Ei~SSKrSSKrSSKJr SSKJr SSKJrJrJ r SSK r SSK J r J r SSK JrJrJr SSKJr SS KJr SS KJr S S KJrJrJr S S KJr S SKJ r J!r! S SK"J#r#J$r$ SSK%Jr& SSKJ'r'J(r( SSK)J*r* SSK+J,r,J-r- SSK.J/r/ /SQr0S\1\ S\ 4Sjr2Sr3Sr4"SS\ Rj5r6"SS\ Rj5r7"S S!\ Rj5r8"S"S#\ Rj5r9\S$S%.r:"S&S'\5r;"S(S)\5r<\"5\!"S*\;Rz4S+\$R|4S,9SS-S\$R|SS..S/\ \;S0\?S1\ \@S2\ \$S3\ \@S4\S\94S5jj55rA\"5\!"S*\%_default_anchorgen..3s,pXoSTS^!34c!n:L6MNXos)+) @i))??g@)tuplelenr) anchor_sizes aspect_ratiosanchor_generators r-_default_anchorgenrN2s3pXoppL&\)::M&|C r/ch^\rSrSrSrS S\\S\R44U4Sjjjr Sr Sr Sr U=r $) RetinaNetHead9aY A regression and classification head for use in RetinaNet. Args: in_channels (int): number of channels of the input feature num_anchors (int): number of anchors to be predicted num_classes (int): number of classes to be predicted norm_layer (callable, optional): Module specifying the normalization layer to use. Default: None norm_layer.c^>[TU]5 [XX4S9Ul[ XUS9Ulg)NrR)super__init__RetinaNetClassificationHeadclassification_headRetinaNetRegressionHeadregression_head)self in_channels num_anchors num_classesrR __class__s r-rVRetinaNetHead.__init__Ds2 #> k$   7{\fgr/ctURRXU5URRXX45S.$)N)classificationbbox_regression)rX compute_lossrZ)r[targets head_outputsanchors matched_idxss r-rdRetinaNetHead.compute_lossKs<#66CCG[gh#33@@X_n  r/cHURU5URU5S.$)N) cls_logitsrcrXrZ)r[r's r-forwardRetinaNetHead.forwardRs$"66q9dNbNbcdNeffr/rlN)__name__ __module__ __qualname____firstlineno____doc__rrrModulerVrdrm__static_attributes__ __classcell__r_s@r-rPrP9sEh(S[\_acajaj\jSkJlhh ggr/rPc|^\rSrSrSrSrS S\\S\R44U4Sjjjr U4Sjr Sr S r S rU=r$) rWWaJ A classification head for use in RetinaNet. Args: in_channels (int): number of channels of the input feature num_anchors (int): number of anchors to be predicted num_classes (int): number of classes to be predicted norm_layer (callable, optional): Module specifying the normalization layer to use. Default: None rrR.c >[T U]5 /n[S5H'nUR[R "XUS95 M) [ R"U6UlURR5Hn[U[ R5(dM$[R RRURSS9 UR cMe[R RR#UR S5 M [ R"XU-SSSS9Ul[R RRUR$RSS9 [R RR#UR$R [&R("SU- U- 5*5 X0lX l[.R0R2Ulg) Nr1rT{Gz?stdrr r kernel_sizestridepadding)rUrVr5append misc_nn_opsConv2dNormActivationr Sequentialconvmodules isinstanceConv2dtorchinitnormal_r2r3 constant_rkmathlogr^r] det_utilsMatcherBETWEEN_THRESHOLDS) r[r\r]r^prior_probabilityrRr_layerr_s r-rV$RetinaNetClassificationHead.__init__ds[ qA KK 88^hi jMM4( YY&&(E%++ %%ell%=::)HHMM++EJJ: ) ))K{1JXYbcmno  doo44$?   4 4txxEVAVZk@k7l6lm&& #,"3"3"F"Fr/c |>URSS5nUbUS:a [X5 [T U] UUUUUUU5 gNversionrgetr<rU_load_from_state_dict r[r7r8local_metadatastrict missing_keysunexpected_keys error_msgsrr_s r-r1RetinaNetClassificationHead._load_from_state_dictL!$$Y5 ?gk j 1 %        r/c P/nUSn[XU5HyupgnUS:n U R5n [R"U5n SU U USX4'XR:gn UR [ X|XSS9[SU 5- 5 M{ [U5[U5- $)NrkrrHlabelssum) reductionr) ziprr zeros_likerrr maxr.rJ) r[rerfrhlossesrktargets_per_imagecls_logits_per_imagematched_idxs_per_imageforeground_idxs_per_imagenum_foregroundgt_classes_targetvalid_idxs_per_images r-rd(RetinaNetClassificationHead.compute_losss!,/ ORSZhtOu K 5K(>!