^j.dZddlmZmZmZmZddlZddlZddl m Z ddl m cm Z Gdde jZdeeeeege jffdedee jfd Zdd ZGd d e jZGd de jZGdde jZGdde jZGdde jZy)z Projector.)CallableOptionalSequenceUnionNc*eZdZdZdfd ZdZxZS) LayerNormaHA LayerNorm variant, popularized by Transformers, that performs point-wise mean and variance normalization over the channel dimension for inputs that have shape (batch_size, channels, height, width). https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119 ct|tjt j ||_tjt j||_||_ |f|_ yN) super__init__nn Parametertorchonesweightzerosbiasepsnormalized_shape)selfrr __class__s k/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/models/backbone/projector.pyr zLayerNorm.__init__sT ll5::.>#?@ LL-=!>? !1 3c|jdddd}tj||j|j|j |j }|jdddd}|S)zb LayerNorm forward TODO: this is a hack to avoid overflow when using fp16 r)permuteF layer_normrrrrrxs rforwardzLayerNorm.forward$sY IIaAq ! LLD114;; 488 T IIaAq !r)gư>__name__ __module__ __qualname____doc__r r# __classcell__rs@rrrs 4rrnorm out_channelsreturncf|yt|trt|dk(ryddi|}||S)z Args: norm: Either one of BN, SyncBN, FrozenBN, GN; or a callable that takes a channel number and returns the normalization layer as a nn.Module. Returns: The normalization layer. NrLNct|Sr )r)channelss rzget_norm..>s 9X#6r) isinstancestrlen)r+r,s rget_normr6/sF |$ t9> 6   rc|dk(rtj|}|S|dk(rtj|}|S|dvrtjd|}|S|tj}|St dj |)zGet activation.siluinplacerelu) LeakyReLU leakyrelulrelug?zUnsupported act type: {})r SiLUReLUr<IdentityAttributeErrorformat)namer:modules rget_activationrFCs v~) M ) M 4 4c73 M  M7>>tDEErc8eZdZdZ dfd ZdZxZS)ConvXzConv-bn module.c dtt| t|ts||f}|ddz|ddzf} t j ||||| ||d|_| rt j||_ n(|r td|nt j||_ t|d|_ y) NrrrF) kernel_sizestridepaddinggroupsdilationrr/Tr9)r rHr r3tupler Conv2dconvRMSNormbnr6 BatchNorm2drFact) r in_planes out_planeskernelrKrMrNrUr rms_normrLrs rr zConvX.__init__Us eT#%&%(f%F!9>6!9>2II      jj,DG4>htZ0BNNS]D^DG!#t4rc|j|j|j|j}|S)forward.)rUrSrQ contiguousrr"outs rr#z ConvX.forwardus-hhtwwtyy89: r)rrrrr;FFr$r*s@rrHrHRs(  5@rrHc*eZdZdZdfd ZdZxZS) BottleneckzStandard bottleneck.c t |t||z} t|| |dd||| |_t| ||dd|||| |_|xr||k(|_y)z1ch_in, ch_out, shortcut, groups, kernels, expand.rrrUr rY)rMrUr rYN)r r intrHcv1cv2add) rc1c2shortcutgkerUr rYc_rs rr zBottleneck.__init__~sg  a[R1qcjS[\R1q ]ef(brc|jr#||j|j|zS|j|j|S)z1'forward()' applies the YOLOv5 FPN to input data.)rfrerdr!s rr#zBottleneck.forwards:,0HHq488DHHQK((O$((488A;:OOr)Trrr?r8FFr$r*s@rr`r`{s)Prr`c*eZdZdZdfd ZdZxZS)C2fzzC2f.__init__..s;  tvvtvvxfYcnv w w s25N) r r rcrvrHrdrer ModuleListrangem) rrgrhnrirjrlrUr rYrs ` `` ```rr z C2f.