9in&SSKrSSKrSSKJr SSKJs Jr "SS\R5r"SS\R5r "SS\R5r "SS \R5r "S S \R5r "S S \R5r "SS\R5r"SS\R5r"SS\R5rSSjrg)Nc2^\rSrSrSrU4SjrSrSrU=r$)Q2VRankerStage1 z Used to calculate the qv_ctx_score with query embedding and multi anchor context embeddings as input. The qv_ctx_score is used to pre-rank and retain top-k related anchors. cd>[TU]5 [R"X"5UlXlgNsuper__init__nnLinearfcnscalesselfr hidden_dim __class__s k/var/www/html/land-doc-ocr/venv/lib/python3.13/site-packages/modelscope/models/multi_modal/soonet/blocks.pyr Q2VRankerStage1.__init__# ))J3 c URU5n[5n[UR5HVn[R "S[ R"XSSS9[ R"USSS95nURU5 MX U$Nz bld,bd->blpdim) r listrangertorcheinsumF normalizeappend)r ctx_featsqfeat qv_ctx_scoresiscores rforwardQ2VRankerStage1.forwardst t||$ALL!"Y\QA!F!"UaQ!?AE   ' % rr r __name__ __module__ __qualname____firstlineno____doc__r r*__static_attributes__ __classcell__rs@rrr s   rrc2^\rSrSrSrU4SjrSrSrU=r$)V2QRankerStage1 zh Used to calculate the vq_ctx_score with anchor context embeddings and multi query embeddings as input. cd>[TU]5 [R"X"5UlXlgrrrs rr V2QRankerStage1.__init__%rrc [5n[UR5Hen[R"S[ R "URX5SSS9[ R "USSS95nURU5 Mg U$r) rrrr r!r"r#r r$)rr%r& vq_ctx_scoresr(r)s rr*V2QRankerStage1.forward*so t||$ALLakk$''),*?1!L EQA.0E   ' % rr,r-r5s@rr7r7 s rr7c6^\rSrSrSrSU4SjjrSrSrU=r$)Q2VRankerStage25z Used to calculate the qv_ctn_score with query embedding and video sequence embedding as input. The qv_ctn_score is used to re-rank anchors. c>[TU]5 XlX0l[R "X"5Ul[5Ulgr) r r rsnippet_lengthr r qfc V2VAttentionencoder)rrrrBrs rr Q2VRankerStage2.__init__;s2  ,99Z4#~ rc URU5n[5n[5nUR5upxn [5n [UR5GHen UR SU --n X-XKRS5:Xde[ R"XKSX;5n URX-U U 5R5n[ R"USX;5nURU[ R"UR5SSURS95nU RU5 [ R"S[R "UR#S5SSS9[R "USSS95n[ R$"USS9upURU5 URX-5 GMh XeU 4$) Nrrr)device bkld,bd->bklrr)rCrsizerrrBr index_selectviewdetachrEonesrHr$r!r"r# unsqueezemax)rvfeatsr& hit_indicesr' qv_ctn_scoresqv_merge_scores_LD ctn_featsr( anchor_length qv_ctx_scorectn_feat qv_ctn_scores rr*Q2VRankerStage2.forwardBs &++-aF t||$A //!Q$6M%)9)>)>q)AA AA --m.>.9n>L{{1#5}#$&&,fh ))(A{~FH|| 8==?2A.xGIH   X & << H,>,>q,AQA N EQA.0L$ii !
    ?'%*y88r)rErrCrB) r-r5s@rr?r?5s &99rr?c2^\rSrSrSrU4SjrSrSrU=r$)V2QRankerStage2bzh Used to calculate the vq_ctn_score with anchor content embeddings and multi query embeddings as input. cd>[TU]5 [R"X"5UlXlgrrrs rr V2QRankerStage2.__init__grrc ^[5n[UR5Hn[R"S[ R "URX5RS5SSS9[ R "USSS95n[R"USS9nURU5 M U$)NrIrrrJrrrK) rrrr r!r"r#r rQmeanr$)rrZr& vq_ctn_scoresr(r)s rr*V2QRankerStage2.forwardls t||$ALL DGGIL1;;A>!K EQA.0EJJu!,E   ' %rr,r-r5s@rrbrbbs  rrbc2^\rSrSrSrU4SjrSrSrU=r$)rDyzT Self-attention encoder for anchor frame sequence to encode intra-anchor knowledge. c>[TU]5 [SSSS9Ul[ SSSS9Ul[ R"S5Ulg)Nii)max_lenrdropoutg?)rn_headsro) r r PositionEncodingposembMultiHeadAttentionrEr Dropoutro)rrs rr V2VAttention.