^jdZddlmZddlmZddlmZddlZddlZddl Z ddl m Z ddlm Z ddl mZmZmZdd lmZdd lmZmZdd lmZeZdd Zdd ZddZ d d!dZd"dZd#dZGdde j>j@Z Gdde!Z"dd d$dZ# d% d&dZ$ d% d&dZ%d'dZ&d(dZ'd'dZ(y))zCOCO dataset which returns image_id for evaluation. Mostly copy-paste from https://github.com/pytorch/vision/blob/13b35ff/references/detection/coco_utils.py ) annotations)Path)AnyN)Image)Tensor)ComposeToDtypeToImage) AUG_CONFIG)AlbumentationsWrapper Normalize) get_loggerc@t|dz dz jS)Ntrain_annotations.coco.json)rexists) dataset_dirs _/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/datasets/coco.pyis_valid_coco_datasetr's!   '*B B J J LLc|jjDchc]3\}}|jds|jds)t|5}}}|r t |S|j j D]O}|jdst|jdddkDs3|jt|dQt |Scc}}w)zLReturn sorted COCO category ids that carry keypoint metadata or annotations. keypointsskeleton num_keypointsr category_id)catsitemsgetintsortedannsvaluesadd)cococat_idcategory category_ids annotations r_category_ids_with_keypointsr)+s-1IIOO,=(kAZ^f^j^jku^vF Ll##ii&&(= >>+ &#jnn_a.P*QTU*U   SM!:; <= , s *C Cc $t|xsg}t|Dcgc] \}}|dkDs |}}}|s tdt|}|s tdt |t |kDr$tdt |dt |dt d|j jD}tt |t|dz}t} i} t||D]>\} } | |k\rtd | d | d t |d | | | <| j| @t|D cgc] } | | vs|  } } |D],} | | vr| std | d| jd| | <.| Scc}}wcc} w)aMap COCO category ids onto model label slots that have keypoint capacity. RF-DETR keypoint schemas are indexed by model label. The preview person-keypoint schema is ``[17]``: label slot ``0`` owns the 17 COCO person keypoints. Legacy checkpoints may still use a background-first ``[0, 17]`` schema where slot ``0`` is reserved (0 keypoints) and slot ``1`` is person. This helper maps keypoint-bearing categories onto slots with a non-zero keypoint count (``count > 0``), so both layouts keep supervision aligned. For multi-class keypoint training supply e.g. ``[17, 4]`` where each non-zero entry corresponds to a keypoint-bearing category in ascending COCO category ID order. rzKeypoint COCO dataset requested, but num_keypoints_per_class has no active keypoint slots. Provide a schema such as [17] for the keypoint preview model.zKeypoint COCO dataset has no keypoint category metadata and no keypoint annotations. Expected COCO categories with a 'keypoints' field or annotations with 'keypoints'/'num_keypoints'.zz,_build_keypoint_cat2label..XsGFCKGszKeypoint schema slot z for category_id z# exceeds the detected class count (zE). Pass num_classes large enough to include this keypoint label slot.z>No free model label slots remain for non-keypoint category_id .)list enumerate ValueErrorr)lenr rkeysmaxsetzipr#rangepop)r$num_keypoints_per_classschemaidxcount active_slotskeypoint_cat_idssorted_cat_idsrequired_slotsassigned_slots cat2labelr%slot free_slotss r_build_keypoint_cat2labelrF9s)/R 0F*3F*;IJCuqyCILI  L  4D9  q   s<00 $%&&B3|CTBUV^ ^  Gdiinn6FGGN^,c,.?!.CDN"uN "I,l;!  > !'v->vhG'((mo ! &4 !$)#8W4Dll3/ V    '**37KCF{r)FFNF)r str | Pathrrrz Any | Nonerboolrrr:list[int] | NonerrreturnNone)r<rrztuple[Any, Any])rr __module__ __qualname____doc__rr __classcell__)rs@rrrsu"R$"'48#(            "2  !     D  rrc<eZdZdZ d ddZddZy)ruConvert a raw COCO annotation dict into model-ready tensors. Accepts the ``(image, target)`` pair produced by ``torchvision.datasets.CocoDetection`` and returns the same image alongside a target dict containing: - ``"boxes"`` – ``(N, 4)`` float32 tensor in absolute ``[x_min, y_min, x_max, y_max]`` format. - ``"labels"`` – ``(N,)`` int64 tensor of class indices. - ``"image_id"`` – scalar int64 tensor. - ``"area"`` – ``(N,)`` float32 tensor of annotation areas (used by COCO eval). - ``"iscrowd"`` – ``(N,)`` int64 tensor (0 = instance, 1 = crowd). - ``"masks"`` – ``(N, H, W)`` bool tensor of binary segmentation masks, only present when ``include_masks=True``. - ``"keypoints"`` – ``(N, K, 3)`` float32 tensor in COCO keypoint format, only present when ``include_keypoints=True``. Crowd annotations (``iscrowd=1``) and degenerate boxes (zero width or height after clamping to image boundaries) are filtered out. Args: