o .wi1&@sddlZddlZddlZddlmZddlmZddlmZddl m Z m Z m Z m Z mZmZmZddlZddlZddlmZddlmZddlmZdd lmZmZmZes\es\d d gZ d?deeeefdeeeefdee de e ddffde!ddf ddZ"dedefddZ# d@dee de e ddffde e ddffddZ$de dde e%e%effddZ&Gd d!d!Z'd"e(ddfd#d$Z) dAdee dee ddffd%ed&e e(d'e'd(eee fd)e!de eeee ffd*d+Z*d,e ed-e ede e(fd.d/Z+d'e'd-e ed0e ed1e ed2e ed3e ed,e ed4e ed5e ed6e edee dee ddffd7e d8d9e e,d:e e,d&e e(d;e!dZ-dS)BN)Sequence)version) ModuleType)AnyDictListLiteralOptionalTupleUnion)apply_to_collection)Tensor)rank_zero_warn)_FASTER_COCO_EVAL_AVAILABLE_PYCOCOTOOLS_AVAILABLE _PYCOCOTOOLS_GREATER_EQUAL_2_0_9CocoBackend.tm_to_cocoCocoBackend.coco_to_tmbboxFpredstargetsiou_typersegm. ignore_scorereturncst|tr|f}dddtfdd|Dr td|dfdd |D}t|ts5td |t|tsAtd |t|t|krWtd t|d t|g|d|sadgngD]tfdd|Drytddqdg|dD]tfdd|Drtddq|D]tfdd|Dstddq|stdd|Dstdtdd|Dstd|D]tfdd|Dstddqtdd|Dstdt|D]3\}}|D],| d |d d kr%td!d"|d#| d d$|d d d% qq|r,d&St|D]J\}}|D]B| d |d d krS|d d kswntd!d'|d(| d d)|d d d*|d d d% q6q0d&S)+z9Ensure the correct input format of `preds` and `targets`.boxesmasksrc3|]}|vVqdSN.0tpname_mapr [/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/torchmetrics/detection/helpers.py 4z#_input_validator..z IOU type z is not supportedcsg|]}|qSr r r!r$r r& 6sz$_input_validator..z:Expected argument `preds` to be of type Sequence, but got z;Expected argument `target` to be of type Sequence, but got zHExpected argument `preds` and `target` to have the same length, but got z and labelsscoresc3|]}|vVqdSrr r"pkr r&r'Br(z.Expected all dicts in `preds` to contain the `z` keyc3r,rr r-r/r r&r'Fr(z/Expected all dicts in `target` to contain the `c3|] }t|tVqdSr isinstancer r"predivnr r&r'Jz Expected all z in `preds` to be of type Tensorcs|] }t|dtVqdS)r+Nr2r4r r r&r'Lr8z3Expected all scores in `preds` to be of type Tensorcsr9r*Nr2r4r r r&r'Nr8z3Expected all labels in `preds` to be of type Tensorc3r1rr2r"targetr6r r&r'Qr8z! in `target` to be of type Tensorcsr9r:r2r;r r r&r'Sr8z4Expected all labels in `target` to be of type TensorrzInput 'z' and labels of sample z. in targets have a different length (expected z labels, got )Nz', labels and scores of sample z2 in predictions have a different length (expected z labels and scores, got z labels and ) r3strany Exceptionr ValueErrorlenall enumeratesize)rrrr item_val_nameiitemr )r7r0r%r&_input_validator)s      8   rIrcCs$|dkr|jdkr|dS|S)zIEmpty tensors can cause problems in DDP mode, this methods corrects them.r)numelndim unsqueeze)rr r r&_fix_empty_tensorsis rNcsBdt|tr |f}tfdd|Drtdd||S)z+Validate that iou type argument is correct.)rrc3rrr r!allowed_iou_typesr r&r'wr(z)_validate_iou_type_arg..z*Expected argument `iou_type` to be one of z or a tuple of, but got )r3r>r?rA)rr rOr&_validate_iou_type_argps rQbackend pycocotoolsfaster_coco_evalcCs||dkr!ts tdddlm}ddlm}ddlm}|||fSts'tdddl m}ddl m }dd l m}|||fS) z-Load the backend tools for the given backend.rTzBackend `pycocotools` in metric `MeanAveragePrecision` metric requires that `pycocotools` is installed. Please install