vjo3ddlZddlZddlZddlZddlmZddlmZddlm Z ddl Z ddl Z ddl Z ddlmcmZddlmcmZddlmZddlmZddlmZddlmZdd lmZmZdd lm Z d d l!m"Z"d d l#m$Z$ ddZ%dZ&dZ' ddZ(ddZ)Gdde*Z+e$j,ej-ej.Gdde"Z/dS)N)exists)TemporaryDirectory)urlparse) VideoReader)Compose) http_get_file) Preprocessors)FieldsModeKeys) type_assert) Preprocessor) PREPROCESSORSct|}|jdvr&t|jrt |||}nt 5}t jj}t|||dtj ||}t |||}dddn #1swxYwY||} t||} n&|j j} t||j j} g} t|dD]X} t| D]F} | jd| | | || GYt)j| dS)aw simple interface to load video frames from file Args: cfg (Config): The global config object. video_path (str): video file path num_spatial_crops_override (int): the spatial crops per clip num_temporal_views_override (int): the temporal clips per video Returns: data (Tensor): the normalized video clips for model inputs )fileN)url local_dir file_namecookiesrr dim)rschemerpath _decode_videoruuiduuid4hexrosjoinkinetics400_tranformTESTNUM_SPATIAL_CROPSrangesize transformsset_spatial_indexappendtorchstack)cfg video_pathnum_spatial_crops_overridenum_temporal_views_override url_parseddatatemporary_cache_dir random_strtemp_file_pathnum_spatial_crops transform data_listijs n/lsinfo/ai/hellotax_ai/data_center/backend/venv/lib/python3.11/site-packages/modelscope/preprocessors/video.py ReadVideoDatar:s*%%JL((V O..(S*.IJJ  ! ! >%8)J -$       W\\*=zJJN n!<>>D > > > > > > > > > > > > > > >"-6(.HII H6(ch.HII I 499Q<< 11()) 1 1A   # 5 5a 8 8 8   YYtAw// 0 0 0 0 1 ;ya ( ( ((s AB66B:=B:ct|jj|jjg|jj|}t j|t j|jj|jjdg}t|S)aI 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'(-sx/BC()+---L  ""L!CHL$ @ @ @ % & &&c||dkrtjd|dz g}n||z|z|z } t|| z d} |dkr| dz } n|tj| |dz z z} |r | | z|z } n| | zdz } t j| | |}t j|d|dz }|S)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:>223!8+g5 B j;.22 >>! II 4:gQ.G#H#HHI  2+-8GG+-1Gy':>> E1j1n55::<< LrKc ttsJ||}n |jj}g}t |D]}t t ||jj|||jj |jj |jj }d}tj tjfd|Dd}||tj|}~ |S)a 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 Nc g|] }| Srf).0r^ frames_lists r9 z-_decode_video_frames_list..sEEEUk%(EEErKr)axis) isinstancelistr"NUM_ENSEMBLE_VIEWSr$rclenrB TARGET_FPSNUM_INPUT_FRAMES SAMPLING_RATEMINUS_INTERVALr) from_numpynpr*tolistr() r+rhrWr.num_clips_per_video frame_listrYlist_framesvrs ` r9_decode_video_frames_listr{s k4 ( ((( (".9!h9J-.."")     H    H % H " H #   ! HEEEEellnnEEEA N N NPP&!!!! [ $ $F MrKc ft|}||}n |jj}g}t|D]h}t t |||jj|||jj |jj |jj }d}| drtjd|dd} tj|tj| |g}|| jdd}nKtj||}||jtj|}~|S)aK Decodes the video given the numpy frames. Args: cfg (Config): The global config object. path (str): video file path. num_temporal_views_override (int): the temporal clips per video Returns: frames (Tensor): video tensor data Nz.avir)rr"rmr$rcrn get_avg_fpsrBrorprqrrendswithr)arangedlpack from_dlpack get_batchcat to_dlpackcloneshaper(r*) r+rr.rzrvrwrYrxry append_lists r9rrs T  B".9!h9J-.."") GG NN   H    H % H " H #    ==  9,q%(A66K' UY (-(/00111:>>>Cegg K-a0112FF' U##--//1116 &!!!! [ $ $F MrKc4eZdZdZ d dZdZdZdZdZdS) rAa@Perform resize and crop for kinetics-400 dataset Args: short_side_range (list): The length of short side range. In inference, this should be [256, 256] crop_size (int): The cropped size for frames. num_spatial_crops (int): The number of the cropped spatial regions in each video. r cXd|_||_t||_||_dS)N)idxr<intr=r4)selfr<r=r4s r9__init__zKineticsResizedCrop.