a  du@sddlZddlZddlZddlmZddlmZddlmZm Z ddl m Z ddl m Z GdddejZd d ZGd d d ejZGd ddejZGdddejZGdddejZGdddejZGdddejZGdddejZGdddejZe GdddejZGdddejZGdd d ejZGd!d"d"ejZGd#d$d$ejZe Gd%d&d&ejZ dS)'N)nn) functional)FusedLeakyReLUfused_leaky_relu) upfirdn2d) ARCH_REGISTRYc@seZdZddZdS) NormStyleCodecCs"|ttj|dddddS)zNormalize the style codes. Args: x (Tensor): Style codes with shape (b, c). Returns: Tensor: Normalized tensor. T)Zdimkeepdim:0yE>)torchrsqrtmeanselfxre/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/basicsr/archs/stylegan2_arch.pyforwards zNormStyleCode.forwardN)__name__ __module__ __qualname__rrrrrr srcCsJtj|tjd}|jdkr:|dddf|dddf}||}|S)zMake resampling kernel for UpFirDn. Args: k (list[int]): A list indicating the 1D resample kernel magnitude. Returns: Tensor: 2D resampled kernel. )Zdtyper N)r Ztensorfloat32ndimsum)krrrmake_resample_kernels    rcs2eZdZdZd fdd ZddZddZZS) UpFirDnUpsampleaUpsample, FIR filter, and downsample (upsampole version). References: 1. https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.upfirdn.html # noqa: E501 2. http://www.ece.northwestern.edu/local-apps/matlabhelp/toolbox/signal/upfirdn.html # noqa: E501 Args: resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. factor (int): Upsampling scale factor. Default: 2. r csXtt|t||d|_||_|jjd|}|dd|d|df|_dS)Nr rr )superr__init__rkernelfactorshapepadrresample_kernelr"r$ __class__rrr 8s zUpFirDnUpsample.__init__cCs"t||j||jd|jd}|SNr )ZupZdownr$rr!type_asr"r$rroutrrrr@szUpFirDnUpsample.forwardcCs|jjd|jdSNz(factor=)r(rr"rrrr__repr__DszUpFirDnUpsample.__repr__)r rrr__doc__r rr2 __classcell__rrr'rr+s rcs2eZdZdZd fdd ZddZddZZS) UpFirDnDownsamplezUpsample, FIR filter, and downsample (downsampole version). Args: resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. factor (int): Downsampling scale factor. Default: 2. r csHtt|t||_||_|jjd|}|dd|df|_dS)Nrr r )rr6r rr!r"r#r$r%r'rrr Qs  zUpFirDnDownsample.__init__cCs"t||j|d|j|jd}|Sr)r*r,rrrrYszUpFirDnDownsample.forwardcCs|jjd|jdSr.r0r1rrrr2]szUpFirDnDownsample.__repr__)r r3rrr'rr6Hsr6cs2eZdZdZd fdd ZddZddZZS) UpFirDnSmoothakUpsample, FIR filter, and downsample (smooth version). Args: resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. upsample_factor (int): Upsampling scale factor. Default: 1. downsample_factor (int): Downsampling scale factor. Default: 1. kernel_size (int): Kernel size: Default: 1. r cstt|||_||_t||_|dkr<|j|d|_|dkr|jjd||d}|dd|d|ddf|_n<|dkr|jjd||d}|dd|df|_nt dS)Nr r r) rr7r upsample_factordownsample_factorrr!r#r$NotImplementedError)rr&r8r9 kernel_sizer$r'rrr ls $zUpFirDnSmooth.__init__cCs t||j|dd|jd}|Sr))rr!r+r$r,rrrr}szUpFirDnSmooth.forwardcCs|jjd|jd|jdS)Nz(upsample_factor=z, downsample_factor=r/)r(rr8r9r1rrrr2szUpFirDnSmooth.