ``z%graph_module_cfg.backend_cfg.token_lm) hasattr ValueErrorr)r*get use_graph_lm backend_cfgregister_artifactupdate_k2_modules)selfr+ token_lm_pathrrr_init_k2=s      zASRK2Mixin._init_k2 input_cfgc Cs0|`t|dr |`t|dr|jjd}n|jjd}|jddo*|jdddv}||j_ d d l m }|||j d d d |dd|dd|dd |jd|_|jj }|dk|_|dd}|rv|dkrvtd|d|dd}||_|jrd dlm}||d dddd|dd |jd|_dSdS)a  Helper function to initialize or update k2 loss and transcribe_decoder. Args: input_cfg: DictConfig to take new parameters from. Schema is expected as in nemo.collections.asr.models.configs.k2_sequence_models_config.GraphModuleConfig transcribe_decoderjointr$rTrr) forced_blankidentityr) LatticeLoss ctc_reduction mean_batchk2r+r, loss_typectcsplit_batch_size) num_classes reductionbackendr+rBrDr*r-transcribe_trainingFz,You do not need to use transcribe_training=`z4` with criterion_type=`z,`. transcribe_training will be set to False.)ViterbiDecoderWithGraphr1best)rErGdec_type return_typereturn_ilabelsoutput_alignedrDr*N)lossr/r:r;num_classes_with_blankdecoderr3r1_werremove_consecutive*nemo.collections.asr.losses.lattice_lossesr>r)r+r2r warningrH*nemo.collections.asr.modules.graph_decoderrI)r6r9rErSr>r+rHrIrrrr5Xs\              zASRK2Mixin.update_k2_modules log_probsencoded_lengthgreedy_predictionsreturncCs2|jr|jr |jr|jj||d\}}}|||fS)a k2-related post-processing parf of .forward() Args: log_probs: The log probabilities tensor of shape [B, T, D]. encoded_length: The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. greedy_predictions: The greedy token predictions of the model of shape [B, T] Returns: A tuple of 3 elements - 1) The log probabilities tensor of shape [B, T, D]. 2) The lengths of the acoustic sequence after propagation through the encoder, of shape [B]. 3) The greedy token predictions of the model of shape [B, T] (via argmax) )rWlog_probs_length)r2trainingrHr:forward)r6rWrXrY_rrr_forward_k2_post_processings  z&ASRK2Mixin._forward_k2_post_processingN) rrrrr8rr5torchTensorrr_rrrrr(4s;r()abcr dataclassesrrtypingrrrr` omegaconfr nemo.utilsr r rr(rrrrs