with `pip install pycocotools` or `pip install torchmetrics[detection]`rN)COCO)COCOevalzBackend `faster_coco_eval` in metric `MeanAveragePrecision` metric requires that `faster-coco-eval` is installed. Please install with `pip install faster-coco-eval`.)COCOeval_faster)mask) rModuleNotFoundErrorpycocotools.maskrYpycocotools.cocorVpycocotools.cocoevalrWrrUrXfaster_coco_eval.core)rR mask_utilsrVrWr r r&_load_coco_backend_tools~s"        r`c@seZdZdZdedddfddZedefdd Zedefd d Z edefd d Z  d7de e de e de e de e de e de e de e de e de e de edeeddffdeddeeeffddZd eed!ed"eedeee ffd#d$Ze  %d8d&ed'ede edeeddffdeddeeeee feeee fff d(d)Z *  d9de e de e de e de e de e de e de e de e de e d+ede edeeddffdedddfd,d-Z       d:d.e e d/e e d0ee e d1ee e d2ee e d3ee e d4ee e de edeeddffdeddefd5d6ZdS); CocoBackendaIBackend implementation for COCO-style Mean Average Precision (mAP) calculation. This class provides the core functionality for evaluating object detection and instance segmentation predictions using the Common Objects in Context (COCO) evaluation protocol. It supports both the standard 'pycocotools' and optimized 'faster_coco_eval' backends. It's used for calculation of mAP in MeanAveragePrecision class. It's a backend that abstracts away the mAP calculation with coco package Args: backend (str): Either 'pycocotools' or 'faster_coco_eval' rRrSrNcCs |dvr td|||_dS)NrSzUExpected argument `backend` to be one of ('pycocotools', 'faster_coco_eval') but got )rArR)selfrRr r r&__init__s  zCocoBackend.__init__cCst|j\}}}|S)z.Returns the coco module for the given backend.r`rR)rbcoco_r r r&rezCocoBackend.cococCst|j\}}}|S)z3Returns the coco eval module for the given backend.rd)rbrfcocoevalr r r&rhrgzCocoBackend.cocoevalcCst|j\}}}|S)z4Returns the mask utils object for the given backend.rd)rbrfr_r r r&r_rgzCocoBackend.mask_utilsrmicrogroundtruth_labelsgroundtruth_boxgroundtruth_maskgroundtruth_crowdsgroundtruth_areadetection_labels detection_boxdetection_maskdetection_scoresrr.averagemacrorjc  Cs2| dkrt|tdd}t|tdd}||} } t|dks)t|dkr6t||ng}|j |t|dkrC|ndt|dkrL|nd||| || d| _ |j |t|dkra|ndt|dkrj|nd| | || d| _ t t | | Wd| | fS1swY| | fS) z=Returns the coco datasets for the target and the predictions.rjcS t|Srtorch zeros_likexr r r& z0CocoBackend._get_coco_datasets..cSrwrrxr{r r r&r}r~rN)r*rrcrowdsarear all_labelsrt)r*rrr+rrrt)r r rerBrycatuniquecputolist_get_coco_formatdataset contextlibredirect_stdoutioStringIO createIndex)rbrkrlrmrnrorprqrrrsrrt coco_target coco_predsrr r r&_get_coco_datasetssD    zCocoBackend._get_coco_datasetsstatsprefixmax_detection_thresholdscCsp|}|dtj|dgtjd|dtj|dgtjd|dtj|dgtjd|dtj|d gtjd|d tj|d gtjd|d tj|d gtjd|d|dtj|dgtjd|d|dtj|dgtjd|d|dtj|dgtjd|dtj|dgtjd|dtj|dgtjd|dtj|dgtjdi S)z;Converts the output of COCOeval.stats to a dict of tensors.maprdtypemap_50rJmap_75 map_small map_medium map_largemar_ mar_small mar_medium mar_large rytensorfloat32)rbrrrmdtr r r&_coco_stats_to_tensor_dicts$$$z&CocoBackend._coco_stats_to_tensor_dictrTrrcCst|}t|\}}}tt||}||}Wdn1s&wY|jd}|jd} i} |D]t} | d| vrcgggd| | d<d|vrWg| | dd<d|vrcg| | dd<d|vrt| | dd| dd|vr| | dd| | | | dd | d | | dd | d | | dd | d q9i} | D]f} | d| vrggd | | d<d|vrg| | dd<d|vrg| | dd<d|vr| | dd| dd|vr| | dd| | | | dd| d| | dd | d q| D]%}|| vr?ggd | |<d|vr4g| |d<d|vr?g| |d<qgg}}| D]}t j | |dt j dt j | |d t j dd }d|vryt j t| |dt j d|d<d|vrt j t| |dt jd|d<||t j | |d t j dt j | |d t j dt j | |d t j dd}d|vrt j | |dt j d|d<d|vrt j t| |dt jd|d<||qH||fS)aUtility function for converting .json coco format files to the input format of the mAP metric. The function accepts a file for the predictions and a file for the target in coco format and converts them to a list of dictionaries containing the boxes, labels and scores in the input format of mAP metric. Args: coco_preds: Path to the json file containing the predictions in coco format coco_target: Path to the json file containing the targets in coco format iou_type: Type of input, either `bbox` for bounding boxes or `segm` for segmentation masks backend: Backend to use for the conversion. Either `pycocotools` or `faster_coco_eval`. Returns: A tuple containing the predictions and targets in the input format of mAP metric. Each element of the tuple is a list of dictionaries containing the boxes, labels and scores. Example: >>> # File formats are defined at https://cocodataset.org/#format-data >>> # Example files can be found at >>> # https://github.com/cocodataset/cocoapi/tree/master/results >>> from torchmetrics.detection import MeanAveragePrecision >>> preds, target = MeanAveragePrecision().coco_to_tm( ... "instances_val2014_fakebbox100_results.json", ... "val2014_fake_eval_res.txt.json" ... iou_type="bbox" ... ) # doctest: +SKIP N annotationsimage_id)r*iscrowdrrrrrr* category_idrr)r+r*r+scorer)rQr`rrrrloadResrappend annToMaskryrrint32nparrayuint8)rrrrRrerfgtdt gt_dataset dt_datasetr<trr.r0 batched_predsbatched_targetkeybpbtr r r& coco_to_tm s"            " "   "r tm_map_inputnamec  Cst|dks t|dkrt||ng} |j|t|dkr&|ndt|dkr/|nd||| | | d}|j|t|dkrC|ndt|dkrL|nd| | | | d}d| vrt|dtddd |d<t|dt j t j fd dd |d<t|td dd }t|t j t j fd dd }t j |dd d}t j |d d}t| dd }||Wdn1swYt| dd}||WddS1swYdS)aUtility function for converting the input for mAP metric to coco format and saving it to a json file. This function should be used after calling `.update(...)` or `.forward(...)` on all data that should be written to the file, as the input is then internally cached. The function then converts to information to coco format and writes it to json files. Args: groundtruth_labels: List of tensors containing the ground truth labels groundtruth_box: List of tensors containing the ground truth bounding boxes groundtruth_mask: List of tensors containing the ground truth segmentation masks groundtruth_crowds: List of tensors indicating whether ground truth annotations are crowd annotations groundtruth_area: List of tensors containing the area of ground truth annotations detection_labels: List of tensors containing the predicted labels detection_box: List