(C %6::%;# a( ) !Pv2F|c'l**r/cf/nUHnURU5nURU5nURupVpxURUSURXx5nUR SSSSS5nUR USUR5nURU5 M [R"USS9$)Nrr r1rrdim) rrkshapeviewr^permutereshaperrcat) r[r'all_cls_logitsfeaturesrkNrHWs r-rm#RetinaNetClassificationHead.forwardsH8,J4J$))JA!#B0@0@!GJ#++Aq!Q:J#++Ar43C3CDJ  ! !* -yyQ//r/)rrkrr]r^)r|N)rprqrrrsrt_versionrrrrurVrrdrmrvrwrxs@r-rWrWWsVH9= G Xc299n56 GGB 2+B00r/rWc^\rSrSrSrSrS\R0rS S\ \ S\ R44U4Sjjjr U4SjrS rS rS rU=r$) rYa  A regression head for use in RetinaNet. Args: in_channels (int): number of channels of the input feature num_anchors (int): number of anchors to be predicted norm_layer (callable, optional): Module specifying the normalization layer to use. Default: None r box_coderrR.c >[TU]5 /n[S5H'nUR[R "XUS95 M) [ R"U6Ul[ R"XS-SSSS9Ul [R RRURRSS9 [R RRURR 5 URR#5Hn[%U[ R5(dM$[R RRURSS9 UR cMe[R RRUR 5 M [&R("SS 9UlS Ulg) Nr1rTr rrr|r}rHrHrHrHweightsl1)rUrVr5rrrrrrrbbox_regrrrr2zeros_r3rrrBoxCoderr _loss_type)r[r\r]rRrrrr_s r-rV RetinaNetRegressionHead.__init__s3 qA KK 88^hi jMM4(  +QAVWabc   dmm22=  T]]//0YY&&(E%++ %%ell%=::)HHMM((4 ) #++4HIr/c |>URSS5nUbUS:a [X5 [T U] UUUUUUU5 grrrs r-r-RetinaNetRegressionHead._load_from_state_dictrr/c /nUSn[XX45Hupxp[R"U S:5Sn U R5n USXn XSS24nXSS24n UR [ UR URU U U5[SU 5- 5 M [U5[S[U55- $)Nrcrr r) rrwherenumelrrrrrr.rJ)r[rerfrgrhrrcrbbox_regression_per_imageanchors_per_imagerrrmatched_gt_boxes_per_images r-rd$RetinaNetRegressionHead.compute_losss&'89gj gh c :K). 4Ja4O(PQR(S %6<<>N*;7)CDZDu)v &(A]^B^(_ % 1Q2N O  MMOONN%.- a() h 0F|c!S\222r/c>/nUHnURU5nURU5nURupVpxURUSSXx5nUR SSSSS5nUR USS5nUR U5 M [R"USS9$)Nrr1rr rrr) rrrrrrrrr) r[r'all_bbox_regressionrrcrrrrs r-rmRetinaNetRegressionHead.forward0s H"ii1O"mmO>> import torch >>> import torchvision >>> from torchvision.models.detection import RetinaNet >>> from torchvision.models.detection.anchor_utils import AnchorGenerator >>> # load a pre-trained model for classification and return >>> # only the features >>> backbone = torchvision.models.mobilenet_v2(weights=MobileNet_V2_Weights.DEFAULT).features >>> # RetinaNet needs to know the number of >>> # output channels in a backbone. For mobilenet_v2, it's 1280, >>> # so we need to add it here >>> backbone.out_channels = 1280 >>> >>> # let's make the network generate 5 x 3 anchors per spatial >>> # location, with 5 different sizes and 3 different aspect >>> # ratios. We have a Tuple[Tuple[int]] because each feature >>> # map could potentially have different sizes and >>> # aspect ratios >>> anchor_generator = AnchorGenerator( >>> sizes=((32, 64, 128, 256, 512),), >>> aspect_ratios=((0.5, 1.0, 2.0),) >>> ) >>> >>> # put the pieces together inside a RetinaNet model >>> model = RetinaNet(backbone, >>> num_classes=2, >>> anchor_generator=anchor_generator) >>> model.eval() >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)] >>> predictions = model(x) )rproposal_matcherc `>[TU]5 [U5 [US5(d [ S5eXl[ U[[S545(d[S[U535eUc [5nXpl Uc([URUR5SU5nXlU c[ R""U USS9n Xl[ R&"SS9UlUc/S QnUc/S Qn[+X4XV40UD6UlXlXlXlXlS Ulg) N out_channelszbackbone should contain an attribute out_channels specifying the number of output channels (assumed to be the same for all the levels)zFanchor_generator should be of type AnchorGenerator or None instead of rT)allow_low_quality_matchesrr)g ףp= ?gv/?gCl?)gZd;O?gy&1?g?F)rUrVrhasattr ValueErrorbackbonerrr9 TypeErrorrNrMrPrnum_anchors_per_locationheadrrrrrr! transform score_thresh nms_threshdetections_per_imgtopk_candidates _has_warned)r[rr^min_sizemax_size image_mean image_stdrMrrrrr fg_iou_thresh bg_iou_threshrkwargsr_s r-rVRetinaNet.