a þd0ã@sädZddlZddlmZddlmmZddlZzddlm Z Wne y`ddl m Z Yn0dZ Gdd„dejƒZdd „Zd d „ZGd d „d ejjjƒZGdd„dejjjƒZGdd„dejjjƒZGdd„dejjjƒZdS)z4 File from: https://github.com/mseitzer/pytorch-fid éN)Úload_state_dict_from_url)Úload_urlzjhttps://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pthcsHeZdZdZdZdddddœZefddddf‡fd d „ Zd d „Z‡ZS) Ú InceptionV3z5Pretrained InceptionV3 network returning feature mapséréé)é@éÀééTFc sVtt|ƒ ¡||_||_t|ƒ|_t|ƒ|_|jdks@Jdƒ‚t   ¡|_ |rVt ƒ}n t dd}|j|j|jt jdddg}|j  t j|Ž¡|jdkrÂ|j|jt jdddg}|j  t j|Ž¡|jdkr|j|j|j|j|j|j|j|jg} |j  t j| Ž¡|jdkr<|j|j|j t j!dd g} |j  t j| Ž¡| "¡D] } || _#qDd S) aLBuild pretrained InceptionV3 Parameters ---------- output_blocks : list of int Indices of blocks to return features of. Possible values are: - 0: corresponds to output of first max pooling - 1: corresponds to output of second max pooling - 2: corresponds to output which is fed to aux classifier - 3: corresponds to output of final average pooling resize_input : bool If true, bilinearly resizes input to width and height 299 before feeding input to model. As the network without fully connected layers is fully convolutional, it should be able to handle inputs of arbitrary size, so resizing might not be strictly needed normalize_input : bool If true, scales the input from range (0, 1) to the range the pretrained Inception network expects, namely (-1, 1) requires_grad : bool If true, parameters of the model require gradients. Possibly useful for finetuning the network use_fid_inception : bool If true, uses the pretrained Inception model used in Tensorflow's FID implementation. If false, uses the pretrained Inception model available in torchvision. The FID Inception model has different weights and a slightly different structure from torchvision's Inception model. If you want to compute FID scores, you are strongly advised to set this parameter to true to get comparable results. rz%Last possible output block index is 3T)Ú pretrainedr)Ú kernel_sizeÚstrider)rr)Z output_sizeN)$ÚsuperrÚ__init__Ú resize_inputÚnormalize_inputÚsortedÚ output_blocksÚmaxÚlast_needed_blockÚnnZ ModuleListÚblocksÚfid_inception_v3Ú _inception_v3Z Conv2d_1a_3x3Z Conv2d_2a_3x3Z Conv2d_2b_3x3Z MaxPool2dÚappendZ SequentialZ Conv2d_3b_1x1Z Conv2d_4a_3x3ÚMixed_5bÚMixed_5cÚMixed_5dZMixed_6aÚMixed_6bÚMixed_6cÚMixed_6dÚMixed_6eZMixed_7aÚMixed_7bÚMixed_7cZAdaptiveAvgPool2dÚ parametersÚ requires_grad) Úselfrrrr&Zuse_fid_inceptionÚ inceptionZblock0Zblock1Zblock2Zblock3Úparam©Ú __class__©úc/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/cleanfid/inception_pytorch.pyr$sX#   ÿ   ü  ý ø   ü zInceptionV3.