9i.SSKrSSKrSSKrSSKrSSKJr SSKrSSKrSSK r SSK r SSK J r SSK Jr SSKJrJr SSKJr SSKJr SS KJrJrJr SS KJr SS KJr "S S 5rg)N)path)Image)File) OFATokenizerOFATokenizerZH) load_image)Trie)_get_kaldi_fbank_get_torchaudio_fbankconvert_waveform)OFA_TASK_KEY_MAPPING)set_torch_seedc\rSrSrSrSrSSjr\SSj5r\S5r Sr S r S r SS jr S \R4S jrSrg)OfaBasePreprocessorz OFA base preprocessor for c XlX0lURRRSS5Ul[ R RU5(a[ R RU5nURS:Xa[R"U5nO-URS;a[R"U5nO[eUR[S5Vs/sHnSRU5PM sn5 UR[S5Vs/sHnSRU5PM sn5 URRRSS 5S :XaURS S /5 X`l["R$"UR&/5Ul["R$"UR*/5Ul["R$"UR./5UlUR35R55VV s0sHupX_M sn n=UlUlURRS S5UlURRSS5UlURRSS5UlURRRSS5UlURRRSS5Ul URRRSS5n [BRDRGU 5 [IU 5 URRRSS5n U (a/SQUl%/SQUl&O/SQUl%/SQUl&URRRSS5Ul'[PURRRVs0sHnX_M snUl*[WURS5(afURRXRTbEURRXRTR55HupXRTU 'M [ZR][^R`Vs0sHnUS_M sn5Ul1SUl2URRRSS5(Ga/[fRh"X RRRj5n[mUS S!S"9n[nRp"U5nSSS5 WUl9URrR55V V s0sHupX_M sn n Ul:[wUR.5Ul2[yUR{55H_unnUR}S#U-SSS$9nURdRUR&/UR5-UR./-5 Ma SUlASUlBgs snfs snfs sn nfs snfs snf!,(df  GN =fs sn n f)%zpreprocess the data via the vocab.txt from the `model_dir` path Args: cfg(modelscope.utils.config.ConfigDict) : model config model_dir (str): model path languageen)zhcni z izmultimodal_typedefaulttext2sqlz>=z<=max_src_lengthmax_tgt_lengthmax_image_sizei prompt_typenoneseedimagenet_default_mean_and_stdF)g ףp= ?gv/?gCl?)gZd;O?gy&1?g?)?r$r$patch_image_sizeidatasetN answer2labelrzutf-8)encoding )add_bosadd_eos)Ccfgmodemodelgetrosrexistsabspathrfrom_pretrainedrNotImplementedError add_tokensrangeformat tokenizertorch LongTensor bos_token_idbos_item pad_token_idpad_item eos_token_ideos_item get_vocabitemstgt_dictsrc_dictrrrrnprandomr!rmeanstdr%rtask column_maphasattrr&str maketransstring punctuationtranstabconstraint_trieospjoinr'openjsonload ans2label label2ansr enumeratekeys tokenize_textinserttolisttrain_audio_feature_transformstest_audio_feature_transforms)selfr- model_dirr.argskwargsr9ikeyvaluer!r#kvans2label_filereaderans2label_dictanswer answer_items a/var/www/html/land-doc-ocr/venv/lib/python3.13/site-packages/modelscope/preprocessors/ofa/base.py__init__OfaBasePreprocessor.__init__s **:t< 77>>) $ $ 2I ==D $44Y?I ]]l *&66yAI% % U4[I[k003[IJE$KHKqj//2KHI 88>>  / ;z I  $ ."(()*@*@)AB (()*@*@)AB (()*@*@)AB (11399;) ;  J;)    "iimm,>--mVDxx~~!!&!, tt(,(:(: +U)4% (-DI,DH'DI&DH $ 2 23Es K,DHHMM: : H:  488  "&(("2"2"="="I((3399;%&"< ** ,*$Y* ,-  $ 88>>  nd 3 3 XXi1L1LMNncG<!%6!2=+DN/3~~/C/C/EF/Etqad/EFDN#' (>(>#?D &~':':'<= 6"00&L%1@ $$++Y-C-C,D.9.@.@.B-C/8/E/E.F-GH>/3+-1*qJH) *  , =<Gs0'X!X!X& X, X1X6 Y6 YcUcgURUURSSSS9SRS5nU(a"[R"UR U/5nU(a![R"X@R /5nU$)a Using `OFATokenizer` to tokenize text input. Args: text (`str`): Input text. add_bos ('bool', **optional**, default to `True`) Whether or not to add beginning of sentence token in the front of sentence. add_eos ('bool', **optional**, default to `True`) Whether or not to add ending of sentence token in the end of sentence. Returns: A list of tokens with the max length of `max_src_length + 2` NFTpt) max_lengthadd_special_tokens truncationreturn_tensors input_idsr)r9rsqueezer:catr=rA)ratextr+r,inputss ror\!OfaBasePreprocessor.tokenize_textks < **$  ! "- . /6gaj  YY v67F YY 67F NcUR5RS5RSS5RSS5RSS5n[R"SSU5nUR S5nUR S5nURS5nUb#[U5U:aSRUS U5nU$) a Preprocessing for text sentence. step 1. Get the lower case of input text. step 2. Remove the words within `,.!?