o oi?@sddlmZddlZddlmmZddlmZddlmZm Z m Z Gdddej Z GdddeZ Gd d d eZGd d d eZdddZdddZdS)) annotationsN)nn)ModuleTensorpadcs,eZdZ      ddfdd ZZS) ConvNormActreluconvnormact in_channelsint out_channels kernel_sizestridestrgroups conv_naming norm_naming act_namingreturnNonec  st|ddkr$|dd} |d| } t| | | | f|_d} n|dd} t||||| d|d} t|}tjtjtj d|dd}| || | ||| | |dS)NrrF)r silunoneT)inplace) super__init__r ZeroPad2drConv2d BatchNorm2dReLUSiLUIdentity __setattr__)selfr rrrr rrrrp1p2paddingr r activation __class__P/home/ubuntu/.local/lib/python3.10/site-packages/kornia/contrib/models/common.pyrs       zConvNormAct.__init__)rr rr r r )r rrrrrrrr rrrrrrrrrrr)__name__ __module__ __qualname__r __classcell__r-r-r+r.rsrcs,eZdZ ddfd d ZdddZZS)MLPF input_dimr hidden_dim output_dim num_layerssigmoid_outputboolrrcsRt||_|g|d}tddt|g|g||D|_||_dS)Nrcss |] \}}t||VqdSN)rLinear).0nkr-r-r. FszMLP.__init__..)rrr7r ModuleListziplayersr8)r&r4r5r6r7r8hr+r-r.r@s * z MLP.__init__xrcCsLt|jD]\}}||jdkrt||n||}q|jr$t|}|S)Nr) enumeraterBr7Fr r8sigmoid)r&rDilayerr-r-r.forwardIs & z MLP.forward)F) r4rr5rr6rr7rr8r9rrrDrrrr/r0r1rrJr2r-r-r+r.r3?s r3cs.eZdZdZddfd d ZdddZZS)DropPathzYDrop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).T drop_probfloat scale_by_keepr9rrcst||_||_dSr:)rrrOrQ)r&rOrQr+r-r.rVs  zDropPath.__init__rDrcCsh|jdks|js |Sd|j}|jdfd|jd}|||}|dkr0|jr0||||S)NrNrr)r)rOtrainingshapendim new_empty bernoulli_rQdiv_)r&rD keep_probrS random_tensorr-r-r.rJ[s  zDropPath.forward)rNT)rOrPrQr9rrrK)r/r0r1__doc__rrJr2r-r-r+r.rMSsrMcs*eZdZddfdd Zdd d ZZS) LayerNorm2dư> num_channelsrepsrPrrcs8ttt||_tt||_||_ dSr:) rrr Parametertorchonesweightzerosbiasr^)r&r]r^r+r-r.ris  zLayerNorm2d.__init__rDrcCsl|jddd}||djddd}||||j}|jddddf||jddddf}|S)NrT)keepdimr)meanpowr^sqrtrbrd)r&rDusr-r-r.rJos ,zLayerNorm2d.forward)r\)r]rr^rPrrrKrLr-r-r+r.r[hsr[rDr window_sizerrtuple[Tensor, tuple[int, int]]c Cs|j\}}}}||||}||||}|dks|dkr*t|ddd|d|f}||||}} |||||| |||}|ddddddd|||} | || ffS)aCPartition into non-overlapping windows with padding if needed. Args: x: input tokens with [B, H, W, C]. window_size: window size. Returns: windows: windows after partition with [B * num_windows, window_size, window_size, C]. (Hp, Wp): padded height and width before partition rrr)rSrviewpermute contiguous) rDrkBHWCpad_hpad_wHpWpwindowsr-r-r.window_partitionws $ r}r|pad_hwtuple[int, int]hwc Cs|\}}|\}}|jd||||}||||||||d} | dddddd|||d} ||ks=||krO| ddd|d|ddf} | S) ayWindow unpartition into original sequences and removing padding. Args: windows: input tokens with [B * num_windows, window_size, window_size, C]. window_size: window size. pad_hw: padded height and width (Hp, Wp). hw: original height and width (H, W) before padding. Returns: x: unpartitioned sequences with [B, H, W, C]. rrprrmrrnroN)rSrqrrrs) r|rkr~rrzr{rurvrtrDr-r-r.window_unpartitions $$r)rDrrkrrrl) r|rrkrr~rrrrr) __future__rr`torch.nn.functionalr functionalrF kornia.corerrr Sequentialrr3rMr[r}rr-r-r-r.s  $