Zj;*SrSSKJrJrJr SSKJrJr SSKJ r \"5(a SSK r SSK J s J r SSSS SS \S \S \S \\S44 SjjrSS SSSSSS\\S-S \S \S\\\4S-S \S\\\\\-444Sjjr"SS\ 5rS/rg)z"Video processor class for Videomt.)IMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STDPILImageResampling)is_torch_availablerequires_backends)BaseVideoProcessorN mask_labels torch.Tensor mask_probs query_idxmask_thresholdoverlap_mask_area_thresholdreturncX:HnUR5nXU:nUR5nXW-n U R5n US:=(a US:=(a U S:n U (aXh- n U R5U:dSn X4$)aD Checks whether a predicted query produces a valid panoptic segment. Args: mask_labels (`torch.Tensor`): Tensor of shape `(height, width)` containing the winning query index for each pixel. mask_probs (`torch.Tensor`): Tensor of shape `(num_queries, height, width)` containing per-query mask probabilities. query_idx (`int`): Index of the query to validate. mask_threshold (`float`, *optional*, defaults to 0.5): Threshold used to binarize the query mask probabilities. overlap_mask_area_threshold (`float`, *optional*, defaults to 0.8): Minimum overlap ratio required between the assigned query area and the original query mask area. Returns: `tuple[bool, torch.Tensor]`: A tuple containing whether the segment is valid and the final boolean mask for that segment. r F)sumitem) r r r rr query_maskquery_mask_area original_mask original_area final_maskfinal_mask_area mask_exists area_ratios څ/root/GenerationalWealth/GenerationalWealth/venv/lib/python3.13/site-packages/transformers/models/videomt/video_processing_videomt.pycheck_segment_validityrs4)J nn&O)^;M!%%'M+J nn&O!A%S-!*;SRS@SK$4  #>>K  "" pred_scores pred_labelslabel_ids_to_fuse target_sizecpUcURSOUSnUcURSOUSn[R"Xx4[RURS9S- n /n UR 5nUSS2SS4U-R S5n Sn 0n [URS5HnX.R5n[XXU5unnU(dM.U(aX;aX;a XU U'MHXU'XU'[XR5S5nU RU UUS.5 U S- n M X4$)aq Converts per-query mask predictions into a panoptic segmentation map. Args: mask_probs (`torch.Tensor`): Tensor of shape `(num_queries, height, width)` containing per-query mask logits. pred_scores (`torch.Tensor`): Tensor of shape `(num_queries,)` containing the confidence score of each predicted query. pred_labels (`torch.Tensor`): Tensor of shape `(num_queries,)` containing the predicted class ID of each query. label_ids_to_fuse (`set[int]`, *optional*): Label IDs that should be fused across disconnected regions. mask_threshold (`float`, *optional*, defaults to 0.5): Threshold used to binarize the query mask probabilities. overlap_mask_area_threshold (`float`, *optional*, defaults to 0.8): Minimum overlap ratio required to keep a predicted segment. target_size (`tuple[int, int]`, *optional*): Final `(height, width)` of the segmentation map. If unset, uses the spatial size of `mask_probs`. Returns: `tuple[torch.Tensor, list[dict[str, int | float]]]`: The panoptic segmentation map and the metadata for each predicted segment. Nr )dtypedeviceidlabel_idscore) shapetorchzeroslongr'sigmoidargmaxrangerrroundappend)r rr r!rrr"heightwidth segmentationsegmentsr current_segment_idstuff_memory_listr pred_classrr segment_scores rcompute_segmentsr>Gs`@%0$7Z  a [^F#.#6J  Q KNE;;ejjIZIZ[^__LH##%Jq$}- :BB1EK(*;,,Q/0  +002 "8 Y@[#  Z  !@.