hSjSrSSKJr SSKrSSKrSSKrSSKJr Sr S"Sjr "SS\R5r "SS \ 5r "S S \R5r"S S \ 5r"SS\R 5r"SS\R5r"SS\R5r"SS\R5r"SS\R5r"SS\R5r"SS\R5r"SS\R5r"SS\R5r"S S!\R5rg)#zConvolution modules.) annotationsN)ConvConv2 LightConvDWConvDWConvTranspose2d ConvTransposeFocus GhostConvChannelAttentionSpatialAttentionCBAMConcatRepConvIndexcUS:a:[U[5(a X S- -S-OUVs/sH o2US- -S-PM snnUc/[U[5(aUS-OUVs/sHo3S-PM snnU$s snfs snf)zPad to 'same' shape outputs.) isinstanceint)kpdxs U/home/james-whalen/.local/lib/python3.13/site-packages/ultralytics/nn/modules/conv.pyautopadrs}1u)!S11AQK!OQR7SQRAQU aQR7Sy C((AFq.Aq!Avq.A H8T.As A5"A:c^^\rSrSrSr\R "5rSU4SjjrSr Sr Sr U=r $)r'a# Standard convolution module with batch normalization and activation. Attributes: conv (nn.Conv2d): Convolutional layer. bn (nn.BatchNorm2d): Batch normalization layer. act (nn.Module): Activation function layer. default_act (nn.Module): Default activation function (SiLU). c T>[T U]5 [R"XX4[ X5U5XgSS9Ul[R "U5UlUSLaURUl g[U[R5(aXl g[R"5Ul g)a2 Initialize Conv layer with given parameters. Args: c1 (int): Number of input channels. c2 (int): Number of output channels. k (int): Kernel size. s (int): Stride. p (int, optional): Padding. g (int): Groups. d (int): Dilation. act (bool | nn.Module): Activation function. FgroupsdilationbiasTN) super__init__nnConv2drconv BatchNorm2dbn default_actrModuleIdentityact selfc1c2rsrgrr. __class__s rr% Conv.__init__4sw IIbaGA!,[T U]XX4XVXxS9 [R"XSU[ SXW5XgSS9Ulg)a3 Initialize Conv2 layer with given parameters. Args: c1 (int): Number of input channels. c2 (int): Number of output channels. k (int): Kernel size. s (int): Stride. p (int, optional): Padding. g (int): Groups. d (int): Dilation. act (bool | nn.Module): Activation function. r4rr.rFr N)r$r%r&r'rcv2r/s rr%Conv2.__init__ks= q<99RQ71a+;AX]^r7cURURURU5URU5-55$r9)r.r*r(rRr;s rr< Conv2.forward|s1xx ! txx{ :;<r7c[R"URRR5nUR SSVs/sHo"S-PM nnUR RRR5USS2SS2USUSS-2USUSS-24'URR=RU- slURS5 URUl gs snf)zFuse parallel convolutions.rNrrrR) torch zeros_liker(weightdatashaperRclone __delattr__rAr<)r0wris r fuse_convsConv2.fuse_convss   TYY--22 3WWQR[ )[!V[ )48HHOO4H4H4N4N4P!Q!qtax1!q 01 " (( *sC1)rRr<)rNrrT) rErFrGrHrIr%r<rArbrKrLrMs@rrr`s"_" = /))r7rcX^\rSrSrSrS\R "54U4SjjrSrSr U=r $)rz Light convolution module with 1x1 and depthwise convolutions. This implementation is based on the PaddleDetection HGNetV2 backbone. Attributes: conv1 (Conv): 1x1 convolution layer. conv2 (DWConv): Depthwise convolution layer. rc`>[TU]5 [XSSS9Ul[ X"X4S9Ulg)z Initialize LightConv layer with given parameters. Args: c1 (int): Number of input channels. c2 (int): Number of output channels. k (int): Kernel size for depthwise convolution. act (nn.Module): Activation function. rFr.N)r$r%rconv1rconv2)r0r1r2rr.r5s rr%LightConv.__init__s. "!/ BA/ r7cBURURU55$)z Apply 2 convolutions to input tensor. Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Output tensor. )rjrir;s rr<LightConv.forwardszz$**Q-((r7)rirj) rErFrGrHrIr&ReLUr%r<rKrLrMs@rrrs$"# 0 ) )r7rc0^\rSrSrSrSU4SjjrSrU=r$)rzDepth-wise convolution module.c N>[TU]XX4[R"X5XVS9 g)a  Initialize depth-wise convolution with given parameters. Args: c1 (int): Number of input channels. c2 (int): Number of output channels. k (int): Kernel size. s (int): Stride. d (int): Dilation. act (bool | nn.Module): Activation function. rQNr$r%mathgcd)r0r1r2rr3rr.r5s rr%DWConv.