o *Îjn&ã@sÞddlZddlZddlmZddlmmZGdd„dejƒZGdd„dejƒZ Gdd„dejƒZ Gdd „d ejƒZ Gd d „d ejƒZ Gd d „d ejƒZ Gdd„dejƒZGdd„dejƒZGdd„dejƒZddd„ZdS)éNcó(eZdZdZ‡fdd„Zdd„Z‡ZS)ÚQ2VRankerStage1z 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. có"tƒ ¡t ||¡|_||_dS©N©ÚsuperÚ__init__ÚnnÚLinearÚfcÚnscales©Úselfr Ú hidden_dim©Ú __class__©úl/var/www/html/Deteccion_Ine/venv/lib/python3.10/site-packages/modelscope/models/multi_modal/soonet/blocks.pyró  zQ2VRankerStage1.__init__c CsX| |¡}tƒ}t|jƒD]}t dtj||dddtj|ddd¡}| |¡q |S©Nz bld,bd->blé©ÚpÚdimé) r ÚlistÚranger ÚtorchÚeinsumÚFÚ normalizeÚappend)rÚ ctx_featsÚqfeatÚ qv_ctx_scoresÚiÚscorerrrÚforwards þ zQ2VRankerStage1.forward©Ú__name__Ú __module__Ú __qualname__Ú__doc__rr'Ú __classcell__rrrrr s rcr)ÚV2QRankerStage1zt Used to calculate the vq_ctx_score with anchor context embeddings and multi query embeddings as input. crrrr rrrr%rzV2QRankerStage1.__init__c CsTtƒ}t|jƒD]}t dtj| ||¡dddtj|ddd¡}| |¡q|Sr) rrr rrrr r r!)rr"r#Z vq_ctx_scoresr%r&rrrr'*sþ zV2QRankerStage1.forwardr(rrrrr. ó r.có*eZdZdZd‡fdd„ Zdd„Z‡ZS)ÚQ2VRankerStage2z¡ 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. é cs0tƒ ¡||_||_t ||¡|_tƒ|_dSr) rrr Úsnippet_lengthr r ÚqfcÚ V2VAttentionÚencoder)rr rr3rrrr;s  zQ2VRankerStage2.__init__c Cs.| |¡}tƒ}tƒ}| ¡\}}} tƒ} t|jƒD]w} |jd| } || ||  d¡ks0J‚t || d|| ¡} | || | | ¡  ¡}t |d|| ¡}|  |tj | ¡dd…|j d¡}|   |¡t dtj| d¡dddtj|ddd¡}tj|dd\}}|  |¡|  | |¡q||| fS) Nrrr)Údeviceú bkld,bd->bklér©r)r4rÚsizerr r3rZ index_selectÚviewÚdetachr6Zonesr7r!rrr Ú unsqueezeÚmax)rZvfeatsr#Z hit_indicesr$Z qv_ctn_scoresZqv_merge_scoresÚ_ÚLÚDÚ ctn_featsr%Z anchor_lengthZ qv_ctx_scoreZctn_featZ qv_ctn_scorerrrr'Bs<  ÿ ÿÿþ þ  zQ2VRankerStage2.forward)r2r(rrrrr15sr1cr)ÚV2QRankerStage2zt Used to calculate the vq_ctn_score with anchor content embeddings and multi query embeddings as input. crrrr rrrrgrzV2QRankerStage2.__init__c Cshtƒ}t|jƒD])}t dtj| ||¡ d¡dddtj|ddd¡}tj |dd}|  |¡q|S)Nr8rrr9rrr:) rrr rrrr r r>Úmeanr!)rrCr#Z vq_ctn_scoresr%r&rrrr'lsý zV2QRankerStage2.forwardr(rrrrrDbr/rDcr)r5z` Self-attention encoder for anchor frame sequence to encode intra-anchor knowledge. cs:tƒ ¡tdddd|_tdddd|_t d¡|_dS)Niiç)Úmax_lenrÚdropoutégš™™™™™¹?)rÚn_headsrH) rrÚPositionEncodingÚposembÚMultiHeadAttentionr6r ÚDropoutrH)rrrrr~s zV2VAttention.