a  d%W@sddlZddlZddlZddlmZmZddlmZddlmZddl m Z Gdddej Z Gdd d ej ZGd d d ej ZGd d d ej ZGdddej ZGdddej ZeGdddej ZGdddej ZGdddej ZGdddejZGdddej ZdS)N)FusedLeakyReLUfused_leaky_relu) ARCH_REGISTRY)nn) functionalc@seZdZddZdS) NormStyleCodecCs"|ttj|dddddS)zNormalize the style codes. Args: x (Tensor): Style codes with shape (b, c). Returns: Tensor: Normalized tensor. T)Zdimkeepdim:0yE>)torchrsqrtmeanselfxrm/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/gfpgan/archs/stylegan2_bilinear_arch.pyforward s zNormStyleCode.forwardN)__name__ __module__ __qualname__rrrrrr srcs2eZdZdZd 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)superr__init__ in_channels out_channelslr_mul activation ValueErrormathsqrtscaler Parameterr randnZdiv_weightzerosfill_rregister_parameter)rrrr bias_init_valrr  __class__rrr&s zEqualLinear.__init__cCsb|jdurd}n |j|j}|jdkrFt||j|j}t||}ntj||j|j|d}|S)Nr)r)rrr FZlinearr'r$r)rrroutrrrr7s    zEqualLinear.forwardcCs*|jjd|jd|jd|jdudS)N (in_channels=, out_channels=, bias=))r-rrrrrrrr__repr__Cs zEqualLinear.__repr__)Trr Nrrr__doc__rrr5 __classcell__rrr,rrs  rcs2eZdZdZd fdd Zdd Zd d ZZS) ModulatedConv2da|Modulated Conv2d used in StyleGAN2. There is no bias in ModulatedConv2d. Args: in_channels (int): Channel number of the input. out_channels (int): Channel number of the output. kernel_size (int): Size of the convolving kernel. num_style_feat (int): Channel number of style features. demodulate (bool): Whether to demodulate in the conv layer. Default: True. sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None. eps (float): A value added to the denominator for numerical stability. Default: 1e-8. TNr bilinearc stt|||_||_||_||_||_||_||_ |j dkrJd|_ nd|_ dt ||d|_ t||ddddd|_ttd|||||_|d|_dS)NnearestFr rTrr+rr )rr9rrr kernel_size demodulate sample_modeepsinterpolation_mode align_cornersr"r#r$r modulationrr%r r&r'padding) rrrr=num_style_featr>r?r@rAr,rrrZs"   zModulatedConv2d.__init__c Cs*|j\}}}}|||d|dd}|j|j|}|jrtt|d gd|j }||||j ddd}|||j ||j |j }|j dkrtj|d|j|jd}n |j dkrtj|d|j|jd}|j\}}}}|d||||}tj|||j|d} | j||j g| jdd R} | 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)rupsampleZ scale_factormoderB downsample?)rDgroupsrG)shaperCviewr$r'r>r r powsumr@rr=r?r. interpolaterArBconv2drD) rrstylebchwr'Zdemodr/rrrrxs     zModulatedConv2d.forwardc Cs6|jjd|jd|jd|jd|jd|jd S)Nr0r1, kernel_size=z , demodulate=z, sample_mode=r3)r-rrrr=r>r?r4rrrr5szModulatedConv2d.__repr__)TNr r:r6rrr,rr9Hs#r9cs,eZdZdZd fdd Zd ddZZS) StyleConvaStyle 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. TNr:c sFtt|t|||||||d|_ttd|_ t ||_ dS)N)r>r?rAr ) rrZrr9modulated_convrr%r r(r'ractivate)rrrr=rEr>r?rAr,rrrszStyleConv.__init__c CsR|||}|dur6|j\}}}}||d||}||j|}||}|SNr )r[rNZ new_emptyZnormal_r'r\) rrrTnoiser/rU_rWrXrrrrs  zStyleConv.forward)TNr:)Nrrrr7rrr8rrr,rrZs rZcs,eZdZdZd fdd Zd ddZZS) ToRGBzTo 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. Tr:c sftt|||_||_|jdkr,d|_nd|_t|dd|dd|d|_t t dddd|_ dS)Nr;FrFr r=rEr>r?rA) rrarrHrArBr9r[rr%r r(r)rrrErHrAr,rrrs  zToRGB.__init__NcCsF|||}||j}|durB|jr:tj|d|j|jd}||}|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. NrrI)r[rrHr.rRrArB)rrrTskipr/rrrrs   z ToRGB.forward)Tr:)Nr`rrr,rrasracs(eZdZdZfddZddZZS) ConstantInputzConstant input. Args: num_channel (int): Channel number of constant input. size (int): Spatial size of constant input. cs*tt|ttd||||_dSr])rrdrrr%r r&r')rZ num_channelsizer,rrrszConstantInput.