of tensors containing the predicted bounding boxes detection_mask: List of tensors containing the predicted segmentation masks detection_scores: List of tensors containing the confidence scores for predictions name: Name of the output file, which will be appended with "_preds.json" and "_target.json" iou_type: Type of IoU calculation to use. Can be either "bbox" for bounding box or "segm" for segmentation average: Type of averaging to use. Can be either "macro" or "micro" Example: >>> from torch import tensor >>> from torchmetrics.detection import MeanAveragePrecision >>> preds = [ ... dict( ... boxes=tensor([[258.0, 41.0, 606.0, 285.0]]), ... scores=tensor([0.536]), ... labels=tensor([0]), ... ) ... ] >>> target = [ ... dict( ... boxes=tensor([[214.0, 41.0, 562.0, 285.0]]), ... labels=tensor([0]), ... ) ... ] >>> metric = MeanAveragePrecision(iou_type="bbox") >>> metric.update(preds, target) >>> metric.tm_to_coco("tm_map_input") rN)r*rrrrrrrt)r*rrr+rrrtrrcS |dSNzutf-8decoder{r r r&r}r~z(CocoBackend.tm_to_coco..)rfunctioncSt|Srintr{r r r&r}cSrrrr{r r r&r}r~cSrrrr{r r r&r}rr)indentz _preds.jsonwz _target.json)rBryrrrrrr bytesruint32uint64jsondumpsopenwrite)rbrkrlrmrnrorprqrrrsrrrtrtarget_dataset preds_dataset preds_json target_jsonfr r r& tm_to_cocozsX:       "rr*rrrr+rrc  Csg} g} d} t|D]Y\} }|dur|| }|}|dur0|| }t|dkr0|dur0q |}| d| id|vr`t|dkr`|ddd|ddd| dd<| dd<t|D]\}}|durp||}|durt|dkr||}|d|dd }d |vrt|d krtd | d |dt|dt|tstd| d |dt|dd}d}|dur|| |dkr|| |}n&d|vr|j |n|d|d}t|dkr|d|d}|j |}| | |||dur|| |ndd}|dur||d<||d<|dur(||d <|dur1||d<|durZ|| |}t|t sVtd| d |dt|d||d<| || d7} qdq | dkrqdd|Dndddg}| | |d }t rd!d"t d#i|d$<|S)%zTransforms and returns all cached targets or predictions in COCO format. Format is defined at https://cocodataset.org/#format-data rJNridrheightwidth)rEcountsrrzInvalid input box of sample z , element z (expected 4 values, got r=zInvalid input class of sample z+ (expected value of type integer, got type rr)rrrrr area_bbox area_segm segmentationzInvalid input score of sample z) (expected value of type float, got type rrjcSsg|] }|t|dqS)rr)r>)r"rGr r r&r)=sz0CocoBackend._get_coco_format..0r)imagesr categories descriptionz:Dummy info generated by tm_to_coco to support pycocotools rTinfo)rDrrrBrrAr3rtyper_rfloatrr)rbr*rrrr+rrrrtrr annotation_idr image_labels image_boxes image_masksr0 image_label image_box image_mask area_stat_boxarea_stat_mask area_stat annotationrclassesresultr r r&rs  2  $          $7zCocoBackend._get_coco_format)rirj)rirT)rrirj)NNNNNrirj)__name__ __module__ __qualname____doc__rrcpropertyobjectrerhr_rr r tuplerlistrr>rdictr staticmethodrrr rr r r r&ras     6  " z      l        ralimitcCstd|dtdS)NzEncountered more than aY detections in a single image. This means that certain detections with the lowest scores will be ignored, that may have an undesirable impact on performance. Please consider adjusting the `max_detection_threshold` to suit your use case. To disable this warning, set attribute class `warn_on_many_detections=False`, after initializing the metric.)r UserWarning)rr r r&_warning_on_too_many_detectionsJs r box_formatr coco_backendrHwarnc sddlm}ddgd|vr%t|d}|dkr!