9io3SSKrSSKrSSKrSSKrSSKJr SSKJr SSKJ r SSK r SSK r SSK r SSKJs Jr SSKJs Jr SSKJr SSKJr SSKJr SSKJr SS KJrJr SS KJ r S S K!J"r" S S K#J$r$ SSjr%Sr&Sr'SSjr(SSjr)"SS\*5r+\$RX"\RZ\R\S9"SS\"55r/g)N)exists)TemporaryDirectory)urlparse) VideoReader)Compose) http_get_file) Preprocessors)FieldsModeKeys) type_assert) Preprocessor) PREPROCESSORSc[U5nURS;a'[UR5(a [ XU5nOi[ 5n[ R"5Rn[UUUSS9 [RRXg5n[ XU5nSSS5 UbUn [X5n O5URRn [XRR5n /n [WR!S55HLn [U 5H:n U R"SR%U 5 U R'U "X\55 M< MN [(R*"U SS9$!,(df  N=f)aVsimple interface to load video frames from file Args: cfg (Config): The global config object. video_path (str): video file path num_spatial_crops_override (int): the spatial crops per clip num_temporal_views_override (int): the temporal clips per video Returns: data (Tensor): the normalized video clips for model inputs )fileN)url local_dir file_namecookiesrr dim)rschemerpath _decode_videoruuiduuid4hexrosjoinkinetics400_tranformTESTNUM_SPATIAL_CROPSrangesize transformsset_spatial_indexappendtorchstack)cfg video_pathnum_spatial_crops_overridenum_temporal_views_override url_parseddatatemporary_cache_dir random_strtemp_file_pathnum_spatial_crops transform data_listijs ^/var/www/html/land-doc-ocr/venv/lib/python3.13/site-packages/modelscope/preprocessors/video.py ReadVideoDatar:sE*%JL(V OO..S.IJ  !%8))J -$    WW\\*=JN !<>D""-6(I HH66(hh.H.HI I 499Q< ()A   # 5 5a 8   Ytw/ 0*! ;;ya ((-" !s AE33 FcR[URRURR/URRUS9n[R "5U[R "URRURRSS9/n[U5$)a) Configs the transform for the kinetics-400 dataset. We apply controlled spatial cropping and normalization. Args: cfg (Config): The global config object. num_spatial_crops (int): the spatial crops per clip Returns: transform_function (Compose): the transform function for input clips )short_side_range crop_sizer4T)meanstdinplace) KineticsResizedCropDATA TEST_SCALETEST_CROP_SIZEr& ToTensorVideoNormalizeVideoMEANSTDr)r+r4 resize_videostd_transform_lists r9r!r!Es'((--sxx/B/BC(())+-L   "L!!CHHLL$ @ % &&c|US:Xa[R"SUS- 5/nU$XV-U-U- n [X - S5n US:XaU S- n OU[R"XS- - 5-n U(aX-U- n OX-S- n [ R "XU5n[ R"USUS- 5R5nU$)a Generates the frame index list using interval based sampling. Args: vid_length (int): the length of the whole video (valid selection range). vid_fps (int): the original video fps target_fps (int): the normalized video fps clip_idx (int): -1 for random temporal sampling, and positive values for sampling specific clip from the video num_clips (int): the total clips to be sampled from each video. combined with clip_idx, the sampled video is the "clip_idx-th" video from "num_clips" videos. num_frames (int): number of frames in each sampled clips. interval (int): the interval to sample each frame. minus_interval (bool): control the end index Returns: index (tensor): the sampled frame indexes r r) randomrandintmaxmathfloorr)linspaceclamplong) vid_lengthvid_fps target_fpsclip_idx num_clips num_framesintervalminus_intervalindex clip_lengthmax_idx start_idxend_idxs r9_interval_based_samplingrc[s,Q:>23$ L!+g5 B j.2 >! I 4::gQ.G#HHI -8G-1Gy:> E1j1n5::< LrKc \[U[5(deUbUnOURRn/n[ U5Hn[ [ U5UURRUUURRURRURR5nSn[R"[R"UR!5V s/sHoU PM sn SS95nUR#U5 M [R"U5nA U$s sn f)ai Decodes the video given the numpy frames. Args: cfg (Config): The global config object. frames_list (list): all frames for a video, the frames should be numpy array. vid_fps (int): the fps of this video. num_temporal_views_override (int): the temporal clips per video Returns: frames (Tensor): video tensor data Nr)axis) isinstancelistr"NUM_ENSEMBLE_VIEWSr$rclenrB TARGET_FPSNUM_INPUT_FRAMES SAMPLING_RATEMINUS_INTERVALr) from_numpynpr*tolistr() r+ frames_listrWr.num_clips_per_video frame_listrYlist_framesr^vrs r9_decode_video_frames_listrws k4 ( (( (".9!hh99J-.)    HH     HH % % HH " " HH # #  !! HHellnEnU%(nEA NP&!!