a  dD@sddlZddlZddlmZddlmZmZmZmZm Z ddl m Z m Z m Z ddlmZddlmZeeeddd Zd)eeeeeeedddZeedddZeedddZee e dddZe eeefee eeefe ededfee eeefee eeeffdddZd*ee eeefeeeee ddd Zd+e eeefee eeefed!eeeeeee ee d" d#d$Zd,e eeefee eeefed!eeeeee e ee e ffd& d'd(ZdS)-N)inf)ListOptionalSequenceTupleUnion)Tensorstacktensor)Literal)_validate_inputs) preds_word target_wordreturncCs t||kS)a.Distance measure used for substitutions/identity operation. Code adapted from https://github.com/rwth-i6/ExtendedEditDistance/blob/master/EED.py. Args: preds_word: hypothesis word string target_word: reference word string Return: 0 for match, 1 for no match )int)r rri/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/torchmetrics/functional/text/eed.py_distance_between_wordsds r@333333?皙??)hyprefalpharhodeletion insertionrc sjdgt|d}dgt|d}d|d<tgt|d}tdt|dD]} tdt|dD]d} | dkrt|| d||| dt|| d|| d|| ||| <qf|| d|| <qf|t|} || d7<|| ddkr||| fdd|D}|}tgt|d}qP|td d |D} td|d| tt|| S) a8Compute extended edit distance score for two lists of strings: hyp and ref. Code adapted from: https://github.com/rwth-i6/ExtendedEditDistance/blob/master/EED.py. Args: hyp: A hypothesis string ref: A reference string alpha: optimal jump penalty, penalty for jumps between characters rho: coverage cost, penalty for repetition of characters deletion: penalty for deletion of character insertion: penalty for insertion or substitution of character Return: Extended edit distance score as float rr csg|]}t|qSr)min.0xZjumprr z!_eed_function..css|]}|dkr|ndVqdS)rrNrr#rrr r(z _eed_function..)lenrranger"rindexsumfloat) rrrrrrZnumber_of_visitsrowZnext_rowwiZ min_indexZcoveragerr&r _eed_functionts,$  r2)sentencercCst|tstdt|d|}gd}|D]\}}|||}q2gd}|D]\}}t|||}qTgd}|D]\}}|||}qxd|dS)zPreprocess english sentences. Copied from https://github.com/rwth-i6/ExtendedEditDistance/blob/master/util.py. Raises: ValueError: If input sentence is not of a type `str`. 6Only strings allowed during preprocessing step, found  instead)).z .)!z !)?z ?),z ,))z\s+r!)z(\d) ([.,]) (\d)z\1\2\3)z#(Dr|Jr|Prof|Rev|Gen|Mr|Mt|Mrs|Ms) .z\1.))ze . g .ze.g.)zi . e .zi.e.)zU . S .zU.S.r!) isinstancestr ValueErrortyperstripreplaceresub)r3Zrules_interpunctionpattern replacementZrules_rerrr_preprocess_ens    rDcCs2t|tstdt|d|}td|S)zPreprocess japanese sentences. Copy from https://github.com/rwth-i6/ExtendedEditDistance/blob/master/util.py. Raises: ValueError: If input sentence is not of a type `str`. r4r5NFKC)r:r;r<r=r> unicodedata normalize)r3rrr_preprocess_jas rH)sentence_level_scoresrcCs(t|dkrtdSt|tt|S)zReduction for extended edit distance. Args: sentence_level_scores: list of sentence-level scores as floats Return: average of scores as a tensor rr )r*r r-)rIrrr _eed_computes rJenja)predstargetlanguagercsft||d\}}|dkrtn|dkr,tntd|fdd|D}fdd|D}||fS)auPreprocess strings according to language requirements. Args: preds: An iterable of hypothesis corpus. target: An iterable of iterables of reference corpus. language: Language used in sentences. Only supports English (en) and Japanese (ja) for now. Defaults to en Return: Tuple of lists that contain the cleaned strings for target and preds Raises: ValueError: If a different language than ``'en'`` or ``'ja'`` is used ValueError: If length of target not equal to length of preds ValueError: If