*#:;~ ` in the beginning of the sentence. step 3. Replace the words within `-/` or pattern `\s{2,}` with word ` ` and replace tag `` with `person`. step 4. Remove the `\n` in the end of the sentence. step 5. Split the sentence with token ` `, If `max_words` is not None, make a length truncation. Args: caption (`str`): Input text. max_words (`int`, **optional**, default `None`): The max length of input text. If None, do nothing, else make a truncation. Returns: A sequence of `str`. ,.!?*#:;~-r*/zperson\s{2,} N lowerlstripreplaceresubrstripstripsplitlenrT)caption max_words caption_wordss ro pre_captionOfaBasePreprocessor.pre_captions.--/((5==c3G WS# wwz8< &&     ..&--$ c*  S%7)%Chh}Zi89Gr~c`UR5RS5RSS5RSS5n[R"SSU5nUR S5nUR S5nURS5n[U5U:aSRUSU5nU$)a Preprocessing for text sentence. Note that this function is very similar to `pre_caption`, should be merged in the future version. step 1. Get the lower case of input text. step 2. Remove the words within `,.!?*#:;~ ` in the beginning of the sentence. step 3. Replace the words within `-/` or pattern `\s{2,}` with word ` `. step 4. Remove the `\n` in the end of the sentence. step 5. Split the sentence with token ` `, If `max_words` is not None, make a length truncation. Args: question (`str`): Input text. max_ques_words (`int`, **optional**, default `None`): The max length of input text. If None, do nothing, else make a truncation. Returns: A sequence of `str`. rrr*rrrNr)questionmax_ques_wordsquestion_wordss ro pre_question OfaBasePreprocessor.pre_questions.>>#**;7??@CEELWDGFN 66     ??4(>>#&", ~  /xx ?@Hr~cUSnURUR5RSS9R5nUR(a[ R "[U5[UR545R5n[U5U- n[U[U55HRnURR5X%UR5-nURRU5nSXFU'MT XAS'gg)z Add constraint mask. targetr)dimTconstraint_maskN)ner?sumitemrRr:zerosrrDboolr7r=r^get_next_layer) rasample target_itm len_label_itmr start_idxreconstraint_prefix_tokenconstraint_nodess roadd_constraint_mask'OfaBasePreprocessor.add_constraint_masksH% " dmm488Q8?DDF   #kkZ#dmm"4577;tv J-7I9c*o6*.--*>*>++224+5'#'#7#7#F#F+$- 7;"#34 7 )8$ % r~cb[U[R5(aUnU$[U5nU$)a& Get the pillow image. If the input is not a pillow image ,it will load image from a local path or an external url. Args: path_or_url_or_pil (`Union[str, Image]`): Can be: - A path or url reference to an image - A pillow image. Returns: A pillow image. ) isinstancerr)rapath_or_url_or_pilimages ro get_img_pilOfaBasePreprocessor.get_img_pils5'11CU[[&Q&Q" ./  r~c[U[5(a[R"U5nU$[U[5(a.[ R "U5n[R"U5nU$[S[U5S35e)NzUnsupported input type: .) rbytesioBytesIOrMrread TypeErrortype)ra path_or_url audio_bytes file_bytess roget_audio_bytes#OfaBasePreprocessor.get_audio_bytessu k5 ) )**[1K   S ) );/J**Z0K6tK7H6IKL Lr~c [RRXS[U5/S[U5//5up[ XSSS9upgUS-nUR 5n[ XbS5nUc [XbS5nUc [S5eU(aURbURU5nO,U)(a$URU5bURU5n[R"U5R5nURU5nU$)NspeedrateT)to_mononormalize_volumeiPzGPlease install pyKaldi or torchaudio to enable fbank feature extraction) torchaudio sox_effectsapply_effects_tensorrMr numpyr r ImportErrorr_r`r: from_numpyfloat pack_frames) rawaveform sample_ratertarget_sample_rateis_train _waveform_fbanks ro prepare_fbank!OfaBasePreprocessor.prepare_fbanks !+ 6 6 K K s5z "VS1C-D$E F!H( 4$H ' OO%  < =))"EE =Y  ;;G77>EY4=="#66u=E  '--/  ' r~featurecURRS:XaU$URSURR-nUSURRU-nURUS5$)Nr r)r-n_frames_per_stepshapereshape)rarn_packed_framess rorOfaBasePreprocessor.pack_frames'sc 88 % % *N!--*dhh.H.HHG48855GH33r~)rXr=r-rKrRrArYrrrrrHr.r?r%rrErIr`rDr9r_rQ)TT)N)i>F)__name__ __module__ __qualname____firstlineno____doc__rpr\ staticmethodrrrrrrr:Tensorr__static_attributes__r~rorrsiL2\:&&P''R8$"*/$ >45<<4r~r)rr1rrOrrSrVrrFr:rPILrmodelscope.fileior!modelscope.models.multi_modal.ofarrmodelscope.preprocessors.imagermodelscope.utils.trier utils.audio_helperr r r utils.constantrutils.random_helprrrr~rorsK   "J5&330-T4T4r~