|hoddlZddlmZddlmZddlmZddlmZddl m Z m Z m Z m Z mZddlZddlZddlZddlmZddlmZdd lmZmZmZmZmZdd lmZdd lm Z m!Z!dd l"m#Z#d dl$m%Z%m&Z&m'Z'm(Z(m)Z)m*Z*m+Z+d dl,m-Z-d dl.m/Z/d dl0m1Z1m2Z2m3Z3m4Z4m5Z5m6Z6m7Z7m8Z8dZ9Gdde-Z:Gdde:Z;Gdde:Z<GddeZ=Gdde-Z>GddZ?y)N) defaultdict)repeat) ThreadPool)Path)AnyDictListOptionalTuple)Image) ConcatDataset) LOCAL_RANKLOGGER NUM_THREADSTQDMcolorstr) Instances)resample_segmentssegments2boxes)TORCHVISION_0_18)ComposeFormat LetterBoxRandomLoadTextclassify_augmentationsclassify_transforms v8_transforms) BaseDataset)merge_multi_segment)HELP_URLcheck_file_speedsget_hashimg2label_pathsload_dataset_cache_filesave_dataset_cache_file verify_imageverify_image_labelz1.0.3ceZdZdZddddeedeffdZedfd ed efd Z d e efd Z dd eed e fdZ d ed dfdZded efdZede ed efdZxZS) YOLODataseta Dataset class for loading object detection and/or segmentation labels in YOLO format. This class supports loading data for object detection, segmentation, pose estimation, and oriented bounding box (OBB) tasks using the YOLO format. Attributes: use_segments (bool): Indicates if segmentation masks should be used. use_keypoints (bool): Indicates if keypoints should be used for pose estimation. use_obb (bool): Indicates if oriented bounding boxes should be used. data (dict): Dataset configuration dictionary. Methods: cache_labels: Cache dataset labels, check images and read shapes. get_labels: Return dictionary of labels for YOLO training. build_transforms: Build and append transforms to the list. close_mosaic: Set mosaic, copy_paste and mixup options to 0.0 and build transformations. update_labels_info: Update label format for different tasks. collate_fn: Collate data samples into batches. Examples: >>> dataset = YOLODataset(img_path="path/to/images", data={"names": {0: "person"}}, task="detect") >>> dataset.get_labels() Ndetectdatataskr-r.c|dk(|_|dk(|_|dk(|_||_|jr|jrJdt ||d|jdi|y)ap Initialize the YOLODataset. Args: data (dict, optional): Dataset configuration dictionary. task (str): Task type, one of 'detect', 'segment', 'pose', or 'obb'. *args (Any): Additional positional arguments for the parent class. **kwargs (Any): Additional keyword arguments for the parent class. segmentposeobbz(Can not use both segments and keypoints.channelsN) use_segments use_keypointsuse_obbr-super__init__selfr-r.argskwargs __class__s W/var/www/html/test/engine/venv/lib/python3.12/site-packages/ultralytics/data/dataset.pyr8zYOLODataset.__init__Ism!I-!V^u}  %%$*<*<i?ii= $I:)>I&I./labels.cachepathreturncbdgi}ddddgf\}}}}}|jd|j|jz d}t|j} |j j dd\} } |jr| dks| dvr tdtt5} | jtt|j|jt|jt|jtt|j d t| t| t|j  } t#| || }|D]{\ }}}}}}}}}}||z }||z }||z }||z }|r/|dj%|||d d dd f|d d d d f||ddd|r|j%||d|d||zd|d|_}|j)d d d |r$t+j,dj/||dk(r+t+j0|jd|dt2t5|j|jz|d<||||t|jf|d<||d<t7|j||t8|S#1swYxYw)z Cache dataset labels, check images and read shapes. Args: path (Path): Path where to save the cache file. Returns: (dict): Dictionary containing cached labels and related information. labelsr Scanning ... kpt_shape)rr>z'kpt_shape' in data.yaml missing or incorrect. Should be a list with [number of keypoints, number of dims (2 for x,y or 3 for x,y,visible)], i.e. 