a  þdö-ã@s”ddlZddlmZddlmZddlmZddlmZddl m Z ddl m Z ddl mZmZdd lmZd d lmZe ¡Gd d „d eƒƒZdS)éN)Ú OrderedDict)Úpath)Útqdm)Ú build_network)Ú build_loss)Úcalculate_metric)ÚimwriteÚ tensor2img)ÚMODEL_REGISTRYé)ÚSRModelcs^eZdZdZdd„Zdd„Zdd„Zedd „ƒZd d „Z d‡fd d„ Z dd„Z dd„Z ‡Z S)ÚHiFaceGANModelz§HiFaceGAN model for generic-purpose face restoration. No prior modeling required, works for any degradations. Currently doesn't support EMA for inference. cCsD|jd}| dd¡|_|jdkr*tdƒ‚|j ¡t|jdƒ|_| |j¡|_|  |j¡| d¡r€t |dƒ  |j ¡|_ nd|_ | d¡r¨t |dƒ  |j ¡|_nd|_| d¡rÐt |dƒ  |j ¡|_nd|_|j durò|jduròtd ƒ‚| d ¡rt |d ƒ  |j ¡|_| d d ¡|_| d d¡|_| ¡| ¡dS)NÚtrainÚ ema_decayrz(HiFaceGAN does not support EMA now. PassZ network_dZ pixel_optZperceptual_optZfeature_matching_optz*Both pixel and perceptual losses are None.Zgan_optÚ net_d_itersr Únet_d_init_iters)ÚoptÚgetrÚNotImplementedErrorÚnet_grrÚnet_dZmodel_to_deviceZ print_networkrÚtoZdeviceÚcri_pixÚcri_perceptualÚcri_featÚ ValueErrorÚcri_ganrrÚsetup_optimizersZsetup_schedulers)ÚselfÚ train_opt©r úg/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/basicsr/models/hifacegan_model.pyÚinit_training_settingss2        z%HiFaceGANModel.init_training_settingscCs†|jd}|d d¡}|j||j ¡fi|d¤Ž|_|j |j¡|d d¡}|j||j ¡fi|d¤Ž|_ |j |j ¡dS)NrZoptim_gÚtypeZoptim_d) rÚpopZ get_optimizerrÚ parametersÚ optimizer_gZ optimizersÚappendrÚ optimizer_d)rrZ optim_typer r r!r?s   zHiFaceGANModel.setup_optimizerscCsÒ|jdd…\}}|jdd…|jdd…kr|tjj |||f¡}tjj |||f¡}tj||gdd}tj||gdd} n$tj||gdd}tj||gdd} tj|| gdd} | | ¡} | | ¡\} } | | fS)a+ This is a conditional (on the input) discriminator In Batch Normalization, the fake and real images are recommended to be in the same batch to avoid disparate statistics in fake and real images. So both fake and real images are fed to D all at once. éþÿÿÿNr ©Zdimr)ÚshapeÚtorchÚnnZ functionalZ interpolateÚcatrÚ _divide_pred)rZinput_lqÚoutputZ ground_truthÚhÚwÚlqÚrealZ fake_concatZ real_concatZ fake_and_realZdiscriminator_outÚ pred_fakeÚ pred_realr r r!Ú discriminateJs zHiFaceGANModel.discriminatecCs|t|ƒtkrHg}g}|D],}| dd„|Dƒ¡| dd„|Dƒ¡qn,|d| d¡d…}|| d¡dd…}||fS)zÁ Take the prediction of fake and real images from the combined batch. The prediction contains the intermediate outputs of multiscale GAN, so it's usually a list cSs"g|]}|d| d¡d…‘qS)Nré©Úsize©Ú.0Ztensorr r r!