+<+H Z(0B*-#5Z k499;Q? (&&   a516  !!rc6\rSrSr\R r\r\ r SSS.r Sr Sr SrSrSrSrSrS/rSS S \\\\4S \S 4S jrS \\\\4S \S 4S jrSS \\\\4S\S \\4SjjrSS \\\\4S\S\S\S\\S-S \\4 SjjrSrg)VideomtVideoProcessori)r6r7TFgp?pixel_values_videosmasks_queries_logitsr target_sizesrc/n[U5H5upE[R"XSUSSS9SnURU5 M7 U$)z=Interpolates mask logits to each frame's original resolution.)N.bilinearF)sizemode align_cornersr ) enumerateF interpolater5)selfrCrDresizedidx original_size upsampleds r_resize_mask_logits)VideomtVideoProcessor._resize_mask_logitss\ "+L"9 C $))4"#   I NN9 %#:rc6[US/5 URnURnUR5R SS9SSS24nUR5R 5n[ R"URSS5URS55nURURSURSURS URS5nURXr5nUV s/sHoRSS9PM sn $s sn f) a@ Converts the output of [`VideomtForUniversalSegmentation`] into semantic segmentation predictions. Args: outputs ([`VideomtForUniversalSegmentationOutput`]): Raw outputs of the model. target_sizes (`list[tuple[int, int]]`): List of `(height, width)` tuples corresponding to the requested final size of each prediction. Length should match the number of frames in the output. Returns: `list[torch.Tensor]`: A list of tensors, each of shape `(height, width)`, where each value is the predicted class index for the corresponding pixel. r.dim.Nr$r%r )rrCclass_queries_logitsfloatsoftmaxr1r.matmul transposeflattenreshaper-rRr2) rMoutputsrDrCrY masks_classes masks_probssegmentation_logits output_logitslogits r"post_process_semantic_segmentation8VideomtVideoProcessor.post_process_semantic_segmentations& $ *&;;&;;-224<<<DS#2#XN *002::< #ll=+B+B1a+H+J]J]^_J`a199   a -"5"5b"9;;L;LR;PR]RcRcdfRg 001DS 1>?  #???s;D thresholdc [US/5 URnURnURXR5nURnUR SnUR Sn /n [ U5GHn Xkn XKn U R5RSS9SSS24nURS5unnU S:nU R5R5nURS5URS5-RS5URS5RS5S -- nUU-n[R"X+S[RUS 9n/nSn[ U 5H{nUUR!5n[R""UU5(dM6UU:dM>UUUU'UR%UUUR!5['US 5S .5 US- nM} U R%UUS .5 GM U $)a0 Converts the output of [`VideomtForUniversalSegmentation`] into instance segmentation predictions. Args: outputs ([`VideomtForUniversalSegmentationOutput`]): Raw outputs of the model. target_sizes (`list[tuple[int, int]]`): List of `(height, width)` tuples corresponding to the requested final size of each prediction. Length should match the number of frames in the output. threshold (`float`, *optional*, defaults to 0.5): Minimum combined score to keep an instance. Returns: `list[dict]`: A list of dicts (one per frame), each containing: - `"segmentation"` -- A `torch.Tensor` of shape `(height, width)` with instance IDs (or -1 for background). - `"segments_info"` -- A list of dicts with `"id"`, `"label_id"`, and `"score"` for each instance. r.r rXrUrV.Nr$gư> fill_valuer&r'r(r)r8 segments_info)rrYrCrRr'r-r3rZr[maxr1r^rr.fullr0ranyr5r4)rMr`rDrhrYrCmask_probs_batchr' num_frames num_queriesresults frame_idx mask_pred mask_class class_probsscores pred_classes pred_masksr mask_scoresrr8r9r:r r,s r"post_process_instance_segmentation8VideomtVideoProcessor.post_process_instance_segmentations . $ *&;;&;;334HW%,,)//2 *004 z*I(3I-8J$**,444)rMr`rDrhrrr!rCrYrr num_labelsrqpred_scores_batchpred_labels_batchrtrur rr to_keepr6r7r8r9s r"post_process_panoptic_segmentation8VideomtVideoProcessor.post_process_panoptic_segmentations@ $ *&;;&;;)//2 )//3a7 334HW/C/I/I/K/S/SXZ/S/[/_/_`b/c,z*I)4J+6K+6K$$Q';+<+r@__all__rrrrs,)ZZ:8## ), *#*#*#*# *# "' *#  4  *#d ),*.G"G"G" G"3x$ G"  G" "' G"sCx4'G" >4S#+%5 67 78G"TX.Xv # #r