biuL ddlmZmZddlZddlmZddlmcmZddl m Z ddl m Z ddl mZGdd ejZGd d ejZGd d ejZGddejZGddejZ ddej*dej*dededeeefdej*f dZ ddej*deej*dededej*f dZy))OptionalTupleN) deprecate)is_torch_version)RMSNormc eZdZdZ d dedededeedef fd Zde jd e jfd Z xZ S) Upsample1DaA 1D upsampling layer with an optional convolution. Parameters: channels (`int`): number of channels in the inputs and outputs. use_conv (`bool`, default `False`): option to use a convolution. use_conv_transpose (`bool`, default `False`): option to use a convolution transpose. out_channels (`int`, optional): number of output channels. Defaults to `channels`. name (`str`, default `conv`): name of the upsampling 1D layer. channelsuse_convuse_conv_transpose out_channelsnamec>t|||_|xs||_||_||_||_d|_|r)tj||jddd|_y|r3tj|j|jdd|_yy)Nrr)padding) super__init__r rr rrconvnnConvTranspose1dConv1d)selfr r rrr __class__s V/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/models/upsampling.pyrzUpsample1D.__init__*s   (4H  "4  **8T5F5F1aPDI  $--1B1BAqQDIinputsreturnc|jd|jk(sJ|jr|j|St j |dd}|j r|j|}|S)Nr@nearest scale_factormode)shaper rrF interpolater )rroutputss rforwardzUpsample1D.forward?s`||A$--///  " "99V$ $--SyI ==ii(Gr)FFNr __name__ __module__ __qualname____doc__intboolrstrrtorchTensorr+ __classcell__rs@rr r sp $#(&* RRR! R sm R  R* ell u|| rr ceZdZdZ d dedededeededeef fd Zdd e jd eed e jfd Z xZ S) Upsample2DaA 2D upsampling layer with an optional convolution. Parameters: channels (`int`): number of channels in the inputs and outputs. use_conv (`bool`, default `False`): option to use a convolution. use_conv_transpose (`bool`, default `False`): option to use a convolution transpose. out_channels (`int`, optional): number of output channels. Defaults to `channels`. name (`str`, default `conv`): name of the upsampling 2D layer. r r rrr kernel_sizec t|||_|xs||_||_||_||_| |_|dk(rtj|| | |_ n0|dk(rt|| | |_ n|d|_ ntd|d} |r*|d}tj||j|d|| } n4|r2|d}tj|j|j||| } |d k(r| |_y| |_y) Nln_normrms_normzunknown norm_type: rr)r:striderbiasr)r:rr?r)rrr rr rrr)r LayerNormnormr ValueErrorConvTranspose2dConv2drConv2d_0)rr r rrrr:r norm_typeepselementwise_affiner?r)rrs rrzUpsample2D.__init__\s   (4H  "4 &  ! Xs4FGDI * $#/ABDI  DI29+>? ? " %%$++QX_fjD" 99T]]D,=,=;`gnrsD 6>DI DMr hidden_states output_sizer ctt|dkDs|jddd}tdd||jd|jk(sJ|j 5|j |j ddddj dddd}|jr|j|S|j}|tjk(r+tdd r|jtj}|jdd k\r|j}|j r|dn7t#t%||jd dDcgc] \}}||z  c}}} |j'| zt)dd kDr|j}|t+j |d d}nt+j ||d}|tjk(rtdd r|j|}|j,r3|j.dk(r|j|}|S|j1|}|Scc}}w)NrscalezThe `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`.z1.0.0rrrr)rrrrDrEr rer)rr rr rers rrzFirUpsample2D.__init__sL '3| IIh !TU_`aDM  $(rrIweightkernelfactorgainr ct|tr|dk\sJ|dg|z}tj|tj}|j dk(rtj ||}|tj|z}|||dzzz}|jr|jd}|jd}|jd}|jd|z |dz z } ||f} |jddz |z|z|jddz |z|zf} | d|jddz | dzz |z | d|jddz | dzz |z f} | ddk\r| ddk\sJ|jd|z} tj|| d|||f}tj|ddgjddddd}tj|| |zd||f}tj||| | d }t|tj||j | dzdz|zdz | dzdzf }|S|jd|z } t|tj||j || dzdz|zdz | dzf }|S) a Fused `upsample_2d()` followed by `Conv2d()`. Padding is performed only once at the beginning, not between the operations. The fused op is considerably more efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of arbitrary order. Args: hidden_states (`torch.Tensor`): Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`. weight (`torch.Tensor`, *optional*): Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`. kernel (`torch.Tensor`, *optional*): FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which corresponds to nearest-neighbor upsampling. factor (`int`, *optional*): Integer upsampling factor (default: 2). gain (`float`, *optional*): Scaling factor for signal magnitude (default: 1.0). Returns: output (`torch.Tensor`): Tensor of the shape `[N, C, H * factor, W * factor]` or `[N, H * factor, W * factor, C]`, and same datatype as `hidden_states`. rrUrrrr)dims)r>output_paddingrdevice)paduprs) isinstancer1r4tensorrXndimoutersumr r'reshapefliprTr(conv_transpose2dupfirdn2d_nativerr)rrIrhrirjrkconvHconvWinC pad_valuer> output_shaperp num_groups inverse_convoutputs r _upsample_2dzFirUpsample2D._upsample_2ds@&#&6Q;66 >S6\FfEMM: ;;! [[0F%))F##4619-. ==LLOELLOE,,q/Ca61eai@If%F$$Q'!+v5=$$Q'!+v5=L Q=#6#6q#9A#="JJURQ=#6#6q#9A#="JJURN"!$)nQ.?1.D DD&,,Q/36J]]6JC+NOFZZaV4<!?@1D ??1%*Q. Xy{{Y'>5Qrrr cbtj||jdzdzfdz|j}|j|jd|jd|j jd|j jdg}t j|jd|j}|j j|dddfj|jddd}||||f<tj||d|jdzdzS)Nrrrrrqrn)r>r) r(rsr new_zerosr'rir4arangerrrWexpandr})rrrhindicesris rr+zKUpsample2D.forwardWsvA! 3594==I!! Q Q !!!$ !!!$   ,,v||Av}}E'a077 QRP#)w !!&&DHHqLSTDTUUr)reflect) r-r.r/r0r3rr4r5r+r6r7s@rrrIs4 RR Vell Vu|| VrrceZdZdZ d dededededededd ffd Zd ejdejfd Z xZ S)CogVideoXUpsample3Da A 3D Upsample layer using in CogVideoX by Tsinghua University & ZhipuAI # Todo: Wait for paper release. Args: in_channels (`int`): Number of channels in the input image. out_channels (`int`): Number of channels produced by the convolution. kernel_size (`int`, defaults to `3`): Size of the convolving kernel. stride (`int`, defaults to `1`): Stride of the convolution. padding (`int`, defaults to `1`): Padding added to all four sides of the input. compress_time (`bool`, defaults to `False`): Whether or not to compress the time dimension. in_channelsrr:r>r compress_timer Ncnt|tj||||||_||_y)Nrg)rrrrDrr)rrrr:r>rrrs rrzCogVideoXUpsample3D.__init__zs2 IIk<[Y_ipq *rrc2|jr|jddkDr|jddzdk(rt|dddddf|ddddddf}}tj|d}tj|d}|dddddddddf}t j ||gd}n|jddkDrtj|d}n|j d}tj|d}|dddddddddf}n|j\}}}}}|jdddddj||z|||}tj|d}|j|||g|jddjddddd}|j\}}}}}|jdddddj||z|||}|j|}|j||g|jddjddddd}|S) Nrrrr")r%)dimrr) rr'r(r)r4catsqueezerTr{r) rrx_firstx_restbcthws rr+zCogVideoXUpsample3D.forwards   ||A"v||A':a'?"(Aq/6!Q(3C--cBvC@!!Qa"23GV#4!<a1$vC@*vC@1dAq 01#LLMAq!Q^^Aq!Q2::1q5!QJF]]6?GG1aQRTUVF 1aA1aA.66q1uaAF6"18v||AB'78@@Aq!QO r)rrrF) r-r.r/r0r1r2rr4r5r+r6r7s@rrrgsz,# + + + +  +  + +  +ellu||rrrwrirudownrsr c |x}}|x}}|dx} } |dx} } |j\} }}}|jd||d}|j\} }}}|j\}}|jd|d|d|}tj|ddd|dz ddd|dz g}|jd||z||z|}tj|ddt | dt | dt | dt | dg}|j |j}|ddt | d|jdt | dz t | d|jdt | dz ddf}|jdddd}|jdd||z| z| z||z| z| zg}tj|ddgjdd||}tj||}|jd|||z| z| z|z dz||z| z| z|z dz}|jdddd}|dddd|dd|ddf}||z| z| z|z |zdz}||z| z| z|z |zdz}|jd|||S)Nrrrnrr) r'r{viewr(rsrZrWrrrTr4r|conv2d)rwrirurrsup_xup_ydown_xdown_ypad_x0pad_y0pad_x1pad_y1_channelin_hin_wminorkernel_hkernel_woutrout_hout_ws rr~r~sD4FV!fFV!fFV#\\Awd ^^BdA .F!<<AtT5Hh ++b$4E 2C %%aAtaxAq$(; rs%# 1"//dttnCBIICLV"))V<="))=F! // LL// LL// //  // sCx //  \\ //h&* ,<<, U\\ ", ,  ,  \\ ,r