^j}UddlmZddlZddlZddlZddlZddlZddlmZddl m Z m Z ddl Z ddlm ZddlmZmZddlmZddlmZddlmZmZdd lmZdd lmZGd d eZGd deZej@e!ejDe#ejHe!ejJe!ejLe!ejNe!ejPe!ejRe#iZ*de+d<ej@dgejDdgejHddgejJdgejLdgejNdgejPdgejRdgiZ,de+d<ej@ddgejDdgejHgdejJddgejLddgejNddgejPddgejRdgiZ-de+d<gdZ. d'dZ/ee/ddd'dZ0 d( d)dZ1d*dZ2 d( d+d Z3 d( d)d!Z4 d( d)d"Z5 d,d#Z6 d( d)d$Z7 d( d-d%Z8 d.d&Z9y)/) annotationsN)Enum)Anycast) deprecatedvoid)Image)denormalize_boxes)polygon_to_maskpolygon_to_xyxy)warn_deprecated)_validate_resolutioncTeZdZdZdZdZdZdZdZdZ dZ d Z e d d Z e dd Zy )LMMa Enum specifying supported Large Multimodal Models (LMMs). .. deprecated:: 0.27.0 `LMM` is deprecated and will be removed in `supervision-0.31.0`. Use `VLM` instead. Attributes: PALIGEMMA: Google's PaliGemma vision-language model. FLORENCE_2: Microsoft's Florence-2 vision-language model. QWEN_2_5_VL: Qwen2.5-VL open vision-language model from Alibaba. QWEN_3_VL: Qwen3-VL open vision-language model from Alibaba. GOOGLE_GEMINI_2_0: Google Gemini 2.0 vision-language model. GOOGLE_GEMINI_2_5: Google Gemini 2.5 vision-language model. MOONDREAM: The Moondream vision-language model. paligemma florence_2 qwen_2_5_vl qwen_3_vl deepseek_vl_2 gemini_2_0 gemini_2_5 moondreamc@|Dcgc]}|jc}Scc}wNvalueclscs d/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/supervision/detection/vlm.pylistzLMM.list1!$%A%%%c(tdt||r|St|tr|j} ||St dt |d|jd#t$r t d|d|j wxYw)Nzn`LMM` is deprecated since `supervision-0.27.0` and will be removed in `supervision-0.31.0`. Use `VLM` instead.Invalid value: . Must be one of Invalid value type: . Must be an instance of  or str.)r isinstancestrlower ValueErrorr!type__name__rrs r from_valuezLMM.from_value5s 7  eS !L eS !KKME Y5z!"4;-/H||nH &   Y ?5'9J388:,!WXX Ys A(()BNreturnz list[str])rz LMM | strr3rr/ __module__ __qualname____doc__ PALIGEMMA FLORENCE_2 QWEN_2_5_VL QWEN_3_VL DEEPSEEK_VL_2GOOGLE_GEMINI_2_0GOOGLE_GEMINI_2_5 MOONDREAM classmethodr!r1r rrsY"IJKI#M$$I&&  rBrcTeZdZdZdZdZdZdZdZdZ dZ d Z e d d Z e dd Zy )VLMa  Enum specifying supported Vision-Language Models (VLMs). Attributes: PALIGEMMA: Google's PaliGemma vision-language model. FLORENCE_2: Microsoft's Florence-2 vision-language model. QWEN_2_5_VL: Qwen2.5-VL open vision-language model from Alibaba. QWEN_3_VL: Qwen3-VL open vision-language model from Alibaba. GOOGLE_GEMINI_2_0: Google Gemini 2.0 vision-language model. GOOGLE_GEMINI_2_5: Google Gemini 2.5 vision-language model. MOONDREAM: The Moondream vision-language model. rrrrrrrrc@|Dcgc]}|jc}Scc}wrrrs r r!zVLM.list`r"r#ct||r|St|tr|j} ||Stdt |d|j d#t$r td|d|j wxYw)Nr%r&r'r(r))r*r+r,r-r!r.r/r0s r r1zVLM.from_valueds eS !L eS !KKME Y5z!"4;-/H||nH &   Y ?5'9J388:,!WXX Ys A)BNr2)r VLM | strr3rDr4rArBr rDrDIsY IJKI#M$$I&&    rBrDzdict[VLM, type] RESULT_TYPES resolution_whinput_whzdict[VLM, list[str]]REQUIRED_ARGUMENTSclasses)rJrIrLALLOWED_ARGUMENTS) #c t|tr t|j}t|t |s!t dt|dt |tj|g}|D]}||vst d|tj|g}|D]!