+ jy@Rt^RIHt^RIHtHtHt^RIHtH t ^RI H t ]!4'd^RI t ^RI HuHtR RRlltR RR llt!R R ] 4tR .tR#)z"Video processor class for Videomt.)'SemanticSegmentationPostProcessorOutput)IMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STDPILImageResampling)is_torch_availablerequires_backends)BaseVideoProcessorNcjV^8dQhRRRRR\R\R\R\\R3,/#) mask_labels torch.Tensor mask_probs query_idxmask_thresholdoverlap_mask_area_thresholdreturn)intfloattuplebool)formats"u/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/models/videomt/video_processing_videomt.py __annotate__rsN*#*#*#*#*# *# "' *#  4  *#c"W8HpVP4pW,V8pVP4pWW,p V P4p V^8;'dV^8;'dV ^8p V 'd Wh, p V P4V8gRp W3#)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. F)sumitem) r r rrr query_maskquery_mask_area original_mask original_area final_maskfinal_mask_area mask_exists area_ratios &&&&& rcheck_segment_validityr%s4)J nn&O)^;M!%%'M+J nn&O!A%SS-!*;SSRS@SK$4  #>>K  ""rcV^8dQhRRRRRRR\\,R,R\R\R \\\3,R,R \R\\ \ \\,3,,3,/#) r r r pred_scores pred_labelslabel_ids_to_fuseNrr target_sizer)setrrrlistdictstr)rs"rrrHsG"G"G"G" G"3x$ G"  G" "' G"sCx4'G" >4S#+%5 67 78G"rc VfVP^,MV^,pVfVP^,MV^,p\P!Wx3\PVPR7^, p .p VP 4pVR,V,P ^4p ^p /p \VP^,4FpW.,P4p\WWV4wppV'gK3V'dW9dW9dW,V V&KTWV&WV&\W,P4^4pV PRV RVRV/4V ^, p K W3#)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. )dtypedeviceidlabel_idscore):NNNNN) shapetorchzeroslongr1sigmoidargmaxrangerr%roundappend)r r'r(r)rrr*heightwidth segmentationsegmentsr current_segment_idstuff_memory_listr pred_classr#r! segment_scores&&&&&&& rcompute_segmentsrFHsa@%0$7Z  a [^F#.#6J  Q KNE;;ejjIZIZ[^__LH##%J}- :BB1EK(*;,,Q/0  +002 "8 Y@[#  Z  !@.+<+H Z(0B*-#5Z k499;Q? (J   a516  !!rca]tRt^to]P t]t] t RRRR/t Rt Rt RtRtRtRtRtR.tV3RlRltRV3R lR lltRV3R lR lltRV3RlRlltRtVtR #)VideomtVideoProcessorr>ir?TFpixel_values_videosc^<V^8dQhRRRS[S[S[S[3,,RS[R,/#)r masks_queries_logitsr target_sizesr)r,rr)r __classdict__s"rr"VideomtVideoProcessor.__annotate__s;,5c?+ n  rc .p\V4FDwrE\P!W,R,VRRR7^,pVPV4KF V#)z=Interpolates mask logits to each frame's original resolution.bilinearF)sizemode align_corners)N.) enumerateF interpolater=)selfrKrLresizedidx original_size upsampleds&&& r_resize_mask_logits)VideomtVideoProcessor._resize_mask_logitss\ "+L"9 C $))4"#   I NN9 %#:rcP<V^8dQhRS[S[S[S[3,,RS[RR/#)r rLreturn_segmentation_scoresrzBlist[torch.Tensor] | list[SemanticSegmentationPostProcessorOutput])r,rrr)rrMs"rrrNs96%6%5c?+6%%) 6% N 6%rc  \VR.4VPpVPpVP4P RR7RRR13,pVP4P 4p\ P!VP^^4VP^44pVPVP^,VPR,VPR ,VPR,4pVPW4p V U u.uF#p \RV P^R7RV /R7NK% p p V'gV U u.uFqPNK p p V #uup iuup i) aQ 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. return_segmentation_scores (`bool`, *optional*, defaults to `False`): Whether to return segmentation scores alongside the segmentation map. When `True`, each element of the returned list is a [`SemanticSegmentationPostProcessorOutput`] with fields `segmentation` (class IDs, shape `(height, width)`) and `segmentation_scores` (shape `(num_classes, height, width)`). Returns: `list[torch.Tensor]` or `list[SemanticSegmentationPostProcessorOutput]`: When `return_segmentation_scores=False` (default), a list of length `batch_size` where each item is a segmentation map of shape `(height, width)` with class IDs. When `return_segmentation_scores=True`, a list of [`SemanticSegmentationPostProcessorOutput`] with fields `segmentation` (class IDs, shape `(height, width)`) and `segmentation_scores` (shape `(num_classes, height, width)`). In both cases, `(height, width)` corresponds to the target size. r6dim.Nr@segmentation_scores)data)rrKclass_queries_logitsrsoftmaxr9r6matmul transposeflattenreshaper5r\rr:r@) rWoutputsrLr_rKrg masks_classes masks_probssegmentation_logits output_logitslogitsemantic_segmentationrs &&&& r"post_process_semantic_segmentation8VideomtVideoProcessor.post_process_semantic_segmentationsd8 $ *&;;&;;-224<<<DS#2#XN *002::< #ll=+B+B1a+H+J]J]^_J`a199   a -"5"5b"9;;L;LR;PR]RcRcdfRg 001DS ' ! ' 4$ellql&9;PRWX ' ! *CX$YCX4%6%6CX !$Y$$! %Zs )E! E&cb<V^8dQhRS[S[S[S[3,,RS[RS[S[,/#)r rL thresholdr)r,rrrr-)rrMs"rrrNsAEE5c?+E E d Erc  \VR.4VPpVPpVPWR4pVPpVP ^,pVP R ,p .p \ V4EFp Wk,p WK,p V P4PR R7RRR 13,pVPR 4wppV ^8pV P4P4pVP^4VP^4,P^4VP^4P^4R,, pVV,p\P!W+,R \PVR7p.p^p\ V 4FpVV,P!4p\P"!VV,4'gK?VV8gKHVVVV,&VP%RVRVV,P!4R \'V^4/4V^, pK V P%R VR V/4EK V #)a0 Converts the output of [`VideomtForUniversalSegmentation`] into instance segmentation predictions. Args: outputs ([`VideomtForUniversalSegmentationOutput`]): Raw outputs of the model. target_sizes (`list[tuple[int, int]]`): List of `(height, width)` tuples corresponding to the requested final size of each prediction. Length should match the number of frames in the output. threshold (`float`, *optional*, defaults to 0.5): Minimum combined score to keep an instance. Returns: `list[dict]`: A list of dicts (one per frame), each containing: - `"segmentation"` -- A `torch.Tensor` of shape `(height, width)` with instance IDs (or -1 for background). - `"segments_info"` -- A list of dicts with `"id"`, `"label_id"`, and `"score"` for each instance. r6ra.Ngư> fill_valuer0r1r2r3r4r@ segments_inforfre)rrgrKr\r1r5r;rrhmaxr9rkrr6fullr8ranyr=r<)rWrmrLrwrgrKmask_probs_batchr1 num_frames num_queriesresults frame_idx mask_pred mask_class class_probsscores pred_classes pred_masksr mask_scoresr'r@rArBrr4s&&&& r"post_process_instance_segmentation8VideomtVideoProcessor.post_process_instance_segmentations. $ *&;;&;;334HW%,,)//2 *004 z*I(3I-8J$**,444r?r@rAs&&&&&&& r"post_process_panoptic_segmentation8VideomtVideoProcessor.post_process_panoptic_segmentation0s@ $ *&;;&;;)//2 )//3a7 334HW/C/I/I/K/S/SXZ/S/[/_/_`b/c,z*I)4J+6K+6K$$Q';+<+rsT)OZZ:8##*#ZG"Tk.k\ # #r