o *jj@sddlZddlZddlZddlZddlZddlmZddlZddlZ ddl Z ddl Z ddl mZddlmZdddZdd Zd d Zd d ZddZdddZdddZdS)N) OrderedDict)Image) transformsc Csdttj|tjdt|tddgttj|ddttj ddddtddgd}|S) z The implementation of transforms functions. The default image resolution is 224. The normalize parameter follows the mainstream setting. ) interpolation)g3<4'?gwg M?gy{ ?)gB91?gwt.?g U?)g?g?)scaler)Z brightnessZ saturationhue)Z clip_testZ clip_train) rZComposeZResizerZBICUBICZ CenterCropZ NormalizeZRandomResizedCropZRandomHorizontalFlipZ ColorJitter)Z input_resZ tsfm_dictr o/var/www/html/Deteccion_Ine/venv/lib/python3.10/site-packages/modelscope/models/cv/vop_retrieval/basic_utils.pyinit_transform_dicts" r cCsJtj|dd}|dj|d}|dj|d}|d}|d}||||gS)a7 Loading dataset from 'feature_path' as a retrieval docs. The default dataset is MSRVTT-9K. Args: feature_path: 'VoP_msrvtt9k_features.pkl' mydevice: device(type='cuda', index=0) Returns: [text_embeds, vid_embeds_pooled, vid_ids, texts] TZ weights_only text_embeds)ZdeviceZ vid_embedsvid_idstexts)torchloadto)Z feature_pathZmydeviceZfeature_contentr Zvid_embeds_pooledrrr r r load_data,s   rcCs8t|d }t|WdS1swYdS)z Load json files. rN)openjsonr)filenamefr r r load_json@s $rcCsP|dkr&t|tj|tj|t|dtjj_ dtjj_ dSdS)z Set random seed. rTFN) rZ manual_seednprandomseedcudaZmanual_seed_allbackendsZcudnnZ deterministicZ benchmark)rr r r set_seedHs     rcCsFtj|dd}|d}t}|D] \}}|||dd<q|}|S)z0 Load pre-train parameters for VoP. Tr state_dictzmodule.)rrritemsreplace)Zcheckpoint_path checkpointr Znew_state_dictkvr r r get_state_dictUsr'randcCs@t||}tjd||ddt}g}t|ddD]\}}||||ddfq|dkr:dd|D} nd d|D} g} | D]A} |tj | | \} } | sid }t |D] }| \} } | rhnq\| snd St | tj } t| } | d dd} | | qEt| |kr| | dt| |ks| | fS) a Get indexes of sampled frames. Args: cap: cv2.VideoCapture num_frames: int - number of frames to sample vlen: video length, int(cap.get(cv2.CAP_PROP_FRAME_COUNT)), 325 sample: 'rand' | 'uniform' how to sample Returns: frames: torch.tensor of stacked sampled video frames of dim (num_frames, C, H, W) frame_idxs: list(int) indices of where the frames where sampled r)startstopnumNr(cSs$g|]}tt|d|dqS)rr))rchoicerange.0xr r r zs$z$get_valid_frames..cSs g|] }|d|ddqS)rr)r r0r r r r3|s )NNr4)minrZlinspaceZastypeint enumerateappendsetcv2ZCAP_PROP_POS_FRAMESreadr/ZcvtColorZ COLOR_BGR2RGBrZ from_numpyZpermutelenclone)cap num_framesvlensampleZ acc_samplesZ intervalsrangesidxZinterv frame_idxsframesindexretframeZn_tries_r r r get_valid_framesdsB         rKcCs^t|}|s J|t|tj}t||||\}}t| d}| ||fS)a Get indexes of sampled frames. Args: video_path: the local video path num_frames: Frame number, 12 frames for each video sample: 'rand' | 'uniform' how to sample Returns: frames: torch.tensor of stacked sampled video frames of dim (num_frames, C, H, W) frame_idxs: list(int) indices of where the frames where sampled ) r;Z VideoCaptureZisOpenedr7getZCAP_PROP_FRAME_COUNTrKrstackfloatrelease)Z video_pathr@rBr?rArFrEr r r load_frames_from_videos rQ)r)r()ospicklershutilzipfile collectionsrr;numpyrrZujsonrZPILrZ torchvisionrr rrrr'rKrQr r r r s&     2