'kpt_shape: [17, 3]'namesfunciterabledesctotalNrTxywh)im_fileshapeclsbboxessegments keypoints normalized bbox_format images,  backgrounds,  corrupt zNo labels found in z. hashresultsmsgs)prefixparentstemlenim_filesr-getr5 ValueErrorrrimapr(zip label_filesr single_clsrappendrOcloserinfojoinwarningr!r#r&DATASET_CACHE_VERSION)r:rAxnmnfnencrarOrPnkptndimpoolr`pbarrRlbrSrVkeypointnm_fnf_fne_fnc_fmsgs r> cache_labelszYOLODataset.cache_labelsZsrN !Q2~BB++i dii(?'@DDMM"YY]];7 d   419F0Bi  $$ ii'MM$$4;;'4--.3tyy1234L4L4??+  Gd%8DW[ WSUHhdD$PSd d d d hK&&'.%*#%a1f:&(ABi(0)1*.+1  KK$#fAbT27)>"XV ) W* JJLI$ L  KK $ ( 7 NNdkk]*=dV2hZP QT-- =>& 2r2s4=='99) &  T16KL]$ $ s EJ%%J.c t|j|_t|jdjj d} t |d}}|dtk(sJ|dt|j|jzk(sJ |jd\}}}}}|rbtdvrZd |d |d ||zd |d } td|j | z|||dr't#j$dj'|ddD cgc]} |j| c} |d} | st)d|dt*| D cgc]} | d c} |_d| D} dt-| D\}}}|r-||k7r(t#j.d|d|d| D]} g| d< |dk(rt#j.d|dt*| S#tttf$r|j|d}}YiwxYwcc} wcc} w)a8 Return dictionary of labels for YOLO training. This method loads labels from disk or cache, verifies their integrity, and prepares them for training. Returns: (List[dict]): List of label dictionaries, each containing information about an image and its annotations. r.cacheTversionr_Fr`rrEz... r[r\r]NrOrPinitialrar^)r_rrarDzNo valid images found in z8. Images with incorrectly formatted labels are ignored. rRc3nK|]-}t|dt|dt|df/yw)rTrUrVNre).0r|s r> z)YOLODataset.get_labels..s2]PRC5 NC8 $5s2j>7JK]s35c32K|]}t|ywN)sum)rrss r>rz)YOLODataset.get_labels..s+JqCF+Jz@Box and segment counts should be equal, but got len(segments) = z, len(boxes) = z. To resolve this only boxes will be used and all segments will be removed. To avoid this please supply either a detect or segment dataset, not a detect-segment mixed dataset.rVzLabels are missing or empty in z#, training may not work correctly. )r$rfrkrrc with_suffixr%rrr#FileNotFoundErrorAssertionErrorAttributeErrorrpoprrrbrrorp RuntimeErrorr!rjrq)r: cache_pathcacheexistsrurtrvrwndkrDr|lengthslen_cls len_boxes len_segmentss r> get_labelszYOLODataset.get_labelss:+4==9$**1-.55AA(K  A3J?6E#'<< <<=HT-=-= -M$NN NN "IIi0BB jG+J>B A --j956E A < 9s AGH H'G=<G=hypc |jr|jr|js |jnd|_|jr|js |jnd|_|jr|js |jnd|_t ||j |}n-tt|j |j fdg}|jtdd|j|j|jd|j|j|jr |j nd |S)z Build and append transforms to the list. Args: hyp (dict, optional): Hyperparameters for transforms. Returns: (Compose): Composed transforms. F) new_shapescaleuprQT) rY normalize return_maskreturn_keypoint return_obb batch_idx mask_ratio mask_overlapbgr)augmentrectmosaicmixupcutmixrimgszrrrmrr4r5r6r overlap_maskr)r:r transformss r>build_transformszYOLODataset.build_transformss <<'+||DII3CJ%)\\$)) CI'+||DII3CJ&tTZZ=J )tzz4::6NX]"^!