}||vst d|d|j|S#t$r4t d|dtDcgc]}|j ncc}wc}wxYw)a Validates the parameters and result type for a given Vision-Language Model (VLM). Args: vlm: The VLM enum or string specifying the model. result: The result object to validate (type depends on VLM). kwargs: Dictionary of arguments to validate against required/allowed lists. Returns: The validated VLM enum value. Raises: ValueError: If the VLM, result type, or arguments are invalid. zInvalid vlm value: r&zInvalid VLM result type: z . Must be zMissing required argument: z Argument z is not allowed for ) r*r+rDr,r-rrHr.rKgetrMname)vlmresultkwargse required_argsarg allowed_argss r _validate_vlm_parametersrbs."#s ciik"C fl3/ 0'V ~Z S@Q?R S  '**33MB f :3%@A AB%((b1LN l "y-A#((LM MN J) %cU*;c) ([\w\s\-]+))rrrrdtypeN)rirk)rrecompilefindallnparrayemptyshapeintr+floatastypecharstripboolindex) r\rIrLwhpatternmatches matches_arrxyxy_arr class_nameclass_idrZmasks r from_paligemmarsm( .DAqjjVGoof%G9@BHHW$5bhhvFVKq xx$c!:BHHQc% 88ZHTW]]40HI Xz )) AIs # F<Gc |jd}|dk(ry||d}|jd}|dk(ry|d|dz}|jd}|dk(ry|d|dz}||dzdj}|jdr|ddj}||zdz}t j |S#t $rYywxYw)a Attempt to recover and parse a truncated or malformed JSON snippet from Qwen-2.5-VL output. This utility extracts a JSON-like portion from a string that may be truncated or malformed, cleans trailing commas, and attempts to parse it into a Python object. Args: text: Raw text containing the JSON snippet possibly truncated or incomplete. Returns: Parsed Python object (usually list) if recovery and parsing succeed; otherwise `None`. [N}ro,])findrfindrstripendswithjsonloads Exception)text first_bracketsnippet last_brace prefix_endprefixbodyrepaireds r &recover_truncated_qwen_2_5_vl_responsers  # B }~&]]3'  *JN+\\#&  ):>*zA~'(//1 == 9##%DD=3&zz(## s"B3B3B3AB33 B?>B?ct|\}}t|\}}|j}tjdd|tjj}tjdd|j}|j d} |j d} | dk7r | dk7r| | kDr|| | dzj} tj|} t)| t*sJt!j"d t!j"d t$ t!j"d t& fSg} g}| D]C}t)|t,rd |vsd|vr| j/|d |j/|dE| sJt!j"d t!j"d t$ t!j"d t& fSt!j0| t2 }t!j0|t& }|||||gz }|||||gz}d }|mt!j0|Dcgc]}||vc}t4 }||}||}t!j0|Dcgc]}|j7|c}t$ }|||fS#tj$rt|} | | } n} tj|} nf#tttf$rOt!j"d t!j"d t$ t!j"d t& fcYcYSwxYwYSwxYwcc}wcc}w)a Parse and rescale bounding boxes and class labels from Qwen-2.5-VL JSON output. The JSON is expected to be enclosed in triple backticks with the format: ```json [ {"bbox_2d": [x1, y1, x2, y2], "label": "some class name"}, ... ] ``` Args: result: String containing Qwen-2.5-VL JSON bounding box and label data. input_wh: Width and height of the coordinate space where boxes are normalized. resolution_wh: Target width