__init__cCs|g}|}|jr(tdƒ‚tj|dddd}|jr:d|d}t|jƒD]2\}}||ƒ}||jvrh| |¡||j krDqxqD|S)agGet Inception feature maps Parameters ---------- inp : torch.autograd.Variable Input tensor of shape Bx3xHxW. Values are expected to be in range (0, 1) Returns ------- List of torch.autograd.Variable, corresponding to the selected output block, sorted ascending by index zshould not resize here)é+r.ZbilinearF)ÚsizeÚmodeZ align_cornersrr) rÚ ValueErrorÚFZ interpolaterÚ enumeraterrrr)r'ÚinpZoutpÚxÚidxÚblockr,r,r-Úforward…s$ ý    zInceptionV3.forward) Ú__name__Ú __module__Ú __qualname__Ú__doc__ZDEFAULT_BLOCK_INDEXZBLOCK_INDEX_BY_DIMrr8Ú __classcell__r,r,r*r-rsüûarcOs\z"ttttj d¡dd…ƒƒ}Wnty8d}Yn0|dkrJd|d<tjj|i|¤ŽS)z¶Wraps `torchvision.models.inception_v3` Skips default weight inititialization if supported by torchvision version. See https://github.com/mseitzer/pytorch-fid/issues/28. Ú.Nr)r)réFZ init_weights) ÚtupleÚmapÚintÚ torchvisionÚ __version__Úsplitr1ÚmodelsZ inception_v3)ÚargsÚkwargsÚversionr,r,r-r©s"  rcCsžtdddd}tddd|_tddd|_td dd|_td d d |_td d d |_td d d |_td dd |_ t dƒ|_ t dƒ|_ ttdd}| |¡|S)a_Build pretrained Inception model for FID computation The Inception model for FID computation uses a different set of weights and has a slightly different structure than torchvision's Inception. This method first constructs torchvision's Inception and then patches the necessary parts that are different in the FID Inception model. iðF)Z num_classesZ aux_logitsr r é )Ú pool_featureséri r é€)Ú channels_7x7é ir )Úprogress)rÚ FIDInceptionArrrÚ FIDInceptionCrr r!r"ÚFIDInceptionE_1r#ÚFIDInceptionE_2r$rÚFID_WEIGHTS_URLZload_state_dict)r(Z state_dictr,r,r-rºs þ    rcs(eZdZdZ‡fdd„Zdd„Z‡ZS)rQz,InceptionA block patched for FID computationcstt|ƒ ||¡dS©N)rrQr)r'Ú in_channelsrKr*r,r-rÕszFIDInceptionA.__init__cCsr| |¡}| |¡}| |¡}| |¡}| |¡}| |¡}tj|ddddd}| |¡}||||g}t   |d¡S©NrrF©r rÚpaddingZcount_include_pad) Ú branch1x1Z branch5x5_1Z branch5x5_2Úbranch3x3dbl_1Úbranch3x3dbl_2Zbranch3x3dbl_3r2Ú avg_pool2dÚ branch_poolÚtorchÚcat)r'r5r[Z branch5x5Ú branch3x3dblr_Úoutputsr,r,r-r8Øs       ÿ  zFIDInceptionA.forward©r9r:r;r<rr8r=r,r,r*r-rQÓs rQcs(eZdZdZ‡fdd„Zdd„Z‡ZS)rRz,InceptionC block patched for FID computationcstt|ƒ ||¡dSrV)rrRr)r'rWrNr*r,r-rîszFIDInceptionC.__init__cCs| |¡}| |¡}| |¡}| |¡}| |¡}| |¡}| |¡}| |¡}| |¡}t j |ddddd}|  |¡}||||g}t   |d¡SrX)r[Z branch7x7_1Z branch7x7_2Z branch7x7_3Zbranch7x7dbl_1Zbranch7x7dbl_2Zbranch7x7dbl_3Zbranch7x7dbl_4Zbranch7x7dbl_5r2r^r_r`ra)r'r5r[Z branch7x7Z branch7x7dblr_rcr,r,r-r8ñs          ÿ  zFIDInceptionC.forwardrdr,r,r*r-rRìs rRcs(eZdZdZ‡fdd„Zdd„Z‡ZS)rSz2First InceptionE block patched for FID computationcstt|ƒ |¡dSrV)rrSr©r'rWr*r,r-r szFIDInceptionE_1.__init__cCsž| |¡}| |¡}| |¡| |¡g}t |d¡}| |¡}| |¡}| |¡|  |¡g}t |d¡}t j |ddddd}|  |¡}||||g}t |d¡S)NrrFrY) r[Ú branch3x3_1Ú branch3x3_2aÚ branch3x3_2br`rar\r]Úbranch3x3dbl_3aÚbranch3x3dbl_3br2r^r_©r'r5r[Z branch3x3rbr_rcr,r,r-r8 s$  þ   þ  ÿ  zFIDInceptionE_1.forwardrdr,r,r*r-rSs rScs(eZdZdZ‡fdd„Zdd„Z‡ZS)rTz3Second InceptionE block patched for FID computationcstt|ƒ |¡dSrV)rrTrrer*r,r-r+szFIDInceptionE_2.__init__cCsœ| |¡}| |¡}| |¡| |¡g}t |d¡}| |¡}| |¡}| |¡|  |¡g}t |d¡}t j |dddd}|  |¡}||||g}t |d¡S)Nrr)r rrZ) r[rfrgrhr`rar\r]rirjr2Z max_pool2dr_rkr,r,r-r8.s   þ   þ   zFIDInceptionE_2.forwardrdr,r,r*r-rT)s rT)r<r`Ztorch.nnrZtorch.nn.functionalZ functionalr2rCZtorchvision.models.utilsrÚ ImportErrorZtorch.utils.model_zoorrUÚModulerrrrFr(Z InceptionArQZ InceptionCrRZ InceptionErSrTr,r,r,r-Ús"  !