_`J " -- $ 2 2<<>> --#||CGG  r?chd|_d|_d|_d|_|j ||_y)z Disable mosaic, copy_paste, mixup and cutmix augmentations by setting their probabilities to 0.0. Args: hyp (dict): Hyperparameters for transforms. rN)r copy_pasterrrr)r:rs r> close_mosaiczYOLODataset.close_mosaics2   //4r?labelc|jd}|jdg}|jdd}|jd}|jd}|jrdnd}t|d kDrAtd |D}||kr|d zn|}t j t || d }n(t jd |dftj}t||||||d<|S)a Update label format for different tasks. Args: label (dict): Label dictionary containing bboxes, segments, keypoints, etc. Returns: (dict): Updated label dictionary with instances. Note: cls is not with bboxes now, classification and semantic segmentation need an independent cls label Can also support classification and semantic segmentation by adding or removing dict keys there. rUrVrWNrYrXdirc32K|]}t|ywrr)rss r>rz1YOLODataset.update_labels_info..s3Q#a&3rr)raxisrHdtype)rYrX instances) rr6remaxnpstackrzerosfloat32r) r:rrUrVrWrYrXsegment_resamplesmax_lens r>update_labels_infozYOLODataset.update_labels_infos8$99Z,IIk40 ii . YY|, $(<O PWXYHxx$5q 9LH&vxP[hrsk r?batchc i}|Dcgc]$}tt|j&}}|dj}t t |Dcgc]}t |j c}}t|D]~\}}||}|dvrtj|d}n:|dk(r5tjjjj|d}|dvrtj|d}|||<t |d|d<tt!|dD]}|d|xx|z cc<tj|dd|d<|Scc}wcc}w) Collate data samples into batches. Args: batch (List[dict]): List of dictionaries containing sample data. Returns: (dict): Collated batch with stacked tensors. r>img text_featsvisualsT) batch_first>rTr2masksrUrVrWr)dictsorteditemskeyslistrjvalues enumeratetorchrnnutilsrnn pad_sequencecatrangere)r new_batchbrrirvalues r> collate_fnzYOLODataset.collate_fnsV 278QfQWWY'(88Qx}}ce<D,<=>dO !DAq1IE)) E1-i**7747PNN %+ IaL !"&i &r*r*/s26:xJHTNJJ")--=(>BB4BH0DJ0dHTNg@ 5 5 5B$t*r?r*c eZdZdZddddeedeffdZded effd Zdd eed e ffd Z e d Z e dZ eddeded eefdZxZS)YOLOMultiModalDataseta Dataset class for loading object detection and/or segmentation labels in YOLO format with multi-modal support. This class extends YOLODataset to add text information for multi-modal model training, enabling models to process both image and text data. Methods: update_labels_info: Add text information for multi-modal model training. build_transforms: Enhance data transformations with text augmentation. Examples: >>> dataset = YOLOMultiModalDataset(img_path="path/to/images", data={"names": {0: "person"}}, task="detect") >>> batch = next(iter(dataset)) >>> print(batch.keys()) # Should include 'texts' Nr+r,r-r.c*t||||d|y)ax Initialize a YOLOMultiModalDataset. Args: data (dict, optional): Dataset configuration dictionary. task (str): Task type, one of 'detect', 'segment', 'pose', or 'obb'. *args (Any): Additional positional arguments for the parent class. **kwargs (Any): Additional keyword arguments for the parent class. r,Nr7r8r9s r>r8zYOLOMultiModalDataset.__init__Ns $T??r?rrBct||}|jdjDcgc]\}}|j dc}}|d<|Scc}}w)a Add text information for multi-modal model training. Args: label (dict): Label dictionary containing bboxes, segments, keypoints, etc. Returns: (dict): Updated label dictionary with instances and texts. rJ/texts)r7rr-rsplit)r:rrD_vr=s r>rz(YOLOMultiModalDataset.update_labels_infoZsS+E259IIg4F4L4L4NODAq1773<Ow PsArct||}|jrPtt |j ddd|j |j}|jd||S)a! Enhance data transformations with optional text augmentation for multi-modal training. Args: hyp (dict, optional): Hyperparameters for transforms. Returns: (Compose): Composed transforms including text augmentation if applicable. rwPT max_samplespadding padding_valuer) r7rrrminr-_get_neg_texts category_freqinsertr:rr transformr=s r>rz&YOLOMultiModalDataset.build_transformskshW-c2 << ' $4"11$2D2DEI   b) ,r?c|jdj}|Dchc](}|jdD]}|j*c}}Scc}}w)zw Return category names for the dataset. Returns: (Set[str]): List of class names. rJr)r-rrstrip)r:rJnamers r>category_namesz$YOLOMultiModalDataset.category_namessI '"))+&+EdTZZ_E E EEEs-AcX|jdjDcgc]}|jd}}tt}|j D]M}|dj dD]4}|t |}|D]}|j}||xxdz cc<!6O|Scc}w)1Return frequency of each category in the dataset.rJrrTrr)r-rrrintrDsqueezer )r:rrrrctextts r>rz#YOLOMultiModalDataset.category_freqs(,yy'9'@'@'BC!CC#C( [[ *E5\))"- *SV}*A A!!$)$* * * DsB'r thresholdc`|jDcgc] \}}||k\s |c}}Scc}}wz7Get negative text samples based on frequency threshold.rrrrrs r>rz$YOLOMultiModalDataset._get_neg_texts+,113FdaqI~FFF **rr)rrrrr rrr8rrrpropertyr rrrr rrrs@r>rr=s 6:x @HTN @ @"HTNg0FF  GdGsGT#YGGr?rc eZdZdZddddedeffdZded efd Zd eeee fd d fd Z e dfde d eee ffdZ d eefdZ ddeed effd ZedZedZeddeded eefdZxZS)GroundingDataseta' Dataset class for object detection tasks using annotations from a JSON file in grounding format. This dataset is designed for grounding tasks where annotations are provided in a JSON file rather than the standard YOLO format text files. Attributes: json_file (str): Path to the JSON file containing annotations. Methods: get_img_files: Return empty list as image files are read in get_labels. get_labels: Load annotations from a JSON file and prepare them for training. build_transforms: Configure augmentations for training with optional text loading. Examples: >>> dataset = GroundingDataset(img_path="path/to/images", json_file="annotations.json", task="detect") >>> len(dataset) # Number of valid images with annotations r+)r. json_filer.rcR|dvsJd||_t|||ddid|y)a Initialize a GroundingDataset for object detection. Args: json_file (str): Path to the JSON file containing annotations. task (str): Must be 'detect' or 'segment' for GroundingDataset. *args (Any): Additional positional arguments for the parent class. **kwargs (Any): Additional keyword arguments for the parent class. >r+r0zEGroundingDataset currently only supports `detect` and `segment` tasksr3rI)r.r-N)rr7r8)r:r.rr;r<r=s r>r8zGroundingDataset.__init__s;,,u.uu," $TQJ6Jr?img_pathrBcgS)a  The image files would be read in `get_labels` function, return empty list here. Args: img_path (str): Path to the directory containing images. Returns: (list): Empty list as image files are read in get_labels. )r:r!s r> get_img_fileszGroundingDataset.get_img_filess  r?rDNc  ddddd}td|D}|jD]5\}}||jvs||k(sJd|jd|d |d y tjd |jdy ) a Verify the number of instances in the dataset matches expected counts. This method checks if the total number of bounding box instances in the provided labels matches the expected count for known datasets. It performs validation against a predefined set of datasets with known instance counts. Args: labels (List[Dict[str, Any]]): List of