__repr__)r r r r3rrr'rr7as r7cs2eZdZdZd fdd Zdd Zd d ZZS) EqualLinearaEqualized Linear as StyleGAN2. Args: in_channels (int): Size of each sample. out_channels (int): Size of each output sample. bias (bool): If set to ``False``, the layer will not learn an additive bias. Default: ``True``. bias_init_val (float): Bias initialized value. Default: 0. lr_mul (float): Learning rate multiplier. Default: 1. activation (None | str): The activation after ``linear`` operation. Supported: 'fused_lrelu', None. Default: None. Trr Ncstt|||_||_||_||_|jdvr@td|ddt |||_ t t ||||_|rt t |||_n |dddS)N) fused_lreluNz'Wrong activation value in EqualLinear: z*Supported ones are: ['fused_lrelu', None].r bias)rr<r in_channels out_channelslr_mul activation ValueErrormathsqrtscaler Parameterr randnZdiv_weightzerosfill_r>register_parameter)rr?r@r> bias_init_valrArBr'rrr s zEqualLinear.__init__cCsb|jdurd}n |j|j}|jdkrFt||j|j}t||}ntj||j|j|d}|S)Nr=)r>)r>rArBFZlinearrIrFr)rrr>r-rrrrs    zEqualLinear.forwardcCs*|jjd|jd|jd|jdudS)N (in_channels=, out_channels=, bias=r/)r(rr?r@r>r1rrrr2s zEqualLinear.__repr__)Trr Nr3rrr'rr<s  r<cs2eZdZdZd fdd Zdd Zd d ZZS) ModulatedConv2daModulated Conv2d used in StyleGAN2. There is no bias in ModulatedConv2d. Args: in_channels (int): Channel number of the input. out_channels (int): Channel number of the output. kernel_size (int): Size of the convolving kernel. num_style_feat (int): Channel number of style features. demodulate (bool): Whether to demodulate in the conv layer. Default: True. sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None. resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. Default: (1, 3, 3, 1). eps (float): A value added to the denominator for numerical stability. Default: 1e-8. TNr rTr r c stt|||_||_||_||_||_||_|jdkrPt |dd|d|_ n<|jdkrnt |dd|d|_ n|jdurznt d|jddt ||d|_t||ddddd |_ttd|||||_|d|_dS) Nupsampler r r8r9r; downsamplezWrong sample mode z6, supported ones are ['upsample', 'downsample', None].Tr>rMrArB)rrRr r?r@r; demodulate sample_modeepsr7smoothrCrDrErFr< modulationrrGr rHrIpadding) rr?r@r;num_style_featrYrZr&r[r'rrr s0       zModulatedConv2d.__init__c Cs|j\}}}}|||d|dd}|j|j|}|jrtt|d gd|j }||||j ddd}|||j ||j |j }|j dkr(|d||||}|||j ||j |j }|dd|||j |j |j }tj||dd|d} | j||j g| jddR} || } n|j dkr||}|jd||g|jddR}tj||dd|d} | j||j g| jddR} nH|d||||}tj|||j|d } | j||j g| jddR} | S) zForward function. Args: x (Tensor): Tensor with shape (b, c, h, w). style (Tensor): Tensor with shape (b, num_style_feat). Returns: Tensor: Modulated tensor after convolution. r r r rTrUr)r^stridegroupsrarW)r^rc)r#r]viewrFrIrYr rpowrr[r@r;rZZ transposeZreshaperNZconv_transpose2dr\conv2dr^) rrstylebchwrIZdemodr-rrrrs.   "    "" zModulatedConv2d.forwardc Cs6|jjd|jd|jd|jd|jd|jd S)NrOrP, kernel_size=z , demodulate=z, sample_mode=r/)r(rr?r@r;rYrZr1rrrr2szModulatedConv2d.__repr__)TNrSr r3rrr'rrRs%*rRcs,eZdZdZd fdd Zd ddZZS) StyleConva=Style conv. Args: in_channels (int): Channel number of the input. out_channels (int): Channel number of the output. kernel_size (int): Size of the convolving kernel. num_style_feat (int): Channel number of style features. demodulate (bool): Whether demodulate in the conv layer. Default: True. sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None. resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. Default: (1, 3, 3, 1). TNrSc sFtt|t|||||||d|_ttd|_ t ||_ dS)N)rYrZr&r ) rrmr rRmodulated_convrrGr rJrIractivate)rr?r@r;r_rYrZr&r'rrr /szStyleConv.