9iLVSSKJrJr SSKrSSKJr SSKJr SSKJ r "SS\ 5r g) )AnyDictN) transforms)ModeKeys)OfaBasePreprocessorc^\rSrSrSr\R 4U4SjjrS\\ \ 4S\\ \ 44Sjr S\\ \ 4S\\ \ 44Sjr S\\ \ 4S\\ \ 44Sjr S rU=r$) OfaImageCaptioningPreprocessor z- OFA preprocessor for image captioning task. c p>[[U] "XU/UQ70UD6 [R"S[R "UR UR 4[RRS9[R"5[R"URURS9/5Ul g)zpreprocess the data Args: cfg(modelscope.utils.config.ConfigDict) : model config model_dir (str): model path, mode: preprocessor mode (model mode) c$URS5$)NRGB)convert)images m/var/www/html/land-doc-ocr/venv/lib/python3.13/site-packages/modelscope/preprocessors/ofa/image_captioning.py9OfaImageCaptioningPreprocessor.__init__..!s %--.) interpolation)meanstdN)superr __init__rComposeResizepatch_image_sizeInterpolationModeBICUBICToTensor Normalizerrpatch_resize_transform)selfcfg model_dirmodeargskwargs __class__s rr'OfaImageCaptioningPreprocessor.__init__s ,   T D48 D(::BB D    !  diiTXX > : ' #rdatareturncUR[R:XaURU5$UR U5$)N)r%rTRAIN_build_train_sample_build_infer_sample)r"r*s r__call__'OfaImageCaptioningPreprocessor.__call__)s4 99 &++D1 1++D1 1rcxURU5nUSnURUR5R5nUR5R 5nSR USUR 5nURUSS9US'[R"URUSSS/5US'U$) a Building training samples. step 1. Preprocess the data using the logic of `_build_infer_sample` and make sure the label data in the result. step 2. Preprocess the label data. Contains: - remove tokens within `!"#$%&'()*+,-./:;<=>?@[\]^_`{|}~` and tripe - tokenize the label as `target` value without `bos` token. - add `bos` token and remove `eos` token of `target` as `prev_output_tokens`. Args: data (`Dict[str, Any]`): Input data, should contains the key of `image`, `prompt` and `label`, `image` refers the image input data, `prompt` refers the text input data the `label` is the supervised data for training. Return: A dict object, contains source, image, mask, label, target tokens, and previous output tokens data. label NF)add_bostargetprev_output_tokens) r/ translatetranstabstripsplitjoinmax_tgt_length tokenize_texttorchcatbos_item)r"r*sampler6target_token_lists rr.2OfaImageCaptioningPreprocessor._build_train_sample/s&))$/!!$--0668"LLN002+,@T-@-@AB--fe-Dx',yy ]]F8,Sb1 2(4#$ rczURXRS5nURU5nURRR SS5nUR U5nUU[R"S/5S.nSUR;a'URSU;aXRSUS'U$)a Building inference samples. step 1. Get the pillow image. step 2. Do some transforms for the pillow image as the image input, such as resize, normalize, to tensor etc. step 3. Tokenize the prompt as text input. step 4. Determine Whether or not to add labels to the sample. Args: data (`Dict[str, Any]`): Input data, should contains the key of `image` and `prompt`, the former refers the image input data, and the later refers the text input data. Return: A dict object, contains source, image, mask and label data. rpromptz what does the image describe?T)source patch_image patch_masktextr3) get_img_pil column_mapr!r#modelgetr?r@tensor)r"r*rrIrGinputsrCs rr/2OfaImageCaptioningPreprocessor._build_infer_sampleLs   oog&>!?@11%8 ##H.NO##F+&,,v.  T__ $)@D)H"??6#:;F7O r)r!)__name__ __module__ __qualname____firstlineno____doc__r INFERENCErrstrrr0r.r/__static_attributes__ __classcell__)r(s@rr r s(( 22T#s(^2S#X2 S#X4S>:S#X4S>rr ) typingrrr@ torchvisionrmodelscope.utils.constantrbaserr rrras# ".%\%8\r