__init__~s= &sSI )c1cJ zz# rc[R"SUU5RS5nUnXRU5-nUR XXS9nUR UU-5URS5R 5-nU$)Nz bm,bn->bmnr)querykeyvaluemaskr)r r!rQrsrErofloat)r video_feats video_masksr{residualouts rr*V2VAttention.forwards||L+'))21 !KK $<< llkNll8%($)*,7,A,A!,D,J,J,LM r)rorErsr-r5s@rrDrDys'  rrDc6^\rSrSrSrSU4SjjrSrSrU=r$) BboxRegressorz? Predict the offset of bounding box for each candidate anchor. cl>[TU]5 [R"X5Ul[R"X5UlU(a[R"X5Ul[U5Ul[R"[R"SU-U5[R"5[R"US55Ul OY[R"[R"X5[R"5[R"US55Ul X l g)Nr) r r r r fc_ctxfc_qfc_ctn SelfAttentionattn SequentialReLU predictor enable_stage2)rrrrs rr BboxRegressor.__init__s ii 7 IIj5 ))J;DK%j1DI]] !j.*5rwwy *a(*DN ]] *12779 *a(*DN+rcURU5n[R"USS9n[R"UR U55[R"UR S55-nUR(aU(a[5n[[U55Hn[R"URX&5R S55[R"UR S5R S55-nURU5nURU5 M [R"USS9n[R"XE/SS9nOUnURU5nU$)NrrKr)rr catr"relurrQrrrlenrrr$r) rr%rZr&ctx_fuse_featsctn_fuse_featsr(r fuse_featss rr*BboxRegressor.forwards % IIiQ/  I 67!&& OOA ;     )!VN3y>*ffT[[6@@CDqvvOOA&003H55iin%%c* + #YY~1=NN#CLJ'JnnZ( r)rrrrrr)Fr-r5s@rrrs+"rrc2^\rSrSrSrU4SjrSrSrU=r$)rz3 Obtain pooled features by self-attentive pooling. c>[TU]5 [R"XS-5Ul[R "5Ul[R"US-S5Ulg)Nrr)r r r r fc1rrfc2)rrrs rr SelfAttention.__init__sE 99Zq9GGI 99Z1_a0rcURURURU555RS5n[R "USS9R S5n[R"X-SS9nU$)NrJrrK) rrrsqueezer"softmaxrQr sum)rxattrs rr*SelfAttention.forwards^hhtyy!-.66q9ii#--a0iiQ' r)rrrr-r5s@rrrs1 rrc6^\rSrSrSrSU4SjjrSrSrU=r$)rrz An implementation of trainable positional embedding which is added to sequence features to inject time/position information. Args: max_len: The max number of trainable positional embeddings. dim: the dimension of positional embedding. c>[[U] 5 [R"X5Ul[R "5Ul[R"U5Ul gr) r rrr r Embeddingembedrrruro)rrnrrors rr PositionEncoding.__init__s> .0\\'/ GGI zz'* rc$URSSup#[R"U[RURS9nUR S5R US5nURURURU555nU$)Nr)dtyperHrr) shaper arangelongrHrQrepeatrorr)rr batch_sizeseq_lenpos_idspos_embs rr*PositionEncoding.forwardspggbqk ,,wejjJ##A&--j!<,,tyyG)<=>r)rorrrmr-r5s@rrrrrs+rrrc<^\rSrSrSrSU4SjjrSrSrSrU=r $)rtz An implementation of multi-head attention module, as described in 'Attention Is All You Need ' Args: dim: the dimension of features of hidden layers. n_heads: the number of head. c^>[[U] 5 XlX lX-Ul[ R"X5Ul[ R"X5Ul [ R"X5Ul [ R"U5Ul [ R"SS9Ulg)NrrK)r rtr rrqhead_dimr r to_qto_kto_vruroSoftmaxr)rrrqrors rr MultiHeadAttention.__init__st  $02  IIc' IIc' IIc' zz'* zzb) rcUR5SSURUR4-nUR"U6nUR SSSS5$)NrrrrrJ)rLrqrrNpermute)rr new_x_shapes rtranspose_for_scores'MultiHeadAttention.transpose_for_scoressHffhsmt||T]]&CC FFK yyAq!$$rcURU5nURU5nURU5nURU5nURU5n URU5n [R "XR SS55n U [R"UR5- n [X5n URU 5n URU 5n [R "X5n U RSSSS5R5n U R5SSUR 4-n U R""U 6n U $)NrrrrrJ)rrrrr matmul transposemathsqrtr mask_logitsrror contiguousrLrrN)rrxryrzr{qkvq_transk_transv_transrctx_vrs rr*MultiHeadAttention.forwards IIe  IIcN IIe ++A.++A.++A.ll7$5$5b68%:;DIIdmm,,#$ll3ll3 S* aAq)446 Sb!TXXL0 E" r)rrorrqrrrrr) r.r/r0r1r2r rr*r3r4r5s@rrtrts *% rrtcVUR[R5nUSU- U--$)Ng?)typer float32)inputsr{ mask_values rrrs( 99U]] #D S4Z:- --r)gꌠ9Y>))rr torch.nnr torch.nn.functional functionalr"Modulerr7r?rbrDrrrrrtrrrrs bii.bii**9bii*9Zbii.299.*BII*ZBII$ryy411h.r