__init__c CsR|||}|dur6|j\}}}}||d||}||j|}||}|SNr )rnr#Z new_emptyZnormal_rIro) rrrgnoiser-rh_rjrkrrrrCs  zStyleConv.forward)TNrS)Nrrrr4r rr5rrr'rrm s rmcs,eZdZdZd fdd Zd ddZZS) ToRGBaRTo RGB from features. Args: in_channels (int): Channel number of input. num_style_feat (int): Channel number of style features. upsample (bool): Whether to upsample. Default: True. resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. Default: (1, 3, 3, 1). TrScsZtt||r"t|dd|_nd|_t|dd|ddd|_tt dddd|_ dS)Nr )r"rTr F)r;r_rYrZ) rrtr rrUrRrnrrGr rJr>)rr?r_rUr&r'rrr [s zToRGB.__init__NcCs:|||}||j}|dur6|jr.||}||}|S)aForward function. Args: x (Tensor): Feature tensor with shape (b, c, h, w). style (Tensor): Tensor with shape (b, num_style_feat). skip (Tensor): Base/skip tensor. Default: None. Returns: Tensor: RGB images. N)rnr>rU)rrrgskipr-rrrres   z ToRGB.forward)TrS)Nrsrrr'rrtPs  rtcs(eZdZdZfddZddZZS) ConstantInputzConstant input. Args: num_channel (int): Channel number of constant input. size (int): Spatial size of constant input. cs*tt|ttd||||_dSrp)rrvr rrGr rHrI)rZ num_channelsizer'rrr szConstantInput.__init__cCs|j|ddd}|Srp)rIrepeat)rbatchr-rrrrszConstantInput.forwardrsrrr'rrvys rvcsDeZdZdZdfdd Zd d Zd d ZddZdddZZ S)StyleGAN2GeneratoraStyleGAN2 Generator. Args: out_size (int): The spatial size of outputs. num_style_feat (int): Channel number of style features. Default: 512. num_mlp (int): Layer number of MLP style layers. Default: 8. channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1). lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01. narrow (float): Narrow ratio for channels. Default: 1.0. r rS{Gz?r c s8tt|||_tg}t|D]} |t||dd|ddq$tj ||_ t d|t d|t d|t d|t d||t d||t d||t d ||t d ||d } | |_ t | d d d|_t| d | d d|dd|d|_t| d |d|d|_t t|d|_|jddd|_|jdd|_t|_t|_t|_| d } t|jD]<} d| dd} dd| | g}|jd| tj|qxtd|jdD]l} | d| }|jt| |d|dd|d|jt||d|dd|d|jt||d|d|} qdS)NTrr=rXr{@  48Z16Z32Z64Z128Z256Z512Z1024rra)rwrT)r;r_rYrZr&F)rUr&r r rqrU) rrzr r_rrangeappendr<r Sequential style_mlpintchannelsrvconstant_inputrm style_conv1rtto_rgb1rDloglog_size num_layers num_latentZ ModuleList style_convsto_rgbsModulenoisesZregister_bufferr rH)rout_sizer_Znum_mlpchannel_multiplierr&Zlr_mlpnarrowZstyle_mlp_layersirr?Z layer_idx resolutionr#r@r'rrr s              zStyleGAN2Generator.__init__c Csj|jjj}tjdddd|dg}td|jdD]4}tdD]&}|tjddd|d||dqz.StyleGAN2Generator.forward..Ncsg|]}tjd|qSrq)getattrr)rrr1rrr!rr rrTr r)rrrlenrrZ unsqueezerxrandomrandintr catrr#rrziprr)rZstylesZinput_is_latentrqZrandomize_noiseZ truncationZtruncation_latentZ inject_indexZreturn_latentsZstyle_truncationrgrZlatent1Zlatent2r-rurconv1conv2Znoise1Znoise2Zto_rgbimagerr1rrsJ   * zStyleGAN2Generator.forward)r{r|r rSr}r )FNTr NNF) rrrr4r rrrrr5rrr'rrzs&R rzcs*eZdZdZdfdd ZddZZS)ScaledLeakyReLUz_Scaled LeakyReLU. Args: negative_slope (float): Negative slope. Default: 0.2. 皙?cstt|||_dSr)rrr negative_slope)rrr'rrr TszScaledLeakyReLU.