label dictionaries, where each dictionary contains dataset annotations. Each label dict must have a 'bboxes' key with a numpy array or tensor containing bounding box coordinates. Raises: AssertionError: If the actual instance count doesn't match the expected count for a recognized dataset. Note: For unrecognized datasets (those not in the predefined expected_counts), a warning is logged and verification is skipped. iL7i+8i i )final_mixed_train_no_coco_segmfinal_mixed_train_no_coco"final_flickr_separateGT_train_segmfinal_flickr_separateGT_trainc3@K|]}|djdyw)rUrN)rS)rrs r>rz1GroundingDataset.verify_labels..sJ%U8_2215Js'z' has z instances, expected .Nz?Skipping instance count verification for unrecognized dataset ')rrrrrq)r:rDexpected_countsinstance_count data_namecounts r> verify_labelszGroundingDataset.verify_labelss,/6)028-3  J6JJ / 5 5 7  IuDNN*%.w!DNN3C6.IYYnotnuuv0ww.  XY]YgYgXhhijkr?r@rAc  dgi}tjdt|j5}t j |}ddddDcic] }|dd| }}t t}|dD]}||dj|t|jd |j D]X\}} ||d} | d | d | d }} } t|j|z } | jsF|jjt| g}g}i}g}| D]}|dr t!j"|dt j$}|ddxxx|dddz z ccc|ddgxxt'| zcc<|ddgxxt'| zcc<|ddks|ddkr| d}dj)|dDcgc] }||d|dc}j+j-}|s||vr t/|||<|j|g||}|g|j1z}||vs|j||j3d=t/|ddk(r|j|at/|ddkDrkt5|d}t!j6|dt!j"| | gt j$z j9dj1}n|dDcgc] }|D]}| }}}t!j"|t j$j9ddt!j"| | gt j$z j9dj1}|g|z}|j|t/|r%t!j"|t j$n$t!j:dt j$}|rt!j"|Dcgc]}|d c}t j$}|Dcgc]:}t!j"|ddt j$j9dd<}}t!j6|j9ddt=|fd}t!j"|t j$}dj| | | f|ddddf|ddddf|dd|d[t?|jd<tA|jB||tD|S#1swY xYwcc}wcc}wcc}}wcc}wcc}w) a! Load annotations from a JSON file, filter, and normalize bounding boxes for each image. Args: path (Path): Path where to save the cache file. Returns: (Dict[str, Any]): Dictionary containing cached labels and related information. rDzLoading annotation file...Nimagesidr annotationsimage_idzReading annotations )rOheightwidth file_nameiscrowdbboxrrHrrrIcaptionrZtokens_positive segmentationrr)rTrQ)rRrSrTrUrVrXrYrr_)#rroopenrjsonloadrrrmrrrr!rrfrrarrayrfloatrplowerr retolistrgr concatenatereshaperrr#r&rbrr)r:rArsfr5r3 img_to_annsannimg_idannsrhwrRrUrVcat2idrboxr<rcat_namerTrrjr|classess r>rzGroundingDataset.cache_labelssrN 01 $.. ! 'Q))A,K '-8-BCQtWQK!#CC!$' }- 5C J ( / / 4 5 !2!2!4=QRVR`R`Qa;bcD LFDF1:'C(mS\3{3C!qA4==)A-G>># MM W .FHFE% +y>hhs6{"**=BQ3qr7Q;&QF uQx' QF uQx' q6Q;#a&A+i.883GXCY$ZaWQqTAaD%9$Z[aaciik6)'*6{F8$LL(,X&ecjjl*f$MM#&ww~.:s>23q8$OOC0$ ^!459 3C4G HA!#!:RXXq!fTVT^T^=_!_ h hik l s s uA,/,? Kq KA K KA K!#!2::!>!F!Fr1!MPRPXPXZ[]^Y_gigqgqPr!r!(!' !EAI *K% +L8;6{& 3QW_a_i_iHjB(((#;QAaD#;2::NV^_QRBHHQqrU"**=EEb!L__^^W__R%;^H=U$VXYZ"BJJ/B hK  &Va1f: ABi ("&#)"  sD JT^^,&  T16KL[ ' 'C4%[&!L$<_s)UUU! U& ' U, ?U1Uc@t|jjd} t|d}}|dtk(sJ|dt |jk(sJ dDcgc]}|j|c}|d}|j||Dcgc]}t|dc}|_td vr%t!j"d |jd ||S#t tttf$r|j|d}}YwxYwcc}wcc}w) z Load labels from cache or generate them from JSON file. Returns: (List[dict]): List of label dictionaries, each containing information about an image and its annotations. rTrr_F)r_rrDrRrzLoad z from cache file )rrrr%rrr#rrrModuleNotFoundErrorrrr1rrfrrro)r:rrrrrDrs r>rzGroundingDataset.get_labelsMs $..)55h?  <.z:D1E#'<< <<=HT^^$<< << 33!13x 6">CVW <((4e1E <3Ds8C$#DD$,DDrct||}|jr9tdd|j |j }|j d||S)a  Configure augmentations for training with optional text loading. Args: hyp (dict, optional): Hyperparameters for transforms. Returns: (Compose): Composed transforms including text augmentation if applicable. rTrr)r7rrrrrrrs r>rz!GroundingDataset.build_transformscsYW-c2 << '"11$2D2DEI   b) ,r?c |jDchc]#}|dD]}|D]}|j%c}}}Scc}}}w)z.Return unique category names from the dataset.r)rDr )r:rrrs r>r zGroundingDataset.category_names{sC(,{{\\euW~\tW[\RS \ \ \\\s(>ctt}|jD]0}|dD]&}|D]}|j}||xxdz cc<!(2|S)r rr)rrrDr )r:rrrrs r>rzGroundingDataset.category_freqsf$C( [[ *Eg **A A!!$)$* * * r?rrc`|jDcgc] \}}||k\s |c}}Scc}}wrrrs r>rzGroundingDataset._get_neg_textsrrrr)rrrrrr8r r$rrr1rrrr rrrr rrrrrrs@r>rrs &+3R KC Ks K c d !lDc3h$8!lT!lF)--=(>YY4S>YvDJ,HTNg0]]GdGsGT#YGGr?rc@eZdZdZedeedefdZdeddfdZy)YOLOConcatDataseta Dataset as a concatenation of multiple datasets. This class is useful to assemble different existing datasets for YOLO training, ensuring they use the same collation function. Methods: collate_fn: Static method that collates data samples into batches using YOLODataset's collation function. Examples: >>> dataset1 = YOLODataset(...) >>> dataset2 = YOLODataset(...) >>> combined_dataset = YOLOConcatDataset([dataset1, dataset2]) rrBc,tj|S)r)r*r)rs r>rzYOLOConcatDataset.collate_fns%%e,,r?rNcb|jD] }t|ds|j|"y)z Set mosaic, copy_paste and mixup options to 0.0 and build transformations. Args: hyp (dict): Hyperparameters for transforms. rN)datasetshasattrr)r:rdatasets r>rzYOLOConcatDataset.close_mosaics2}} &G7N3   % &r?) rrrrrr rrrr#r?r>r\r\s@  -$t* - - - & & &r?r\c"eZdZdZfdZxZS)SemanticDatasetzSemantic Segmentation Dataset.c"t|y)z$Initialize a SemanticDataset object.Nr)r:r=s r>r8zSemanticDataset.__init__s r?)rrrrr8rrs@r>rcrcs(r?rccTeZdZdZd dededefdZdedefdZ defd Z de e fd Z y ) ClassificationDataseta Dataset class for image classification tasks extending torchvision ImageFolder functionality. This class offers functionalities like image augmentation, caching, and verification. It's designed to efficiently handle large datasets for training deep learning models, with optional image transformations and caching mechanisms to speed up training. Attributes: cache_ram (bool): Indicates if caching in RAM is enabled. cache_disk (bool): Indicates if caching on disk is enabled. samples (list): A list of tuples, each containing the path to an image, its class index, path to its .npy cache file (if caching on disk), and optionally the loaded image array (if caching in RAM). torch_transforms (callable): PyTorch transforms to be applied to the images. root (str): Root directory of the dataset. prefix (str): Prefix for logging and cache filenames. Methods: __getitem__: Return subset of data and targets corresponding to given indices. __len__: Return the total number of samples in the dataset. verify_images: Verify all images in dataset. rootrrbc ddl}tr#|jj|d|_n!