__init__s.  0Y!2rKc"|j\}}}}|jd}||krNt|}t||z |z}tjj|||fd}nMt|}t||z |z}tjj|||fd}t||jz } t||jz } |jdkr | dz} | dz} nn|jdkrc|j dkr||kr| dz} d} nJ||krd} | dz} n<|j dkr | dz} | dz} n&|j dkr||kr| dz} | } n ||kr| } | dz} |dddd| | |jz| | |jzfS)zPerform controlled crop for video tensor. Args: clip (Tensor): the video data, the shape is [T, C, H, W] rbilinearr%moder rMN) rr<rr)nn functional interpolater=r4r) rclip_ clip_height clip_widthlengthnew_clip_heightnew_clip_widthnew_clipx_maxy_maxxys r9_get_controlled_cropz(KineticsResizedCrop._get_controlled_crops )- %1k:&q)  # #!&kkO k!9O!KLLNx*66O^<:7OOHH![[N!+ ":^"KLLOx*66O^<:7OOHNT^344Odn455  !Q & & A AA  #q ( (x1}}!V++ AAA$..A AQQJQJQ!V++ AAA$..A A111aDN 22Aa$.6H4HHIIrKc`|j\}}}}t||}t||}tt j|j}t||z |z}||kr|} |} n|} |} tjj || | fd} t| |j z } t| |j z } tt jd| }tt jd| }| dddd|||j z|||j zfS)Nrrr) rminrPrrNuniformr<r)rrrr=)rrrrr short_side long_sidenew_short_side new_long_siderrrrrrrs r9_get_random_cropz$KineticsResizedCrop._get_random_crops<(, %1k:j11  Z00 V^T-BCDDI 2^CDD  # #,O*NN+O+N8&22 8z3KKNT^344Odn455 q%(( ) ) q%(( ) )111aDN 22Aa$.6H4HHIIrKc||_dS)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'z%KineticsResizedCrop.set_spatial_index6s rKc,||SN)r)rrs r9__call__zKineticsResizedCrop.__call__=s((...rKN)r ) __name__ __module__ __qualname____doc__rrrr'rrfrKr9rArAsz 3 3 3 3*J*J*JXJJJ./////rKrA) module_namecTeZdZfdZdZdZeeedZxZ S)"MovieSceneSegmentationPreprocessorctj|i||dd|_|tjd|_|tjd|_|dd|_ ddl m }||j|_ ||j|_ dS)z7 movie scene segmentation preprocessor is_trainTN num_keyframerr ) get_transform)superrpoprr TRAINpreprocessor_train_cfgEVALpreprocessor_test_cfgrmovie_scene_segmentationrtrain_transformtest_transform)rargskwargsr __class__s r9rz+MovieSceneSegmentationPreprocessor.__init__Es $)&))) :t44 &,jj&F&F#%+ZZ t%D%D""JJ~q99;;;;;;,}T-HII+mD,FGGrKcd|_dS)NTrrs r9trainz(MovieSceneSegmentationPreprocessor.trainTs rKcd|_dS)NFrrs r9evalz'MovieSceneSegmentationPreprocessor.evalXs rKc|jr|j}n|j}tj||d}|d|jddd}|S)Nrrrr)rrrr)r*viewr)rresultsr&s r9rz+MovieSceneSegmentationPreprocessor.__call__\s] = --JJ,J+jj11q999,,r4#4acBBrK) rrrrrrr objectr __classcell__)rs@r9rrAs H H H H H[  ! rKr)NNr)0rQrrNros.pathrtempfiler urllib.parsernumpyrtr)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!rcr{rrrAregister_modulecv%movie_scene_segmentation_preprocessorrrfrKr9rsP ''''''!!!!!! #########=========******666666------66666666444444"""""" .2.2))))))))X''',)))^;?((((V++++\^/^/^/^/^/&^/^/^/B I=NPPP"""""""PP"""rK