__init__s$ ")9QHr7rrrTrErFrGrHrIr%rKrLrMs@rrrs( I Ir7rc0^\rSrSrSrSU4SjjrSrU=r$)rz(Depth-wise transpose convolution module.c N>[TU]XX4XV[R"X5S9 g)a Initialize depth-wise transpose convolution with given parameters. Args: c1 (int): Number of input channels. c2 (int): Number of output channels. k (int): Kernel size. s (int): Stride. p1 (int): Padding. p2 (int): Output padding. )r!Nrr)r0r1r2rr3p1p2r5s rr%DWConvTranspose2d.__init__s$ rdhhr6FGr7rv)rrrrrxrMs@rrrs2 H Hr7rc^^\rSrSrSr\R "5rSU4SjjrSr Sr Sr U=r $)r aW Convolution transpose module with optional batch normalization and activation. Attributes: conv_transpose (nn.ConvTranspose2d): Transposed convolution layer. bn (nn.BatchNorm2d | nn.Identity): Batch normalization layer. act (nn.Module): Activation function layer. default_act (nn.Module): Default activation function (SiLU). c ~>[TU]5 [R"XX4XV(+S9UlU(a[R "U5O[R "5UlUSLaURUl g[U[R5(aXpl g[R "5Ul g)a- Initialize ConvTranspose layer with given parameters. Args: c1 (int): Number of input channels. c2 (int): Number of output channels. k (int): Kernel size. s (int): Stride. p (int): Padding. bn (bool): Use batch normalization. act (bool | nn.Module): Activation function. r#TN) r$r%r&ConvTranspose2dconv_transposer)r-r*r+rr,r.) r0r1r2rr3rr*r.r5s rr%ConvTranspose.__init__sz  00qvN(*"..$ '*d{4##z#ryy?Y?Y_a_j_j_lr7c`URURURU555$)z Apply transposed convolution, batch normalization and activation to input. Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Output tensor. r.r*rr;s rr<ConvTranspose.forwards'xx 3 3A 6788r7cBURURU55$)z Apply activation and convolution transpose operation to input. Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Output tensor. )r.rr;s rrAConvTranspose.forward_fusesxx++A.//r7r)rrrTTrDrMs@rr r s*'')Km$ 9 0 0r7r c6^\rSrSrSrSU4SjjrSrSrU=r$)r i z Focus module for concentrating feature information. Slices input tensor into 4 parts and concatenates them in the channel dimension. Attributes: conv (Conv): Convolution layer. c J>[TU]5 [US-X#XEXgS9Ulg)a Initialize Focus module with given parameters. Args: c1 (int): Number of input channels. c2 (int): Number of output channels. k (int): Kernel size. s (int): Stride. p (int, optional): Padding. g (int): Groups. act (bool | nn.Module): Activation function. rhN)r$r%rr() r0r1r2rr3rr4r.r5s rr%Focus.__init__*s& aa9 r7c UR[R"USSSS2SSS24USSSS2SSS24USSSS2SSS24USSSS2SSS244S55$)z Apply Focus operation and convolution to input tensor. Input shape is (B, C, W, H) and output shape is (B, 4C, W/2, H/2). Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Output tensor. .Nrr)r(rYcatr;s rr< Focus.forward;syyAc3Q3!m$4aQTT3Q36G3PSRSPSUVUYXYUY>IZ\]^acdcgfgcgijimlmim^m\n#oqrsttr7)r()rrNrT rErFrGrHrIr%r<rKrLrMs@rr r s:" u ur7r c6^\rSrSrSrSU4SjjrSrSrU=r$)r iKa  Ghost Convolution module. Generates more features with fewer parameters by using cheap operations. Attributes: cv1 (Conv): Primary convolution. cv2 (Conv): Cheap operation convolution. References: https://github.com/huawei-noah/Efficient-AI-Backbones c r>[TU]5 US-n[XX4SXVS9Ul[XwSSSXvS9Ulg)a  Initialize Ghost Convolution module with given parameters. Args: c1 (int): Number of input channels. c2 (int): Number of output channels. k (int): Kernel size. s (int): Stride. g (int): Groups. act (bool | nn.Module): Activation function. rNrhr)r$r%rcv1rR) r0r1r2rr3r4r.c_r5s rr%GhostConv.