|jj||_|jj |_|jj |_|rL|jdkr=|j dtt|j |jz|_|rt|dnd|_ |jduxs&t|jjd k(|_|jrt!j"d d |_t|jjd k(|_|j'|_|j Dcgc]-}t)|t+|dj-d dgz/c}|_d|j.z df}|rjt1|j2||j4|j6|j8|j:|j<|j>|j@ |_"ytC|j2|_"ycc}w)ab Initialize YOLO classification dataset with root directory, arguments, augmentations, and cache settings. Args: root (str): Path to the dataset directory where images are stored in a class-specific folder structure. args (Namespace): Configuration containing dataset-related settings such as image size, augmentation parameters, and cache settings. augment (bool, optional): Whether to apply augmentations to the dataset. prefix (str, optional): Prefix for logging and cache filenames, aiding in dataset identification. rNT)rg allow_empty)rgg?z: rramzClassification `cache_ram` training has known memory leak in https://github.com/ultralytics/ultralytics/issues/9824, setting `cache_ram=False`.Fdiskz.npy) sizescalehflipvfliperasing auto_augmenthsv_hhsv_shsv_v)rl)# torchvisionrr_ ImageFolderbasesamplesrgfractionroundrerrbrrrE cache_ramrrq cache_disk verify_imagesrrrrmrrfliplrflipudrprqrrrsrtrtorch_transforms)r:rgr;rrbrursrms r>r8zClassificationDataset.__init__s  #,,88dPT8UDI#,,88d8CDIyy(( IINN  t}}s*<<(R%DLL0ADMM0Q*RSDL17h&}-R t+Os4::/D/D/F%/O >> NNe #DNdjj///1V;))+ RVR^R^_QQ4!:#9#9&#A4"HH_ tzz!3' #ZZkkkk !..jjjjjj  %$**5 `s2I.rrBc0|j|\}}}}|jr*|tj|x}|j|d<n|jr_|j s9t j|jtj|dt j|}ntj|}tjtj|tj}|j|}||dS)z Return subset of data and targets corresponding to given indices. Args: i (int): Index of the sample to retrieve. Returns: (dict): Dictionary containing the image and its class index. rIF) allow_pickle)rrT)rxr{cv2imreadr|rrsaveas_posixrBr fromarraycvtColor COLOR_BGR2RGBr)r:rrIrSfnimsamples r> __getitem__z!ClassificationDataset.__getitem__s||A 1b" >>z*-**Q-7T\\!_Q' __99; szz!}5IBAB __S\\"c.?.?@ A&&r*a((r?c,t|jS)z2Return the total number of samples in the dataset.)rerx)r:s r>__len__zClassificationDataset.__len__(s4<<  r?c |jd|jd}t|jjd} t |j ddDcgc]\}}| c}}|jt |}|dtk(sJ|dt|j Dcgc]}|d  c}k(sJ|jd \}}} } td vrF|d |d |d} td| | | |dr'tjdj|d| Scc}}wcc}w#tt t"f$rzd d ggif\}}} } }t%t&5} | j)t*t-|j t/|j}t||t1|j }|D]G\}}}}|r| j3||r| j3|||z }||z }|d |d |d|_I|j7dddn #1swYnxYw| r$tjdj| t|j Dcgc]}|d  ncc}wc}d<||t1| | f|d <| |d<t9|j||t| cYSwxYw)z Verify all images in dataset. Returns: (list): List of valid samples after verification. rErFrNr?)rbrr_rr`rrZr[r]rrar^rKrN)rbrgrrr"rxr%rrr#rrrrrorprrrrrrir'rjrrermrOrnr&)r:rOrAfilerrrsrurwrrxrrarzr`r{rrrrs r>r}z#ClassificationDataset.verify_images,s++i {#6DII**84"  T\\"15EF qtFt{{ [+D1E#'<< <<=HDLL-Iqad-I$JJ JJ!&9!5 BAwW$fAbT2$h7TA6=KK %- 89NG.J">>B '(!RR'7 $BD'1K( D))DLLRXY]YdYdRe@f)gG$c$,,6GH/3D+FD$v. C($JB$JB#'&"Yrd( CDID     DIIdO,  !=1!A$!=!=>AfIr3w<8AiLAfI #DKKq:O PN+ sYD1 D& %AD1) D, 5A0D1& D111K"BrfrfsQ,2 S2 2 s2 h)S)T)2!!,tE{,r?rf)@rA collectionsr itertoolsrmultiprocessing.poolrpathlibrtypingrrr r r rnumpyrrPILr torch.utils.datar ultralytics.utilsrrrrrultralytics.utils.instancerultralytics.utils.opsrrultralytics.utils.torch_utilsrrrrrrrrrrwr converterr rr!r"r#r$r%r&r'r(rrr*rrr\rcrfr#r?r>rs #+33  *MM0C:*    K+K\aGKaGHmG{mG`'& '&VkTTr?