__init__cCstj||jd}|tdS)N)rr )rNZ leaky_relurrDrEr,rrrrXszScaledLeakyReLU.forward)rrsrrr'rrMsrcs2eZdZdZd fdd ZddZd d ZZS) EqualConv2daEqualized Linear as StyleGAN2. Args: in_channels (int): Channel number of the input. out_channels (int): Channel number of the output. kernel_size (int): Size of the convolving kernel. stride (int): Stride of the convolution. Default: 1 padding (int): Zero-padding added to both sides of the input. Default: 0. bias (bool): If ``True``, adds a learnable bias to the output. Default: ``True``. bias_init_val (float): Bias initialized value. Default: 0. r rTcstt|||_||_||_||_||_dt ||d|_ t t |||||_|rzt t |||_n |dddS)Nr r r>)rrr r?r@r;rbr^rDrErFrrGr rHrIrJrKr>rL)rr?r@r;rbr^r>rMr'rrr lszEqualConv2d.__init__cCs&tj||j|j|j|j|jd}|S)N)r>rbr^)rNrfrIrFr>rbr^r,rrrr{s zEqualConv2d.forwardcCsB|jjd|jd|jd|jd|jd|jd|jdudS)NrOrPrlz , stride=z , padding=rQr/)r(rr?r@r;rbr^r>r1rrrr2szEqualConv2d.__repr__)r rTrr3rrr'rr]s rcs"eZdZdZdfdd ZZS) ConvLayeraConv Layer used in StyleGAN2 Discriminator. Args: in_channels (int): Channel number of the input. out_channels (int): Channel number of the output. kernel_size (int): Kernel size. downsample (bool): Whether downsample by a factor of 2. Default: False. resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1). bias (bool): Whether with bias. Default: True. activate (bool): Whether use activateion. Default: True. FrSTc sg}|r*|t|dd|dd} d|_nd} |d|_|t|||| |j|oR| d|r|rr|t|n|tdtt|j|dS)Nr r rVr)rbr^r>r) rr7r^rrrrrr ) rr?r@r;rWr&r>roZlayersrbr'rrr s( zConvLayer.__init__)FrSTT)rrrr4r r5rrr'rrs rcs*eZdZdZdfdd ZddZZS)ResBlockaResidual block used in StyleGAN2 Discriminator. Args: in_channels (int): Channel number of the input. out_channels (int): Channel number of the output. resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1). rSc sVtt|t||dddd|_t||dd|ddd|_t||dd|ddd|_dS)NrTTr>ro)rWr&r>ror F)rrr rrrru)rr?r@r&r'rrr szResBlock.__init__cCs4||}||}||}||td}|S)Nr )rrrurDrE)rrr-rurrrrs    zResBlock.forward)rSrsrrr'rrs  rcs*eZdZdZd fdd Zdd ZZS) StyleGAN2Discriminatora0StyleGAN2 Discriminator. Args: out_size (int): The spatial size of outputs. channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1). stddev_group (int): For group stddev statistics. Default: 4. narrow (float): Narrow ratio for channels. Default: 1.0. r rSrar c sjtt|td|td|td|td|td||td||td||td||td||d }tt|d}td ||d d d d g}||} t|dd D],} |d| d } |t | | || } qt j ||_ t| d |dd d d d |_ t t|ddd|dd dd ddt|dd d dd dd|_||_d |_dS)Nr{r~rrrrrr rTr Trrrarr=rX)rrr rrDrrrrrrr conv_body final_convr< final_linear stddev_group stddev_feat) rrrr&rrrrrr?rr@r'rrr s8       zStyleGAN2Discriminator.__init__c Cs||}|j\}}}}t||j}||d|j||j||}t|jdddd}|j gddd d }| |d ||}t ||gd }| |}||d}||}|S) NrrF)Zunbiasedr r`T)Zkeepdimsr r )rr#minrrdrr rEvarrZsqueezerxrrr) rrr-rhrirjrkgroupZstddevrrrrs     zStyleGAN2Discriminator.forward)r rSrar rsrrr'rrs #r)!rDrr rZtorch.nnrrNZbasicsr.ops.fused_actrrZbasicsr.ops.upfirdn2drZbasicsr.utils.registryrrrrrr6r7r<rRrmrtrvregisterrzrrrrrrrrrrs4    %0j0)C12