objects in reference and hypothesis corpus are not strings )Zhypothesis_corpusZ ref_corpusrKrLz?Expected argument `language` to either be `en` or `ja` but got csg|] }|qSrr)r$predZpreprocess_functionrrr'r(z)_preprocess_sentences..csg|]}fdd|DqS)csg|] }|qSrr)r$rrQrrr'r(z4_preprocess_sentences...r)r$ referencerQrrr'r()r rDrHr<)rMrNrOrrQr_preprocess_sentencessrS)r target_wordsrrrrrc Cs4t}|D]"}t||||||}||kr|}qt|S)aCompute scores for ExtendedEditDistance. Args: target_words: An iterable of reference words preds_word: A hypothesis word alpha: An optimal jump penalty, penalty for jumps between characters rho: coverage cost, penalty for repetition of characters deletion: penalty for deletion of character insertion: penalty for insertion or substitution of character Return: best_score: best (lowest) sentence-level score as a Tensor )rr2r ) r rTrrrrZ best_scorerRscorerrr_compute_sentence_statistics"s rV)rKrL) rMrNrOrrrr sentence_eedrc Cslt|||\}}|durg}dt|t|dfvr8|St||D]$\}} t|| ||||} || qB|S)aCompute scores for ExtendedEditDistance. Args: preds: An iterable of hypothesis corpus target: An iterable of iterables of reference corpus language: Language used in sentences. Only supports English (en) and Japanese (ja) for now. Defaults to en alpha: optimal jump penalty, penalty for jumps between characters rho: coverage cost, penalty for repetition of characters deletion: penalty for deletion of character insertion: penalty for insertion or substitution of character sentence_eed: list of sentence-level scores Return: individual sentence scores as a list of Tensors Nr)rSr*ziprVappend) rMrNrOrrrrrWZ hypothesisrTrUrrr _eed_updateBs rZF) rMrNrOreturn_sentence_level_scorerrrrrc Cs|tgd||||gD]4\}} t| tr:t| tr| dkrtd|dqt|||||||} t| } |rx| t| fS| S)uXCompute extended edit distance score (`ExtendedEditDistance`_) [1] for strings or list of strings. The metric utilises the Levenshtein distance and extends it by adding a jump operation. Args: preds: An iterable of hypothesis corpus. target: An iterable of iterables of reference corpus. language: Language used in sentences. Only supports English (en) and Japanese (ja) for now. Defaults to en return_sentence_level_score: An indication of whether sentence-level EED score is to be returned. alpha: optimal jump penalty, penalty for jumps between characters rho: coverage cost, penalty for repetition of characters deletion: penalty for deletion of character insertion: penalty for insertion or substitution of character Return: Extended edit distance score as a tensor Example: >>> from torchmetrics.functional.text import extended_edit_distance >>> preds = ["this is the prediction", "here is an other sample"] >>> target = ["this is the reference", "here is another one"] >>> extended_edit_distance(preds=preds, target=target) tensor(0.3078) References: [1] P. Stanchev, W. Wang, and H. Ney, “EED: Extended Edit Distance Measure for Machine Translation”, submitted to WMT 2019. `ExtendedEditDistance`_ )rrrrrz Parameter `z)` is expected to be a non-negative float.)rXr:r.r<rZrJr ) rMrNrOr[rrrr param_nameparamrIZaveragerrrextended_edit_distancels( r^)rrrr)rrrr)rKrrrrN)rKFrrrr)r@rFmathrtypingrrrrrZtorchrr r Ztyping_extensionsr Z#torchmetrics.functional.text.helperr r;rrr.r2rDrHrJrSrVrZboolr^rrrrYs    :-& ) #  -