__init__Ys?  1WdA71dB8r7crURU5n[R"X RU54S5$)z Apply Ghost Convolution to input tensor. Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Output tensor with concatenated features. r)rrYrrR)r0rys rr<GhostConv.forwardjs- HHQKyy!XXa[)1--r7)rrRrwrrMs@rr r Ks 9" . .r7r c^\rSrSrSr\R "5rS U4SjjrSr Sr Sr \ S5r SrS rS rU=r$) rixa RepConv module with training and deploy modes. This module is used in RT-DETR and can fuse convolutions during inference for efficiency. Attributes: conv1 (Conv): 3x3 convolution. conv2 (Conv): 1x1 convolution. bn (nn.BatchNorm2d, optional): Batch normalization for identity branch. act (nn.Module): Activation function. default_act (nn.Module): Default activation function (SiLU). References: https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py c >[T U]5 US:XaUS:XdeX`lXlX lUSLa UR O5[ U[R5(aUO[R"5Ul U (aX!:XaUS:Xa[R"US9OSUl [XX4XVSS9Ul[XSXEUS-- USS9Ulg) a Initialize RepConv module with given parameters. Args: c1 (int): Number of input channels. c2 (int): Number of output channels. k (int): Kernel size. s (int): Stride. p (int): Padding. g (int): Groups. d (int): Dilation. act (bool | nn.Module): Activation function. bn (bool): Use batch normalization for identity branch. deploy (bool): Deploy mode for inference. rdrT) num_featuresNF)rr4r.r)r$r%r4r1r2r+rr&r,r-r.r)r*rrirj) r0r1r2rr3rr4rr.r*deployr5s rr%RepConv.__init__s Av!q&  '*d{4##z#ryy?Y?Y_a_j_j_l57BHa"..b1UY"!!e< "!QqAv:!G r7cBURURU55$)zw Forward pass for deploy mode. Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Output tensor. r@r;s rrARepConv.forward_fuserCr7cURcSOURU5nURURU5URU5-U-5$)zy Forward pass for training mode. Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Output tensor. r)r*r.rirj)r0rid_outs rr<RepConv.forwardsDggo4771:xx 1  1 5>??r7cURUR5upURUR5up4URUR5upVXR U5-U-X$-U-4$)z Calculate equivalent kernel and bias by fusing convolutions. Returns: (torch.Tensor): Equivalent kernel (torch.Tensor): Equivalent bias )_fuse_bn_tensorrirjr*_pad_1x1_to_3x3_tensor)r0 kernel3x3bias3x3 kernel1x1bias1x1kernelidbiasids rget_equivalent_kernel_bias"RepConv.get_equivalent_kernel_biassp"11$**= !11$**= //866yAAHLgN_bhNhhhr7cbUcg[RRRU/SQ5$)z Pad a 1x1 kernel to 3x3 size. Args: kernel1x1 (torch.Tensor): 1x1 convolution kernel. Returns: (torch.Tensor): Padded 3x3 kernel. r)rrrr)rYr& functionalpad)rs rrRepConv._pad_1x1_to_3x3_tensors*  88&&**9lC Cr7cUcg[U[5(aURRnURR nURR nURRnURRnURRnGO'[U[R5(Ga[US5(dURUR-n[R"URUSS4[R S9n [#UR5H n SXX-SS4'M [$R&"U 5R)URR*5UlUR,nUR nUR nURnURnURnWW-R/5n WU - R1SSSS5n WU -WWU-U - - 4$)z Fuse batch normalization with convolution weights. Args: branch (Conv | nn.BatchNorm2d | None): Branch to fuse. Returns: kernel (torch.Tensor): Fused kernel. bias (torch.Tensor): Fused bias. )rr id_tensorrd)dtyper)rrr(r[r* running_mean running_varr#epsr&r)hasattrr1r4npzerosfloat32rangerY from_numpytodevicersqrtreshape) r0branchkernelrrgammabetar input_dim kernel_valuerastdts rrRepConv._fuse_bn_tensors > fd # #[[''F!9911L ))//KII$$E99>>D))--C  / /4-- GGtvv- !xx)Q(B"**U twwA;[TU]5 [R"S5Ul[R "XSSSSS9Ul[R"5Ulg)z[ Initialize Channel-attention module. Args: channels (int): Number of input channels. rrTrN) r$r%r&AdaptiveAvgPool2dpoolr'fcSigmoidr.)r0channelsr5s rr%ChannelAttention.