__init__s R!VQZA3:X`a Udff b!h   1X  rct|j|j|j|jfdj fd|j D|j tjdS)z.Forward pass using split() instead of chunk().rc34K|]}|dyw)N)rwr|ys rryzC2f.forward..s*a1R5*s) listrdsplitrvextendr|rercat)rr"rs @rr#z C2f.forwards^ !""DFFDFF#3Q7 8 *466**xx !Q((r)rFrrpr8FFr$r*s@rrrrrsF  )rrrcfeZdZdZ ddeededeededededed ed d ffd Zd Z xZ S)MultiScaleProjectorzThis module implements MultiScaleProjector in :paper:`lwdetr`. It creates pyramid features built on top of the input feature map. in_channelsr, scale_factors num_blocksr rY survival_probforce_drop_last_n_featuresr-Nc B tt| ||_||_||_g} g} d|_|D] | jg|D]!} g} dk(rl| jtj| | dzddtd| dztjtj| dz| dzddgn dk(r-| jtj| | dzddgnS dk(rnM d k(r!| jt| | d d| gn' d k(rd |_tdj tj | } | dj| $tj"| d| d<t%t' fd|D} t)| ||| td|g} tj | } | j| tj"| |_tj"| |_y)a2 Args: in_channels: Channels in each input feature map level. out_channels: Number of channels in the output feature maps. scale_factors: List of scaling factors to upsample or downsample the input features for creating pyramid features. Fg@r)rJrKr/g@rurpr)r g?TzUnsupported scale_factor:{}rc3<K|]}|tdzyw)rN)max)rw in_channelscales rryz/MultiScaleProjector.__init__..sWZZ3q%=8WsN)r rr rrruse_extra_poolappendrr ConvTranspose2dr6GELUrHNotImplementedErrorrC Sequentialrzrcsumrrstages_samplingstages)rrr,rrr rYrrrrin_dimlayersrrs @rr zMultiScaleProjector.__init__s$ !413***D'#"9 "E  " "2 &%. 3 C<MM..vv{PQZ[\$T6Q;7GGI..v{FaKUV_`a c\MM..vv{PQZ[\ c\c\MM!&&!Q:N d]*.D'-.K.R.RSX.YZZ/#**62]. 3^#%--0C"DOB W;WWXFFL*L|,F]]F+F MM& !s9 "v "}}_=mmF+ rcBt|}|jdkrd|jrXd|jz }tjj }t d|D]}|||dz z z}||ksd||ddnL|jdkDr=t |jD]%}tj||dz ||dz <'g}t|jD]\}}g} t|j|D]\} } | j| || !t| dkDrtj| d} n| d} |j|| |jr+|jt!j"|dddd|S) a Args: x: Tensor of shape (N,C,H,W). H, W must be a multiple of ``self.size_divisibility``. Returns: dict[str->Tensor]: mapping from feature map name to pyramid feature map tensor in high to low resolution order. Returned feature names follow the FPN convention: "p", where stage has stride = 2 ** stage e.g., ["p2", "p3", ..., "p6"]. rurrN)dimrr)rJrKrL)r5rtrainingnprandomuniformr{rr zeros_like enumeraterrrrrr max_pool2d) rr" num_featuresfinal_drop_probdrop_picritical_drop_probresultsstage feat_fusejstage_samplings rr#zMultiScaleProjector.forwards1v    # $"4"44OYY&&(F1l+ %&/\A=M*N%O"..AaDG  , ,q 04::; <#..q1q5{;AE(  <!$++. -HAuI%.t/C/CA/F%G 7!>  !!56 79~!!IIiQ7 %aL NN5+ , -    NN1<< 1VWX Yr)rFFrur) r%r&r'r(rrcfloatboolr r#r)r*s@rrrs "*+X,c]X,X,  X,  X,  X,X,X,%(X, X,t%rrc&eZdZdfd ZdZxZS)SimpleProjectorctt| |s/t||dzdd|_t|dz|dd|_n*t||ddd|_t||ddd|_t d||_y) NrTr8)r rU)rr)rXr rU)rrr/)r rr rHconvx1convx2r6ln)rrout_dim factor_kernelrs rr zSimpleProjector.__init__+sw ot-/ tPDK G&QDK4U[\DKDV\]DK4)rcl|j|j|j|d}|gS)r[r)rrrr]s rr#zSimpleProjector.forward5s-ggdkk$++ad"345u rF)r%r&r'r r#r)r*s@rrr*s *rrr)r(typingrrrrnumpyrrtorch.nnr torch.nn.functional functionalrModulerr4rcr6rFrHr`rrrrrrrrs66  28E#xryy0@'A"ABCSV[cdfdmdm[n( &BII&R P P )"))).E"))EPbiir