Ú lóz/HiFaceGANModel._divide_pred..cSs"g|]}|| d¡dd…‘qS)rr8Nr9r;r r r!r=mr>Nrr8)r#Úlistr'r:)ÚpredZfaker4Úpr r r!r/as zHiFaceGANModel._divide_predcCsâ|j ¡D] }d|_q |j ¡| |j¡|_d}tƒ}||j dkr2||j kr2|j rz|  |j|j ¡}||7}||d<|j rÄ|  |j|j ¡\}}|dur¬||7}||d<|durÄ||7}||d<| |j|j|j ¡\}} |j|ddd} || 7}| |d<|jr | || ¡} || 7}| |d <| ¡|j ¡|j ¡D] }d|_q<|j ¡| |j|j ¡|j ¡\}} |j| ddd} | |d <|j|ddd} | |d <| | d }| ¡|j ¡| |¡|_|jdkrÞtd ƒdS)NFrÚl_g_pixÚ l_g_percepÚ l_g_styleT)Zis_discÚl_g_ganÚl_g_featÚl_d_realÚl_d_faker8z(HiFaceGAN does not support EMA now. pass)rr%Z requires_gradr&Z zero_gradrr3r0rrrrÚgtrr7rrZbackwardÚstepr(ÚdetachZreduce_loss_dictZlog_dictrÚprint)rÚ current_iterrAZ l_g_totalZ loss_dictrBrCrDr5r6rErFrGrHZ l_d_totalr r r!Úoptimize_parameterstsV         z"HiFaceGANModel.optimize_parametersFcsV|jdddvr|j ¡|jdr8| ||||¡ntdƒtƒ ||||¡dS)a² Warning: HiFaceGAN requires train() mode even for validation For more info, see https://github.com/Lotayou/Face-Renovation/issues/31 Args: dataloader (torch.utils.data.DataLoader): Validation dataloader. current_iter (int): Current iteration. tb_logger (tensorboard logger): Tensorboard logger. save_img (bool): Whether to save images. Default: False. Z network_gr#)Z HiFaceGANZSPADEGeneratorÚdistzwIn HiFaceGANModel: The new metrics package is under development.Using super method now (Only PSNR & SSIM are supported)N)rrrZdist_validationrLÚsuperÚnondist_validation)rÚ dataloaderrMÚ tb_loggerÚsave_img©Ú __class__r r!Ú validationÂs   zHiFaceGANModel.validationc Csþ|jjd}|jd d¡du}|r4tƒ|_g}g}tt|ƒdd} |D]@} t t  | dd¡¡d} |  | ¡|  ¡|  ¡} |  | d ¡d | vr¬|  | d ¡|`|`|`tj ¡|rp|jd rôt |jd d | | ›d|›d¡} nj|jddr4t |jd d || ›d|jdd›d¡} n*t |jd d || ›d|jd›d¡} tt| d ƒ| ƒ|  d¡|  d| ›¡qH|  ¡|rútj|dd}tj|dd}|jdd ¡D]"\}}tt||d|ƒ|j|<qÈ| |||¡dS)a¥ TODO: Validation using updated metric system The metrics are now evaluated after all images have been tested This allows batch processing, and also allows evaluation of distributional metrics, such as: @ Frechet Inception Distance: FID @ Maximum Mean Discrepancy: MMD Warning: Need careful batch management for different inference settings. ÚnameÚvalZmetricsNÚimage)ÚtotalÚunitZlq_pathrÚresultrIZis_trainrZ visualizationÚ_z.pngÚsuffixr zTest r*)Úsr_packÚgt_pack)ZdatasetrrÚdictZmetric_resultsrÚlenÚospÚsplitextÚbasenameZ feed_dataÚtestZget_current_visualsr'rIr3r0r,ZcudaZ empty_cacheÚjoinrr ÚupdateÚset_descriptionÚcloser.ÚitemsrZ_log_validation_metric_values)rrRrMrSrTZ dataset_nameZ with_metricsZ sr_tensorsZ gt_tensorsZpbarZval_dataZimg_nameZvisualsZ save_img_pathr`rarXZopt_r r r!rQØsR     ÿÿÿ z!HiFaceGANModel.nondist_validationcCsBt|dƒrtdƒ| |jd|¡| |jd|¡| ||¡dS)NZ net_g_emazs