__init__cCs|j|ddd}|Sr])r'repeat)rbatchr/rrrr szConstantInput.forwardr`rrr,rrds rdcsDeZdZdZdfdd Zd d Zd d ZddZdddZZ S)StyleGAN2GeneratorBilinearaStyleGAN2 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. lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01. narrow (float): Narrow ratio for channels. Default: 1.0. r{Gz?r r:c s8tt|||_tg}t|D]} |t||dd|ddq$tj ||_ t d|t d|t d|t d|t d||t d||t d||t d ||t d ||d } | |_ t | d d d|_t| d | d d|dd|d|_t| d |d|d|_t t|d|_|jddd|_|jdd|_t|_t|_t|_| d } t|jD]<} d| dd} dd| | g}|jd| tj|qxtd|jdD]l} | d| }|jt| |d|dd|d|jt||d|dd|d|jt||d|d|} qdS)NTrrr<ri@ ) 48Z16Z32Z64Z128Z256Z512Z1024rqrG)rerFrbF)rHrArr r^rH) rrhrrErrangeappendrr Sequential style_mlpintchannelsrdconstant_inputrZ style_conv1rato_rgb1r"loglog_size num_layers num_latentZ ModuleList style_convsto_rgbsModulenoisesZregister_bufferr r&)rZout_sizerEZnum_mlpZchannel_multiplierZlr_mlpZnarrowrAZstyle_mlp_layersiryrZ layer_idx resolutionrNrr,rrrs              z#StyleGAN2GeneratorBilinear.__init__c Csj|jjj}tjdddd|dg}td|jdD]4}tdD]&}|tjddd|d||dqz6StyleGAN2GeneratorBilinear.forward..Ncsg|]}tjd|qSr^)getattrr)rrr4rrrrr rrFrr)rrtrulenrndimZ unsqueezerfrandomrandintr catrzrNr{r|ziprr)rZstylesZinput_is_latentr^Zrandomize_noiseZ truncationZtruncation_latentZ inject_indexZreturn_latentsZstyle_truncationrTrZlatent1Zlatent2r/rcrconv1conv2Znoise1Znoise2Zto_rgbimagerr4rrsJ   * z"StyleGAN2GeneratorBilinear.forward)rirjrrkr r:)FNTr NNF) rrrr7rrrrrr8rrr,rrhs&R rhcs*eZdZdZdfdd ZddZZS)ScaledLeakyReLUz_Scaled LeakyReLU. Args: negative_slope (float): Negative slope. Default: 0.2. 皙?cstt|||_dSr)rrrnegative_slope)rrr,rrrszScaledLeakyReLU.__init__cCstj||jd}|tdS)N)rr)r.Z leaky_relurr"r#rrr/rrrrszScaledLeakyReLU.forward)rr`rrr,rrsrcs2eZdZdZd fdd ZddZd d ZZS) EqualConv2daEqualized Linear as StyleGAN2. Args: in_channels (int): Channel number of the input. out_channels (int): Channel number of the output. kernel_size (int): Size of the convolving kernel. stride (int): Stride of the convolution. Default: 1 padding (int): Zero-padding added to both sides of the input. Default: 0. bias (bool): If ``True``, adds a learnable bias to the output. Default: ``True``. bias_init_val (float): Bias initialized value. Default: 0. r rTcstt|||_||_||_||_||_dt ||d|_ t t |||||_|rzt t |||_n |dddS)Nr rr)rrrrrr=striderDr"r#r$rr%r r&r'r(r)rr*)rrrr=rrDrr+r,rrrszEqualConv2d.__init__cCs&tj||j|j|j|j|jd}|S)N)rrrD)r.rSr'r$rrrDrrrrrs zEqualConv2d.forwardcCsB|jjd|jd|jd|jd|jd|jd|jdudS)Nr0r1rYz , stride=z , padding=r2r3)r-rrrr=rrDrr4rrrr5 szEqualConv2d.__repr__)r rTrr6rrr,rrs 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. bias (bool): Whether with bias. Default: True. activate (bool): Whether use activateion. Default: True. FTr:c sg}||_|r@|jdkr d|_nd|_|tjjd||jdd} |d|_|t|||| |j|oh| d|r|r|t|n|t dt t |j |dS) Nr;FrLrIr r)rrDrr) rArBrur rZUpsamplerDrrrrrr) rrrr=rKrr\rAZlayersrr,rrrs,  zConvLayer.__init__)FTTr:)rrrr7rr8rrr,rrs rcs*eZdZdZdfdd ZddZZS)ResBlockzResidual block used in StyleGAN2 Discriminator. Args: in_channels (int): Channel number of the input. out_channels (int): Channel number of the output. r:c sVtt|t||dddd|_t||dd|ddd|_t||dd|ddd|_dS)NrFT)rr\)rKrArr\r F)rrrrrrrc)rrrrAr,rrrKs(zResBlock.__init__cCs4||}||}||}||td}|S)Nr)rrrcr"r#)rrr/rcrrrr`s    zResBlock.forward)r:r`rrr,rrCsr)r"rr Zbasicsr.ops.fused_actrrZbasicsr.utils.registryrrZtorch.nnrr.rrrr9rZrardregisterrhrrrvrrrrrrs&   0Z.0@11