and height to scale bounding boxes. classes: Optional list of valid class names to filter results. If provided, only boxes with labels in this list are returned. Returns: A tuple of `(xyxy, class_id, class_name)` where `xyxy` is an array of shape `(N, 4)` in `(x_min, y_min, x_max, y_max)` format, `class_id` is an optional array of shape `(N,)` with class indices, and `class_name` is an array of shape `(N,)` with class names. z ^```(json)?flagsz```$rrrroNrjrlrmbbox_2dlabel)rr}rrsub IGNORECASErrrrJSONDecodeErrorrast literal_evalr- SyntaxError TypeErrorrurwryr+r*r!dictappendrvrzr~r)r\rJrIrLin_win_hout_wout_hrstartenddatar boxes_list labels_listitemxyxyrrrrs r from_qwen_2_5_vlr1s@&h/JD$' 6LE5 <<>D 66."d"-- @ F F HD 66'2t $ * * ,D IIcNE **S/C {sbyS5[EC!G$**,zz$ dD ! "((4s";RXXdRU=VWWJK*$%$)>'QUBU $y/*4=) *  "((4s";RXXdRU=VWW 88Je ,D+S1J 4tT* *D 5%. .DHxxzBe')B$ODz% 88zJeW]]51JRUV : %%W    9$?  D ''- Y7 HHV$HHT-HHT-  LCKsCJ5 L;*M#L88K L8AL1,L80L11L87L8c t|d||S)a[ Parse and scale bounding boxes from Qwen-3-VL style JSON output. Args: result: String containing the Qwen-3-VL JSON output. resolution_wh: Target resolution `(width, height)` to scale bounding boxes. classes: Optional list of valid classes to filter results. Returns: A tuple of `(xyxy, class_id, class_name)` where `xyxy` is an array of shape `(N, 4)` in `(x_min, y_min, x_max, y_max)` format scaled to `resolution_wh`, `class_id` is an optional array of class indices, and `class_name` is an array of class names. )r)r\rJrIrL)r)r\rIrLs r from_qwen_3_vlrs& #  rBc |\}}tjd|tj}tjd|tj}t|t|k7r$t dt|dt|dg}g}t ||D]\} } | j } tjd| D]s} tt| j djd \} }}}|j| d z |z|d z |z|d z |z|d z |zg|j| utj|tj }tj|}|htj|Dcgc]}||vc}t }||}||}tj|Dcgc]}|j|c}}nat!t#t%|}t'|Dcic]\}}|| }}}tj|Dcgc]}|| c}}|||fScc}wcc}wcc}}wcc}w) u Parse bounding boxes from deepseek-vl2-formatted text, scale them to the specified resolution, and optionally filter by classes. The DeepSeek-VL2 output typically contains pairs of <|ref|> ... <|/ref|> labels and <|det|> ... <|/det|> bounding box definitions. Each <|det|> section may contain one or more bounding boxes in the form [[x1, y1, x2, y2], [x1, y1, x2, y2], ...] (scaled to 0..999). For example: ``` <|ref|>The giraffe at the back<|/ref|><|det|>[[580, 270, 999, 904]]<|/det|><|ref|>The giraffe at the front<|/ref|><|det|>[[26, 31, 632, 998]]<|/det|><|end▁of▁sentence|> ``` Args: result: String containing deepseek-vl2-formatted locations and labels. resolution_wh: Tuple (width, height) to which we scale the box coordinates. classes: Optional list of valid class names. If provided, boxes and labels not in this list are filtered out. Returns: A tuple of `(xyxy, class_id, class_name)` where `xyxy` is an array of shape `(n, 4)` in format `[x1, y1, x2, y2]`, `class_id` is an optional array of shape `(n,)` with class indices, and `class_name` is an array of shape `(n,)` with class labels. z<\|ref\|>(.