o ߥi @s^ddlmZGdddejZ   dddZd d ZGd d d ejZGd ddejZdS)Ncs*eZdZdZd fdd ZddZZS) FPNFusionModulez8 This is a fpn-style cross-scale feature fusion module" Fcs4t ||_||_||_t||||d|_dS)N)use_bnn_block)super__init__ embed_dimsfuse_dimr_make_multi_scale_layers multi_scaler)selfr r rr __class__X/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/cv/vidt/fpn_fusion.pyr s  zFPNFusionModule.__init__cCs|}t|jdddD]!}t|jd|d||||<t|jd|||||<q g}t|jdddD]6}||jdkrQt|jd|||gd}nt|jd||||g||dd}||q8|S)Nlayer__rnp_norm_ refinenet_)rangergetattrr sizeappend)r x_blocksidxrefined_embedspathrrrforwards0    zFPNFusionModule.forward)rrF__name__ __module__ __qualname____doc__rr! __classcell__rrrrrsrrrFc sfddt|D}t}t|dddD]W} d|d}||tj||||ddd|}||td ||d |}||t|tjjt |d|dj dd tj t |d|dj d q|S) Ncsg|]}qSrr).0_ out_shaperr =sz,_make_multi_scale_layers..rrrr) kernel_sizer r)gainr) rnnModule add_moduleConv2d GroupNorm_make_fusion_blockinitxavier_uniform_rweight constant_bias) in_shaper+rgroupsr out_shapesr r layer_namerr*rr 7s(    r cCst|td|dddS)z. We use a resnet bottleneck structure for fpn FT)bnexpand align_corners)FeatureFusionBlockr0ReLU)featuresrrrrr5Xsr5cs0eZdZdZ   dfdd ZddZZS) rBz Feature fusion block FTcsxtt|||_d|_||_|}|jdur|d}tj||dddd|_t ||||_ t ||||_ tj |_dS)a-Init. Args: features (int): channel dim of the input feature activation: activation function to use bn: whether to use bn expand: whether to expand feature or not align_corners: whether to use align_corners for interpolation rTr)r-r:r<N)rrBrrAr<r@r0r3 smoothingResidualConvUnit resConfUnit1 resConfUnit2 quantizedFloatFunctionalskip_add)r rD activationr?r@rA out_featuresrrrrgs" zFeatureFusionBlock.__init__cCs^|d}t|dkr(tjj||d|jd}||}||d}|j||}| |}|S)aR Forward pass. Args xs: xs[0]: the feature refined from the previous step, xs[1]: the next scale features to fuse up_size: the size for upsampling; xs[0] is upsampled before merging with xs[1] Returns: output: the fused feature, which is fed to the next fusion step as an input rrbilinear)rmoderAr) lenr0 functional interpolaterArErGrKaddrH)r xsup_sizeoutputresrrrr!s   zFeatureFusionBlock.forward)FFTr"rrrrrBds$rBcs(eZdZdZfddZddZZS)rFz Residual convolution module. c st||_d|_tj|ddd|j |jd|_tjddddd|j |jd|_tjd|dd|j |jd|_|jdurPt ||_ t ||_ t ||_ ||_ tj|_dS)zInit. Args: features (int): channel dim of the input activation: activation function bn: whether to use bn r@)r-strider:r<)r-rYpaddingr:r<TN)rrr?r<r0r3conv1conv2conv3 BatchNorm2dbn1bn2bn3rLrIrJrK)r rDrLr?rrrrsD      zResidualConvUnit.__init__cCs||}||}|jdur||}||}||}|jdur(||}||}||}|jdur<||}|jdkrF| |}|j ||S)z Forward pass Args: x (tensor): input feature Returns: tensor: output feature Tr) rLr\r?r`r]rar^rbr< conv_mergerKrS)r xoutrrrr!s             zResidualConvUnit.forwardr"rrrrrFs .rF)rrF)torch.nnr0r1rr r5rBrFrrrrs 1 ! G