include_masks: If ``True``, decode segmentation annotations (polygon or RLE format) into binary masks and include them in the returned target dict. cat2label: Optional mapping from COCO ``category_id`` values to contiguous 0-based label indices. When ``None`` (default) the raw ``category_id`` values are used as labels directly, which is correct for datasets whose IDs are already 0-indexed. Pass a non-``None`` mapping for sparse COCO-style datasets (e.g. IDs 1–90 with gaps) so that labels stay within the model's output range. num_keypoints_per_class: Optional keypoint schema. When provided, keypoints are padded/truncated to ``max(num_keypoints_per_class)`` in each annotation. Ncf||_||_||_|t|d|_yd|_y)Nr)default)rrrCr5r)rrrrCr:s rrzConvertCoco.__init__:s8+!2"H_HkS!8!Dqrrc : |j\}}|d}tj|g}|d}|Dcgc]}d|vs |ddk(s|}}|Dcgc]}|d }}tj|tjj dd}|ddd dfxx|dddd fz cc<|ddddd fj d| |ddd dd fj d| g} |D]s}|d } t |d dN| |jvr!td| d|jdd| j|j| c| j| utj| tj} |dddf|ddd fkD|ddd f|dddfkDz} || }| | } i}||d<| |d<||d<tj|Dcgc]}|d c}} tj|Dcgc] }d|vr|dndc}} | | |d<| | |d<d}|jr|j}|dk(r*|D]%}|jd}|t|dz}ng}|D]}|jd}|7|jtj|dftjMtj|tjj dd}|j d|krKtj|dftj}||d|j d|j||j|d|t|dkDrtj"|d}n(tjd|dftj}|| |d<|j$rt|dkDr{d|dvrt|Dcgc]}|jdg}}t'|||}|j)dkDr || |d<nWtjd||ftj*|d<n+tjd||ftj*|d<|dj-|d<| |d||d<tjt/|t/|g|d<tjt/|t/|g|d<||fScc}wcc}wcc}wcc}wcc}w)NrriscrowdrbboxrerJrMrK)minr5r.rrCzUnknown category_id z for image_id zR encountered in annotations. Check that your category mapping matches the dataset.rLboxeslabelsarearrgrbr~ orig_sizer_)r_rlrvfloat32reshapeclamp_getattrrCKeyErrorrrkint64rrr3rmshaperyrrnumelrnrr)rimagerwhrannoobjrclassesrkeeprr keypoint_keeprrkeypoint_tensors raw_keypointskeypoint_tensorpadded keypoints_outrzr~s r__call__zConvertCoco.__call__Fszz1*%??H:.m$#Sy';s9~QR?RSS(,-V--U]];CCBJ ae a!e $  aAg!+ aAg!+ ,Cm,Kt[$/;dnn4".{m>&**U_J`Iabllt~~k:;{+ ,//'=ad eAqDk)eAqDkE!Q$K.GHd $-w"x%zt<F <=//X\"]QTY#5E3y>1#L"]^dv#DMy'+  ! ! ..M!C # 4I ,(+I!(;   .0  I # 4  ($++EKK8JRWR_R_,`a"'//-u}}"U"]"]^`bc"d"((+m;"[[-);5==QF9HF5_22156$++F3 ''(GH I#$q( % ,?{#a&#a&)9:vf}yT-> ="]T!Ns#TT T  TT*T)FFNN) rrrrrCzdict[int, int] | Noner:rrr)rz Image.Imagerzdict[str, Any]rz"tuple[Image.Image, dict[str, Any]])rrrrrrrrQrrrrsV>$"'+/48 s s  s) s "2 s  sdrr)max_sizec F|rBdd|Dcgc] }d||di c}ii}ddddgdiidd|Dcgc] }d d d g||d ic}iigii}nK|xsd }t|dk(r|dn|}dddd|iidd|iigii}ddddgdiid d d gd d d idd|iidd|iigii}dd||giigScc}wcc}w)uBuild the training resize pipeline as an Albumentations config list. Expresses the ``RandomSelect(resize_a, Compose([resize_b1, crop, resize_b2]))`` pattern as a config-driven ``OneOf``/``Sequential`` for use with :meth:`AlbumentationsWrapper.from_config`. Two branches are selected with equal probability: - **Option A** – direct resize to the target scale(s). - **Option B** – resize to an intermediate scale (400/500/600 px), crop, then resize to the target scale. Divisibility padding (rounding ``H``/``W`` up to a multiple of ``patch_size * num_windows``) is handled by the batch collator via :func:`~rfdetr.utilities.tensors.make_collate_fn`, not here. Args: scales: Target resize scales in pixels. square: If ``True``, produce square output using ``A.Resize`` (one random scale from *scales*). If ``False``, preserve aspect ratio using ``A.SmallestMaxSize`` with an optional long-side cap. max_size: Maximum long-side size for non-square resizes. Defaults to ``1333`` when *square* is ``False``. Returns: A single-element list containing a ``OneOf`` config entry. OneOfrResizer{r| SequentialSmallestMaxSizer)iiXRandomSizedCropir)min_max_heightr{r|5r.rLongestMaxSize)r3)rYsquarersoption_aoption_bcap size_params r_build_train_resize_configrs\> fUQ*CDU$  &_(EF()/+$%"3Sz]^ij4k l+"  $  $'*6{a'7&)V &Z(@A% C'89  &_(EF&C:QT_b(cd&Z(@A% C'89   |h%9: ; <vhGHG,6,Bj /;; &ved K &_%0<<#9LL , 'HXx((  #Hx  O#/;;"Z$<=!J#56   HH(HHHiHIIK/;;hS]hrHs=tr*s3 # ??2G. M  5t"    " b 4%nPK((66PfN&Nj K= K= K= K=  K=b!$37!,0e-e-e-e- e-  e-  e-e-1e-e-*e- e-V!$37!,0M-M-M-M- M-  M-  M-M-1M-M-*M- M-`Rj 1Lr