*?)<\|/ref\|>rz<\|det\|>(.*?)<\|/det\|>zNumber of ref tags (z) and det tags (z) in the result must be equal.z \[(.*?)\]z[]rirm)rrrtSlenr-zipr}maprzsplitrrurvfloat32r~rsortedr!set enumerate)r\rIrLwidthheightlabel_segmentsdetection_segments xyxy_listclass_name_listrdetection_blobcurrent_class_nameboxx1y1x2y2rrrZrrunique_classesi class_to_ids r from_deepseek_vl_2rs=:"ME6ZZ ;V244PN$?rttT >c"455"3~#6"78 !3455S U  $&I!#O!$^5G!H 7~"[[]::lN; 7C  $(=(=c(BCNBB   #X%#X&#X%#X&    " "#5 6 7 7 88IRZZ 0D/*JxxZ@T@MDz% 88ZHTW]]40HIS_ 56.7.GH71dtQwH H88:F4[.FG : %%AIIFs I I I3 Ict|dk(s Jd|tt|j}|tvrt |dt||}|dvrFt j|dt j}t j|d}||ddfS|d k(r.t j|dt j}|dddfS|d k(rt j|d t j}|jd d d}t j|Dcgc] }t|c}}t j|d}||d|fS|dvrg}g}|dD]} | D]}t j|djt j}t||jt} |j| t|} |j| t j|t j}t j|} |d| dfS|dk(rFt j|dt j}t j|d}||ddfS|dvr&t!|t"sJdt%||dk(r result, got zNo object detected.rjz,z1Expected string to end in location tags, but got rzUnimplemented task: )rnextiterkeysSUPPORTED_TASKS_FLORENCE_2r-rurvrreshaper r{int32r r~rr*r+r.rwrrrssearchrgroupsr RuntimeError)r\rItaskrrxyxyxyxypolygonr masks_listpolygons_of_same_classrxyxy_boxmasksrmatchrr result_strings r from_florence_2rs0 v;! TKF8TT  V[[]# $D --f9:T9U V  D\F RRxxx( ;&*+VT4'' ""xxx( ;T4%% ""88F<0 C##B1-xxJg1JK&*+VT8++ RR,. -/ &,Z&8 + "1 +**Wg6==bhhG&w >EEdK!!$'*73  *  + +xx 4$T5$&& ,,xxx( ;&12VT4'' BB&#& B4<. Q & * *88F"**5rxx|T4O O**LMv&  ?x H  $M21xx( ; !Q1&--/ =/*VT4'' -dV4 55aKs,N?c^t|\}}|j}t|D]7\}}|dk(s dj||dzd}|j dd}n t j |}t|tsJtjdtjdt tjdt fSg} g} |D]U} t| trd | vsd | vr| j| d | d } | j| d| d| d | d gWt!| dk(rJtjdtjdt tjdt fSt#tj$| tj& ||fd} tj$| } d}|gtj$| Dcgc]}||vc}t( }| |} | |} tj$| Dcgc]}|j+|c}}| || fS#t j$rMtjdtjdt tjdt fcYSwxYwcc}wcc}w)a Parse and scale bounding boxes from Google Gemini style [JSON output](https://ai.google.dev/gemini-api/docs/vision?lang=python). The JSON is expected to be enclosed in triple backticks with the format: ```json [ {"box_2d": [x1, y1, x2, y2], "label": "some class name"}, ... ] ``` For example: ```json [ {"box_2d": [10, 20, 110, 120], "label": "cat"}, {"box_2d": [50, 100, 150, 200], "label": "dog"} ] ``` Args: result: String containing the JSON snippet enclosed by triple backticks. resolution_wh: (output_width, output_height) to which we rescale the boxes. classes: Optional list of valid class names. If provided, returned boxes/labels are filtered to only those classes found here. Returns: A tuple of `(xyxy, class_id, class_name)` where `xyxy` is an array of shape `(n, 4)` in format `[x1, y1, x2, y2]`, `class_id` is an optional array of shape `(n,)` with class indices, and `class_name` is an array of shape `(n,)` with class labels. ```json roN```rrjrlrmbox_2drrprqrrInormalization_factor)r splitlinesrjoinrrrrrurwryr+r*r!rrrr rvfloat64r~r)r\rIrLrrlinesrlinerrrrrrrrZrs r from_google_gemini_2_0rOs]N .DAq    EU#4 9 YYuQUW~.F\\%(+F   Vzz&! dD !xx$c!:BHHTQT%v,!'3/w7I7I,8 %%bhh1vT&BC%223KLGL$*$4$4W$= #(::bjj.C#D#L#LS#Q(+=+;'Su'*=+;'Su&+em %*U] &?zA~'/!