o wi#@sddlZddlZddlZddlmZzddlmZWne y)ddlm ZYnwddl Z ddl m Z ddlmZddlmZddlmZmZmZmZdd lmZed d Gd d d Zdedede jfddZdededefddZdedefddZ Gddde!Z"dS)N) dataclass)_read_fileobject)_validate_fileobject_and_memmap)LogisticRegression)Pipeline)StandardScaler)ConfidenceConfigConfidenceMethodConfigget_confidence_aggregation_bankget_confidence_measure_bank) HypothesisT)frozenc@s<eZdZUeed<eed<eed<eed<defddZdS) ConfidenceSpec exclude_blank aggregationconfidence_typealphareturnc CsJ|jdkr d}d}d}n |jd\}}}t|j|jt|||j|ddS)aIConverts confidence spec to the confidence config. Internally, the tuning procedure uses this "spec" objects as they are more aligned with how things are implemented. But when it's time to save the models or call transcribe, we need to use the proper object of type ``ConfidenceConfig``. max_probtsallislin_)name entropy_typer entropy_norm)rr method_cfg)rsplitrrrr r)selfrrrrl/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/nemo/collections/asr/models/confidence_ensemble.pyto_confidence_config0s z#ConfidenceSpec.to_confidence_configN) __name__ __module__ __qualname__bool__annotations__strfloatrr rrrrr)s r hypothesisrrcCs$t|jtrKg}|jD]'}|D]"}|s||dq|d|djddkr1||dqq |s<||dt|}tj rJ|}n/|j }tj rW|}|rx|j dd}|||jddk}|jddkrw|dd}n|}t |d tdstj|dd}|S)aReturns logprobs from the hypothesis object with optional blanks filter. This function supports both CTC and Transducer hypotheses. Will place the logprobs on GPU if it's available. Args: hypothesis: generated hypothesis as returned from the transcribe method of the ASR model. exclude_blank: whether to filter out all ```` tokens. Returns: torch.Tensor: of shape [S, V], where S is (filtered) sequence length and V is the vocabulary size. r)dimN?) isinstance alignmentslistappenditemshapetorchstackcuda is_available y_sequenceargmaxallcloseexpsumtensor log_softmax)r(rfiltered_logprobs alignment align_elemlogprobslabelsrrrget_filtered_logprobsJs:       rCconfidence_cfgc Cst||j}|jd}t|j}|jjdkrd}d}nd|jjd|jj}|jj }t |}|||||d }|S)aUComputes confidence score of the full utterance from a given hypothesis. This is essentially a re-implementation of the built-in confidence computation in NeMo. The difference is that we aggregate full-utterance scores, while core functionality only supports word and token level aggregations. Args: hypothesis: generated hypothesis as returned from the transcribe method of the ASR model. confidence_cfg: confidence config specifying what kind of method/aggregation should be used. Returns: float: confidence score. r)rr,entropy_r)vt) rCrr2r rrrrrrr cpur1) r(rDr> vocab_size aggr_func conf_typer conf_func conf_valuerrrcompute_confidenceys     rN file_pathc stj|s td|hdGfdddtj}zttdzt |dO}t ||dd8}t |t r?|d }t |t r`t |d}||}Wdn1sZwYn||}Wdn1spwYWdn1swYt |tstd |jD]\}}t |ttfstd t|qWntjtfy}ztd |d}~wwWd|WS1swY|WSty}z td t |d}~ww) ab Safely load a joblib file containing a scikit-learn pipeline. Args: file_path: Path to the joblib file Returns: Pipeline: A scikit-learn pipeline object Raises: ValueError: If the file doesn't exist or contains unauthorized content SecurityError: If the file contains potentially malicious content zModel file not found: > numpy.dtype numpy.ndarraysklearn.pipeline.Pipeline*sklearn.preprocessing._data.StandardScaler1sklearn.linear_model._logistic.LogisticRegression numpy._picklecseZdZfddZZS)z-safe_joblib_load..RestrictedUnpicklercsN|d|}|vr|dkrddl}t||St||Std|d)N.rUrzUnauthorized class z in joblib file)numpygetattrsuper find_class SecurityError)rmoduler class_pathnp)ALLOWED_CLASSES __class__rrrZs z8safe_joblib_load..RestrictedUnpickler.find_class)r!r"r#rZ __classcell__rr_)r`rRestrictedUnpicklersrcignorerbN) mmap_moderz,Loaded model must be a scikit-learn PipelinezUnauthorized pipeline step: zFailed to safely load model: )ospathexists ValueErrorpickle Unpicklerwarningscatch_warnings simplefilteropen_validate_joblib_filer-tupler&loadr named_stepsitemsrrtypeUnpicklingErrorAttributeErrorr[ Exception) rOrcrawfstreamfmodel step_namestep_objerrbrsafe_joblib_loadsT           rc@seZdZdZdS)r[z-Custom exception for security-related errors.N)r!r"r#__doc__rrrrr[sr[)#os.pathrgrkrm dataclassesrjoblib.numpy_pickle_utilsrrq ImportErrorrr3sklearn.linear_modelrsklearn.pipelinersklearn.preprocessingr5nemo.collections.asr.parts.utils.asr_confidence_utilsrr r r +nemo.collections.asr.parts.utils.rnnt_utilsr rr$TensorrCr'rNr&rryr[rrrrs*       /"K