|||dd}|d<d|vrRg}|d D]} |jt | } | t | d | d fq3t |d <d t dt ffdd } |rn| dsh| d rntdS)a.Convert and return the boxes or masks from the item depending on the iou_type. Args: iou_type: Type of input to process. Supported types are: - "bbox": Process bounding boxes - "segm": Process segmentation masks box_format: Input format of given boxes. Supported formats are: - 'xyxy': boxes are represented via corners, x1, y1 being top left and x2, y2 being bottom right. - 'xywh': boxes are represented via corner, width and height, x1, y2 being top left, w, h being width and height. - 'cxcywh': boxes are represented via centre, width and height, cx, cy being center of box, w, h being width and height. max_detection_thresholds: List of thresholds on maximum detections per image. Used to determine if warnings should be raised when the number of detections exceeds these thresholds. coco_backend: The COCO evaluation backend class type to use for processing the items. item: Input dictionary containing the boxes or masks to be processed, along with other detection information. warn: Whether to warn if the number of boxes or masks exceeds the max_detection_thresholds. Default is False. Returns: A tuple containing processed boxes or masks depending on the iou_type. The first element is the tensor representation, and the second element contains additional metadata if applicable. r) box_convertNrrxywh)in_fmtout_fmtrrrErrJidxrcs$|}|dur dSt|dkS)NFr)rB)rvalroutputr r&_valid_output_lensz0_get_safe_item_values.._valid_output_lenr)torchvision.opsr rNrKrnumpyr_encoderasfortranarrayrrrboolr) rr rr rHr r rrrGrlerr rr&_get_safe_item_valuesTs" &    rrprkcCs6t|dks t|dkrt||SgS)Nr)rBryrrrrrprkr r r& _get_classessrrlrmrnrorqrrrsrtruiou_thresholdsrec_thresholds class_metricsextended_summaryc Cs|j||||||||| | | d \}}i}tt| D]}t| dkr*dn|d}t| dkrF|jdD] }|d||d<q:t|jdksTt|jdkrc||j d d g||d q|j |||d }t j | t j d |j_t j | t j d |j_||j_||||j}||j |||d i}|r|dt|jt jdd|dt|jd|dt|jd|dt|jdi}|||rl|j||||||||| | dd \}}|j |||d }t j | t j d |j_t j | t j d |j_||j_g}g}t||dD]E}|g|j_tt||||j}Wdn 1s>wY|t|dg|t|dgqtj|tjd }tj|tjd }ntjdgtjd }tjdgtjd }t| dkrdn|d}||d||d|dd|iqWdn 1swY|dtjt||dtj d i|S)N)rtrJrfrarea_rr g)rr)iouTyperiouscSstj|tjdS)Nrrr{r r r&r}sz*_calculate_map_with_coco.. precisionrecallr+rvrrr map_per_classr _per_classr)!rrrrrrBrimgsupdaterrhrrfloat64paramsiouThrsrecThrsmaxDetsevaluate accumulate summarizerr r&ndarrayryrevalrcatIdsrrr) r rkrlrmrnrorprqrrrsrrtrrrr r!rr result_dicti_typeranno coco_evalrsummarymap_per_class_listmar_per_class_listclass_id class_statsmap_per_class_valuesmar_per_class_valuesr r r&_calculate_map_with_cocos          \rC)rFri)F).rrrcollections.abcrimportlib.metadatartypesrtypingrrrrr r r rrrylightning_utilitiesr r torchmetrics.utilitiesrtorchmetrics.utilities.importsrrr__doctest_skip__rr>rrrIrNrQrr`rarrrrrrCr r r r&s    $     @ 5  "@