__init__s@* D!x00+  ! *_d4j,IJJXY]^nYoXpq   #13  0 < !6!68H8a8a8cde8fhstD  #(00*.   !1"++4HI  .J  -I1(jf_ef($"4.!r/c,UR(aU$U$ro)training)r[r detectionss r- eager_outputsRetinaNet.eager_outputss ==Mr/c /n[X15HupVUSR5S:XaPUR[R"UR S54S[R URS95 Ml[R"USU5nURURU55 M URRXX45$)Nr rr)dtypedevice) rrrrfullsizeint64rbox_opsbox_iourrrd)r[rerfrgrhrrmatch_quality_matrixs r-rdRetinaNet.compute_losss 474I 0  )//1Q6##JJ 1 6 6q 9;Ru{{[l[s[st#*??3DW3MO`#a    5 56J K L5Jyy%%gWSSr/cnUSnUSn[U5n/n[U5GHnUV s/sHoUPM n n UV s/sHoUPM n n X(X8p/n/n/n[XU 5GH*unnnURSn[R "U5R 5nUUR:nUUn[R"U5Sn[R"UURS5nURU5unnUUn[R"UUSS9nUU-nURRUUUU5n[ R""UU5nUR%U5 UR%U5 UR%U5 GM- [R&"USS9n[R&"USS9n[R&"USS9n[ R("UUUUR*5nUSUR,nUR%UUUUUUS.5 GM U$s sn fs sn f) Nrkrcrrfloor) rounding_moder)r scoresr)rJr5rrrsigmoidflattenrrr _topk_minrtopkdivr decode_singlerclip_boxes_to_imagerr batched_nmsrr)r[rfrg image_shapes class_logitsbox_regression num_imagesrindexbrbox_regression_per_imagecllogits_per_imager image_shape image_boxes image_scores image_labelsbox_regression_per_levellogits_per_levelanchors_per_levelr^scores_per_level keep_idxs topk_idxsnum_topkidxs anchor_idxslabels_per_levelboxes_per_levelkeeps r-postprocess_detections RetinaNet.postprocess_detectionssZ#L1 %&78& .0 :&E !KK 215 %..y$:N:NPQR)9)>)>x)H& $%dO #ii ;gV #,{#: "&..">">,[9;L[;Y##*"="=o{"[""?3##$45##$455R8 ))KQ7K 99\q9L 99\q9L&&{L,PTP_P_`D1$112D   (.*40*40 ['ji(LAs H-H2c &UR(aUc[R"SS5 OUH}nUSn[R"[U[R5S5 [R"[ UR 5S:H=(a UR SS:HS 5 M /nUH^nUR S Sn[R"[ U5S:HS UR S S35 URUS US 45 M` URX5upUb[U5HupUSnUSS2SS24USS2SS24:*n U R5(dM8[R"U RS S95S S n XJR5n [R"SSU SUS35 M URUR5n [U [R5(a[SU 4/5n [!U R#55n UR%U 5n UR'X5n0n/nUR(a0Uc[R"SS5 GOUR)X-U5nGO U Vs/sH&nUR+S5UR+S5-PM( nnS nUHnUU- nM U SR+S 5nUU-nUVs/sHnUU-PM nn0nU H"n[!U UR-US S95UU'M$ UVs/sHn[!UR-U55PM nnUR/UUUR05nURR3UUR0U5n[R4R75(a2UR8(d[:R<"S5 SUlUU4$UR?UU5$s snfs snfs snf)a Args: images (list[Tensor]): images to be processed targets (list[Dict[Tensor]]): ground-truth boxes present in the image (optional) Returns: result (list[BoxList] or dict[Tensor]): the output from the model. During training, it returns a dict[Tensor] which contains the losses. During testing, it returns list[BoxList] contains additional fields like `scores`, `labels` and `mask` (for Mask R-CNN models). NFz0targets should not be none when in training moder z+Expected target boxes to be of type Tensor.rrr1z5Expected target boxes to be a tensor of shape [N, 4].zJexpecting the last two dimensions of the Tensor to be H and W instead got rrrzLAll bounding boxes should have positive height and width. Found invalid box z for target at index r40r rkzBRetinaNet always returns a (Losses, Detections) tuple in scriptingT) rr_assertrr rJrrr enumerateanyrtolistrtensorsrlistvaluesrrMrdrsplitr- image_sizes postprocessjit is_scriptingrwarningswarnr)r[imagesretargetr original_image_sizesimgval target_idxdegenerate_boxesbb_idxdegen_bbrrfrgrrr'num_anchors_per_levelHWvHWAAhwsplit_head_outputska split_anchorss r-rmRetinaNet.forward=s == e%WX%F"7OEMM*UELL"ACpqMMEKK(A-F%++b/Q2FO&79C))BC.C MMCA \]`]f]fgigj]k\lm  ! ' 'QQ(8 9 ..9  &/&8" w#(AB<5BQB<#? #''))"[[)9)=)=!)=)DEaHKF,1M,@,@,BHMM..6Z7LZLXY['9==0 h - -"S(O#45H)*yy* ''9.0 == e%WX**7'JEM$MHqQVVAY%:H !$MB*a+|,11!4Cr A6K$L6KR!V6K !$L;= !(,\!_-B-BCX^_-B-`(a"1%"KRS7aT!''*?"@A7MS445GX^XjXjkJ33J@R@RThiJ 99 ! ! # ### bc#' :% %!!&*55/%N %M Ts-P4P 3$P) rrMrrrrrrrrr) i i5NNNNNg?rGi,rGg?iro)rprqrrrsrtrrrrrVrr<unusedrrdr-rmrvrwrxs@r-r"r"CsVr''%--O %C!J YY T>@f6f6r/r")rr) categoriesrc F\rSrSr\"S\0\ESSSSS00SS S S .ES 9r\rS r g)r#izLhttps://download.pytorch.org/models/retinanet_resnet50_fpn_coco-eeacb38b.pthizJhttps://github.com/pytorch/vision/tree/main/references/detection#retinanet COCO-val2017box_mapg333333B@gzGb@gCH`@zSThese weights were produced by following a similar training recipe as on the paper. num_paramsrecipe_metrics_ops _file_size_docsurl transformsmetar*N rprqrrrsrr _COMMON_METACOCO_V1DEFAULTrvr*r/r-r#r#sN Z"    "bt! !n  G"Gr/r#c F\rSrSr\"S\0\ESSSSS00SS S S .ES 9r\rS r g)r$izOhttps://download.pytorch.org/models/retinanet_resnet50_fpn_v2_coco-5905b1c5.pthiFz+https://github.com/pytorch/vision/pull/5756rWrXgD@gV-c@gw/Ab@zZThese weights were produced using an enhanced training recipe to boost the model accuracy.rYr`r*Nrdr*r/r-r$r$sN ]"    "Ct! !u  G"Gr/r$ pretrainedpretrained_backbone)rweights_backboneT)rprogressr^rktrainable_backbone_layersrrlr^rkrmrc [RU5n[R"U5nUb&Sn[SU[ UR S55nOUcSnUSL=(d USLn[ XdSS5nU(a[RO[Rn[X1US9n[X/SQ[S S 5S 9n[X40UD6n Ub?U RUR!US S 95 U[R":Xa [%U S 5 U $)a Constructs a RetinaNet model with a ResNet-50-FPN backbone. .. betastatus:: detection module Reference: `Focal Loss for Dense Object Detection `_. The input to the model is expected to be a list of tensors, each of shape ``[C, H, W]``, one for each image, and should be in ``0-1`` range. Different images can have different sizes. The behavior of the model changes depending on if it is in training or evaluation mode. During training, the model expects both the input tensors and targets (list of dictionary), containing: - boxes (``FloatTensor[N, 4]``): the ground-truth boxes in ``[x1, y1, x2, y2]`` format, with ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``. - labels (``Int64Tensor[N]``): the class label for each ground-truth box The model returns a ``Dict[Tensor]`` during training, containing the classification and regression losses. During inference, the model requires only the input tensors, and returns the post-processed predictions as a ``List[Dict[Tensor]]``, one for each input image. The fields of the ``Dict`` are as follows, where ``N`` is the number of detections: - boxes (``FloatTensor[N, 4]``): the predicted boxes in ``[x1, y1, x2, y2]`` format, with ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``. - labels (``Int64Tensor[N]``): the predicted labels for each detection - scores (``Tensor[N]``): the scores of each detection For more details on the output, you may refer to :ref:`instance_seg_output`. Example:: >>> model = torchvision.models.detection.retinanet_resnet50_fpn(weights=RetinaNet_ResNet50_FPN_Weights.DEFAULT) >>> model.eval() >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)] >>> predictions = model(x) Args: weights (:class:`~torchvision.models.detection.RetinaNet_ResNet50_FPN_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.detection.RetinaNet_ResNet50_FPN_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool): If True, displays a progress bar of the download to stderr. Default is True. num_classes (int, optional): number of output classes of the model (including the background) weights_backbone (:class:`~torchvision.models.ResNet50_Weights`, optional): The pretrained weights for the backbone. trainable_backbone_layers (int, optional): number of trainable (not frozen) layers starting from final block. Valid values are between 0 and 5, with 5 meaning all backbone layers are trainable. If ``None`` is passed (the default) this value is set to 3. **kwargs: parameters passed to the ``torchvision.models.detection.RetinaNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.detection.RetinaNet_ResNet50_FPN_Weights :members: Nr^rU[r )rrlrRrr r1rFreturned_layers extra_blocksTrl check_hashg)r#verifyrrrJrcr rFrozenBatchNorm2dr BatchNorm2drrrr"load_state_dictget_state_dictrfr) rrlr^rkrmr is_trainedrRrmodels r-r%r%sV-33G`_. :func:`~torchvision.models.detection.retinanet_resnet50_fpn` for more details. Args: weights (:class:`~torchvision.models.detection.RetinaNet_ResNet50_FPN_V2_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.detection.RetinaNet_ResNet50_FPN_V2_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool): If True, displays a progress bar of the download to stderr. Default is True. num_classes (int, optional): number of output classes of the model (including the background) weights_backbone (:class:`~torchvision.models.ResNet50_Weights`, optional): The pretrained weights for the backbone. trainable_backbone_layers (int, optional): number of trainable (not frozen) layers starting from final block. Valid values are between 0 and 5, with 5 meaning all backbone layers are trainable. If ``None`` is passed (the default) this value is set to 3. **kwargs: parameters passed to the ``torchvision.models.detection.RetinaNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.detection.RetinaNet_ResNet50_FPN_V2_Weights :members: Nr^rUrorpr )rrlrqirFrrrrCrTgiou)rMrTru)r$rwrrrJrcr rrrrNrPrrrr GroupNormrZrr"rzr{) rrlr^rkrmrr|rrMrr}s r-r&r&=s7Z066w?G'../?@+M;GLLYeLfHgh   $D(8(DJ ::bcef g 0DH$YUbcgilUmH*+ 113A62<<,  D '-D# h d>N d]c dE g44hSW4XY Lr/)Crr> collectionsr functoolsrtypingrrrrrr opsr rr rr ops.feature_pyramid_networkrtransforms._presetsrutilsr_apirrr_metarrrrresnetrrrrr anchor_utilsrbackbone_utilsrr rr!__all__r7r.r<rNrurPrWrYr"rer#r$rf IMAGENET1K_V1boolr?r%r&r*r/r-rsT #** LL82(77$C/!,)M/ DLV>gBIIg