/"[[ $F M Fs#D)c |[U5nUbUnOURRn/n[U5GHkn[ [ U5UR 5URRUUURRURRURR5nSnURS5(a[R"SUSS5n [R "UR#[R$"U U/55R'55R)5nXR*SSnOA[R "UR#U5R'55R)5nUR-U5 GMn [R."U5nAU$)a Decodes the video given the numpy frames. Args: cfg (Config): The global config object. path (str): video file path. num_temporal_views_override (int): the temporal clips per video Returns: frames (Tensor): video tensor data Nz.avir)rr"rhr$rcri get_avg_fpsrBrjrkrlrmendswithr)arangedlpack from_dlpack get_batchcat to_dlpackcloneshaper(r*) r+rr.rvrrrsrYrtru append_lists r9rrsq T B".9!hh99J-.) G NN  HH     HH % % HH " " HH # #   == ,,q%(A6K'' UYY (-(/011:>>Ceg --a012F'' U#--/116 &!//0[[ $F MrKc<\rSrSrSrS SjrSrSrSrSr Sr g ) rAa,Perform resize and crop for kinetics-400 dataset Args: short_side_range (list): The length of short side range. In inference, this should be [256, 256] crop_size (int): The cropped size for frames. num_spatial_crops (int): The number of the cropped spatial regions in each video. cJSUlXl[U5UlX0lg)N)idxr<intr=r4)selfr<r=r4s r9__init__KineticsResizedCrop.__init__s"  0Y!2rKc@URu p#nURSnX4:aF[U5n[XC- U-5n[RR R XU4SS9nOE[U5n[X4- U-5n[RR R XU4SS9n[XpR- 5n [X`R- 5n URS:Xa U S-n U S-n O~URS:XanURS:XaXu:XaU S-n Sn OQXe:XaSn U S-n ODURS:Xa U S-n U S-n O)URS:XaXu:XaU S-n U n O Xe:XaU n U S-n USS2SS2W XR-2W XR-24$)zmPerform controlled crop for video tensor. Args: clip (Tensor): the video data, the shape is [T, C, H, W] rbilinearr%moder rMN) rr<rr)nn functional interpolater=r4r) rclip_ clip_height clip_widthlengthnew_clip_heightnew_clip_widthnew_clipx_maxy_maxxys r9_get_controlled_crop(KineticsResizedCrop._get_controlled_crops )- %1:&&q)  #!&kO !9O!KLNxx**66^<:7OH![N!+":^"KLOxx**66^<:7OHN^^34Onn45  ! !Q & A A  # #q (xx1}!+ AA$.A AQQJQJQ!+ AA$.A A1aNN 22Aa..6H4HHIIrKcXURu p#n[X45n[X45n[[R "UR 65n[Xe- U-5nX4:aUn Un OUn Un [RRRXU 4SS9n [XR- 5n [XR- 5n [[R "SU 55n[[R "SU 55nU SS2SS2XUR-2XUR-24$)Nrrr) rminrPrrNuniformr<r)rrrr=)rrrrr short_side long_sidenew_short_side new_long_siderrrrrrrs r9_get_random_crop$KineticsResizedCrop._get_random_crops(, %1:1  0 V^^T-B-BCDI2^CD  #,O*N+O+N88&&22 8z3KN^^34Onn45 q%( ) q%( )1aDNN 22A$..6H4HHIIrKcXlg)zSet the spatial cropping index for controlled cropping.. Args: idx (int): the spatial index. The value should be in [0, 1, 2], means [left, center, right], respectively. N)r)rrs r9r'%KineticsResizedCrop.set_spatial_index6s rKc$URU5$N)r)rrs r9__call__KineticsResizedCrop.__call__=s((..rK)r=rr4r<N)r ) __name__ __module__ __qualname____firstlineno____doc__rrrr'r__static_attributes__rKr9rArAs( 3*JXJ./rKrA) module_namecR^\rSrSrU4SjrSrSr\"\\5S5r Sr U=r $)"MovieSceneSegmentationPreprocessoriAc~>[TU]"U0UD6 URSS5UlUR[R S5UlUR[RS5UlURSS5Ul SSK J n U"UR 5Ul U"UR5Ul g)z' movie scene segmentation preprocessor is_trainTN num_keyframerr ) get_transform)superrpoprr TRAINpreprocessor_train_cfgEVALpreprocessor_test_cfgrmovie_scene_segmentationrtrain_transformtest_transform)rargskwargsr __class__s r9r+MovieSceneSegmentationPreprocessor.__init__Es $)&) :t4 &,jj&F#%+ZZ t%D""JJ~q9;,T-H-HI+D,F,FGrKcSUlg)NTrrs r9train(MovieSceneSegmentationPreprocessor.trainTs  rKcSUlg)NFrrs r9eval'MovieSceneSegmentationPreprocessor.evalXs  rKcUR(a URnO URn[R"U"U5SS9nUR SUR SSS5nU$)Nrrrrr)rrrr)r*viewr)rresultsr&s r9r+MovieSceneSegmentationPreprocessor.__call__\sU ==--J,,J++j1q9,,r4#4#4acBrK)rrrrrr) rrrrrrrr objectrr __classcell__)rs@r9rrAs. H !rKr)NNr)0rQrrNros.pathrtempfiler urllib.parsernumpyror)torch.utils.datatorch.utils.dlpackutilsr}(torchvision.transforms._transforms_videor&_transforms_videodecordrtorchvision.transformsrmodelscope.hub.file_downloadrmodelscope.metainfor modelscope.utils.constantr r modelscope.utils.type_assertr baserbuilderrr:r!rcrwrrrAregister_modulecv%movie_scene_segmentation_preprocessorrrrKr9rs '! ##==*6-64" .2.2))X',)^;?(V+\^/&^/B II=NNP""P"rK