+[ 9).)9)9)B)B(7(H>@XX!"Ad>GAC@RUV@VGE%Ku$<=&--g6&--bhh1vT.JKJ 4 *&&tL'9:"Oq8#t  HHV  HHRs # HHRs # HHRu %     88Je ,D+&JxxZ@T@MDz% 88ZHTW]]40HI  !+.z4+@I4D!IJI  &03OT0JSWQdqSOStC O45 09-0HIHAuuaxI I88:F4[.FG3B2M.SW%/$:BHHZ E   e    HHV  HHRs # HHRs # HHRu %     f%H"))"((Aq6*FGHXAIITJFsV%UA W X  X: XX X(X X > X&A8W  W 5XXc~ |\}}|dks|dkrtd|d|dd|vst|dtstjdt Sg}|dD]@ t fdd Ds d } d } d } d }|j||||gBt|dk(r4ttjttjdSttjtttj|jtj ||fS)a  Parse and scale bounding boxes from moondream JSON output. The JSON is expected to have a key "objects" with a list of dictionaries: { "objects": [ {"x_min": 0.1, "y_min": 0.2, "x_max": 0.3, "y_max": 0.4}, ... ] } For Example: { "objects": [ {"x_min": 0.1, "y_min": 0.2, "x_max": 0.3, "y_max": 0.4}, {"x_min": 0.5, "y_min": 0.6, "x_max": 0.7, "y_max": 0.8} ] } Args: result: Dictionary containing the JSON output from the model. resolution_wh: (output_width, output_height) to which we rescale the boxes. Returns: An array of shape `(n, 4)` containing the bounding boxes coordinates in format `[x1, y1, x2, y2]`. rz8Both dimensions in resolution_wh must be positive. Got (z, z).objectsrjrmc3&K|]}|v ywrrA).0krs r z!from_moondream..sK19Ks)rrrrrrrr)rI)r-r*r!rurwrzallrrrnptNDArrayrr rvr{r) r\rIrrrrrrrrs @r from_moondreamr)Ws6@ DAqAvaFqcA3b Q  j 1BD&Ixxe,, Dy! 2K&JKK W W W W  UE5%01 2 4yA~CKK$bhhv&677  C HHTN ! !"** -a&  rB)r[rGr\rr]dict[str, Any]r3rDr)r\r+rItuple[int, int]rLlist[str] | Noner3Btuple[npt.NDArray[Any], npt.NDArray[Any] | None, npt.NDArray[Any]])rr+r3z Any | None) r\r+rJr+rIr+rLr,r3r-)r\r*rIr+r3zbtuple[npt.NDArray[Any], npt.NDArray[Any] | None, npt.NDArray[Any] | None, npt.NDArray[Any] | None])r\r+rIr+rLr,r3zttuple[npt.NDArray[Any], npt.NDArray[Any] | None, npt.NDArray[Any], npt.NDArray[Any] | None, npt.NDArray[Any] | None])r\r*rIr+r3znpt.NDArray[Any]): __future__rrrr rrrenumrtypingrrnumpyru numpy.typingr' deprecaterrPILr !supervision.detection.utils.boxesr &supervision.detection.utils.convertersr r supervision.utils.internalr supervision.validatorsrrrDr8r+r9rr:r;r<r=r>r?rH__annotations__rKrMrrbrgrrrrrrrrr)rArBr r:sw" &?S670 $0 f( $( XMM3NNDOOSMM3s33MM4 ! o MMO$NN_%OOj/2MMO$(O,O,MMO$ ,( MMOY/NN_%OO=MMOY/3OY7OY7MMO$ +'  ( (()7((V # %  % NR** ** /**:J**G**Z*b!% Y& Y&Y&#Y& Y& H Y&~!% "H 8NRD& D& /D&:JD&GD&N]6 ]6+:]6]6F!%T& T&"T&T&H T&t!%n n"nn nb? ?"??rB