__init__-sG ((+ ))H1adC::<r7cdXRURURU555-$)z Apply channel attention to input tensor. Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Channel-attended output tensor. r.rrr;s rr<ChannelAttention.forward9s'88DGGDIIaL1222r7r)rrreturnNone)r torch.TensorrrrrMs@rr r s    3 3r7r c6^\rSrSrSrSU4SjjrSrSrU=r$)r iFa Spatial-attention module for feature recalibration. Applies attention weights to spatial dimensions based on channel statistics. Attributes: cv1 (nn.Conv2d): Convolution layer for spatial attention. act (nn.Sigmoid): Sigmoid activation for attention weights. c>[TU]5 US;dS5eUS:XaSOSn[R"SSXSS9Ul[R "5Ulg ) zo Initialize Spatial-attention module. Args: kernel_size (int): Size of the convolutional kernel (3 or 7). >rdzkernel size must be 3 or 7rrdrrF)rr#N)r$r%r&r'rrr.)r0rrr5s rr%SpatialAttention.__init__QsU f$B&BB$"a'!Q99Q;eL::<r7cXRUR[R"[R"USSS9[R "USSS9S/S555-$)z Apply spatial attention to input tensor. Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Spatial-attended output tensor. rT)keepdimr)r.rrYrmeanmaxr;s rr<SpatialAttention.forward^sV88DHHUYY 1a0NPUPYPYZ[]^hlPmnoPp/qst%uvwwwr7)r.rrrrMs@rr r Fs   x xr7r c6^\rSrSrSrSU4SjjrSrSrU=r$)rika Convolutional Block Attention Module. Combines channel and spatial attention mechanisms for comprehensive feature refinement. Attributes: channel_attention (ChannelAttention): Channel attention module. spatial_attention (SpatialAttention): Spatial attention module. cb>[TU]5 [U5Ul[ U5Ulg)z Initialize CBAM with given parameters. Args: c1 (int): Number of input channels. kernel_size (int): Size of the convolutional kernel for spatial attention. N)r$r%r channel_attentionr spatial_attention)r0r1rr5s rr% CBAM.__init__vs* !1"!5!1+!>r7cBURURU55$)z Apply channel and spatial attention sequentially to input tensor. Args: x (torch.Tensor): Input tensor. Returns: (torch.Tensor): Attended output tensor. )rrr;s rr< CBAM.forwards %%d&<&[TU]5 Xlg)ze Initialize Concat module. Args: dimension (int): Dimension along which to concatenate tensors. N)r$r%r)r0 dimensionr5s rr%Concat.__init__s r7cB[R"XR5$)z Concatenate input tensors along specified dimension. Args: x (list[torch.Tensor]): List of input tensors. Returns: (torch.Tensor): Concatenated tensor. )rYrrr;s rr<Concat.forwardsyyFF##r7)r)rrzlist[torch.Tensor]rrMs@rrrs $ $r7rc:^\rSrSrSrSU4SjjrSSjrSrU=r$)rizd Returns a particular index of the input. Attributes: index (int): Index to select from input. c.>[TU]5 Xlg)zN Initialize Index module. Args: index (int): Index to select from input. N)r$r%index)r0rr5s rr%Index.__init__s  r7cXR$)z Select and return a particular index from input. Args: x (list[torch.Tensor]): List of input tensors. Returns: (torch.Tensor): Selected tensor. rr;s rr< Index.forwards}r7r)rrrrMs@rrrs  r7r)Nr)rI __future__rrsnumpyrrYtorch.nnr&__all__rr,rrrrrrr r r rr r rrrrvr7rr s"   $ 6&2996&r;)D;)|#) #)LITI$H**H$50BII50p'uBII'uV*. *.Zc*biic*L%3ryy%3P"xryy"xJ!A299!AH$RYY$>BIIr7