__init__cCsXt d||¡ d¡}|}|| |¡}|j||||d}| ||¡| d¡ ¡}|S)Nz bm,bn->bmnr)ÚqueryÚkeyÚvalueÚmaskr)rrr>rLr6rHÚfloat)rZ video_featsZ video_masksrRZresidualÚoutrrrr'„s ÿÿÿÿ ÿzV2VAttention.forwardr(rrrrr5yó r5cr0)Ú BboxRegressorzK Predict the offset of bounding box for each candidate anchor. Fcsštƒ ¡t ||¡|_t ||¡|_|r6t ||¡|_t|ƒ|_t  t d||¡t  ¡t |d¡¡|_ nt  t ||¡t  ¡t |d¡¡|_ ||_ dS)Nr) rrr r Úfc_ctxÚfc_qÚfc_ctnÚ SelfAttentionÚattnZ SequentialÚReLUÚ predictorÚ enable_stage2)rrr^rrrr•s   þ þ zBboxRegressor.__init__c CsÒ| |¡}tj|dd}t | |¡¡t | d¡¡}|jr`|r`tƒ}t t |ƒƒD]$}t |  ||¡ d¡¡t | d¡ d¡¡}|  |¡}|  |¡q*tj|dd}tj||gdd}n|}| |¡}|S)Nrr:réÿÿÿÿ)rXrÚcatrÚrelurWr>r^rrÚlenrYr[r!r]) rr"rCr#Zctx_fuse_featsZctn_fuse_featsr%rTZ fuse_featsrrrr'¦s$ ÿ ÿ   zBboxRegressor.forward)Fr(rrrrrVsrVcr)rZz? Obtain pooled features by self-attentive pooling. cs<tƒ ¡t ||d¡|_t ¡|_t |dd¡|_dS)Nrr)rrr r Úfc1r\raÚfc2)rrrrrrÂs  zSelfAttention.__init__cCsF| | | |¡¡¡ d¡}tj|dd d¡}tj||dd}|S)Nr9rr:) rdrarcZsqueezerÚsoftmaxr>rÚsum)rÚxÚattrTrrrr'ÈszSelfAttention.forwardr(rrrrrZ½rUrZcr0)rKa! 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. rFcs6tt|ƒ ¡t ||¡|_t ¡|_t |¡|_ dSr) rrKrr Z EmbeddingÚembedr\rarNrH)rrGrrHrrrrÙs zPositionEncoding.__init__cCsR|jdd…\}}tj|tj|jd}| d¡ |d¡}| | |  |¡¡¡}|S)Nr)Zdtyper7rr) ÚshaperZarangeÚlongr7r>ÚrepeatrHrari)rrgZ batch_sizeZseq_lenZpos_idsZpos_embrrrr'às zPositionEncoding.forward©rFr(rrrrrKÏs rKcs2eZdZdZd ‡fdd„ Zdd„Zdd„Z‡ZS) rMa  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. rFcsltt|ƒ ¡||_||_|||_t ||¡|_t ||¡|_ t ||¡|_ t  |¡|_ tj dd|_dS)Nr_r:)rrMrrrJÚhead_dimr r Úto_qÚto_kÚto_vrNrHZSoftmaxre)rrrJrHrrrrós  zMultiHeadAttention.__init__cCs6| ¡dd…|j|jf}|j|Ž}| dddd¡S)Nr_rrrr9)r;rJrnr<Úpermute)rrgZ new_x_shaperrrÚtranspose_for_scoress z'MultiHeadAttention.transpose_for_scorescCsÄ| |¡}| |¡}| |¡}| |¡}| |¡} | |¡} t ||  dd¡¡} | t |j ¡} t | |ƒ} |  | ¡} |  | ¡} t | | ¡} |   dddd¡ ¡} |  ¡dd…|jf} | j| Ž} | S)Nr_éþÿÿÿrrrr9)rorprqrsrÚmatmulZ transposeÚmathÚsqrtrnÚ mask_logitsrerHrrÚ contiguousr;rr<)rrOrPrQrRÚqÚkÚvZq_transZk_transZv_transrhZctx_vrjrrrr's$       ÿ     zMultiHeadAttention.forwardrm)r)r*r+r,rrsr'r-rrrrrMés  rMçꌠ9Y>)ÆcCs| tj¡}|d||S)Ngð?)ÚtyperZfloat32)ZinputsrRZ mask_valuerrrrxs rx)r})rvrZtorch.nnr Ztorch.nn.functionalZ functionalrÚModulerr.r1rDr5rVrZrKrMrxrrrrÚs --4