From b4f00f4378dc06be5c77b5d8ad01ceb6f67b6d45 Mon Sep 17 00:00:00 2001 From: bishe <123456789@163.com> Date: Sun, 23 Feb 2025 22:40:36 +0800 Subject: [PATCH] debug --- checkpoints/ROMA_UNSB_001/loss_log.txt | 2 + .../roma_unsb_model.cpython-39.pyc | Bin 19227 -> 18663 bytes models/roma_unsb_model.py | 57 +++++++++--------- 3 files changed, 31 insertions(+), 28 deletions(-) diff --git a/checkpoints/ROMA_UNSB_001/loss_log.txt b/checkpoints/ROMA_UNSB_001/loss_log.txt index 2bc6f9f..d27cb3a 100644 --- a/checkpoints/ROMA_UNSB_001/loss_log.txt +++ b/checkpoints/ROMA_UNSB_001/loss_log.txt @@ -44,3 +44,5 @@ ================ Training Loss (Sun Feb 23 22:29:52 2025) ================ ================ Training Loss (Sun Feb 23 22:30:40 2025) ================ ================ Training Loss (Sun Feb 23 22:33:48 2025) ================ +================ Training Loss (Sun Feb 23 22:39:16 2025) ================ +================ Training Loss (Sun Feb 23 22:39:48 2025) ================ diff --git a/models/__pycache__/roma_unsb_model.cpython-39.pyc b/models/__pycache__/roma_unsb_model.cpython-39.pyc index e819f4454b60684604d42e7a80e4749c23d9717d..bd981a861591db2901c5931be472be7dc126039d 100644 GIT binary patch delta 1137 zcmZvbO>7fK6vt=m_1BUQs1qknV#gqovO+0wDI$h^Kp-J0MFPc##t2p$uL&m3+L|?` z1?PfVfmEpiJt-F|HEEAk53NNn6&yGfLHYiWmLdTH@eM*q94fWpy;&qKwKPA!_vX#) zoBzByG0Ptw=XFMHt;+DXa`hjR%ZtC(9q3$?@n(&p*p0zV^;l;#z^}pOc({71%kZ{6 z#;DjzSWLK&&;sHUvLC>@hRm5rdWz8eM85(0+ zeRX_oEpI%F32BWx>SV60my1(5%hm?q+n$}=frmXi_&rFaR`FYqPaX3Ajk)4BJW1V1 zXi@jOFEHfrbVkWrSzEjz&og0*f8gZ$t*s8_-y*mx$?|f9!eL1y;{Oh9`)kZCl?g9q|`VYKkn&QT7;9 zImgW;Q!Js+cF|Wi@AI!O8G>}YEj5*rzD7+rj(sb#WK;dJZE-0fGjOhu}(6ePrT)`Plb(=?} zrLv9>$Z8>k35|pZBI6rM>z|Lyiv~X0tX{_D5f8S_M{ssim)Gy>hTZ+qYGm_a9&V;p zblCg`%t&iq@f?00`mo(caX%qI2oj_OZDhsZ)lkC!5|;@MO~W0HQl|t>$<}on7Iq9b zrv7iEI3{LlWFOikqntMiVi|>0_busJt1Q~#hvDye|4+SJIB$l&k#@cmc8yrP1I~|Z z^TjZI0^*~E)qheER|$6r^0~{`FitMDtDr6uB!cLq3S)3;v?pE))&R8IH7Ri04fqFl=QGlThtlWbiS=T2E->Dz@*1~0mH6`chOCU5#m%{-kxy5#xDofj)vNiY$}d?J5)cXj;77N{@_{1B3(|OA*qQ*zIOh`jB+%_}5*I9>o#K4wV}2T} ztAU$>)0VF&TES)p=$S%2;B=t@^s)RzJr6d^nFdau0(B1@eD=a+e3#Jm^nnyC|0jJ$ zaXlwQ{H4vOy(IJ{)HfF)hI=8cOL}f8t!lt|QZy~=X=#DVE9WC;*mlp0WRC@UpC#Y3 zrQQ?dHrwod&2a~O>0K7=yR~76@~WN#bA_Hk>GKGmBTOPpAbbH}xXwwboT{f)opu19 ztXh-lN?MU1Usf8C)Lceej~TuL={b2REsaJE|6yK=C2J2!Hlk7tLTH&+)5+>Xn&`jc zhUHZ~la-Dr&5DyXrt%;i#5(9i&>W6OcyltBv}M$kMt?`nGOy9UxL7yr3Tx;e?cM{H z-bOGRgOq~+I02Gg9vnNA!Q#tN;Ulb~rpEC$(AstOZvVyV8f%l)2VXz&TdOP%;vgR4 zVRZvL;?EE+;-L@Oqk+ms3wnQw3Hkt%W!apjN#hcxpm$j7V32&xh6jUzIPAM0g9~6oM6845!(NnBmIi7PQsn^eZd$W7P7*Ye->Y`VG4=6t(XF z>1KZq4f)D|a>C(TvsaoKj*<;qouG)7LXlJnJ%UA$%3&BSi}i|a$#$90mtlxusuyvK zf<@Q~&NTwpRJ7Cz1y|msxccC!+7ugb=-YhY`+?p9ybVt41xvxLw{1NEbbH>7aikE? zzP!M0o~#aps)Sl0Z+h2rT9Cq^4>J<)n*9ZCybp(e<|wW`oks8=@Wyvg5nFi$1#cqY zNSIy2pAy4HWmTa9VTE|WVj=8myfgY=cd@6m0UPU{m-Wacx1zk zybGZWAnDG^%kzpXrF50=3-zL^9+MvfNQ&l*f`xEqV-k)5{TAerGIMY%(3)R0fK(fI zMrcH60!Z2?=A;+S8s-Qs9M9(FUDn$eF=&`rotWr3NEv1=opqL289-x7sc z9a}g{3m933+H1i(v;&_=@q1nxLq(W?ati|ygxKSIsKNd>y zg*t{RzEzKRDN@d>GelL&%Ij&WnJ?!^;4nBQJ$y5?67S^6Z3+~#geDf|gyOXJtc z!299^AuY_6XeJRBNvNcgT}_;}hrnE74-(n#8+fT*gnJ15f8-%jXAy9tC@ulT&d}rF qMni|S(I>ja4zZo8ppFC$ySlo(vQBxW_e*ytMe+|1Im{*i diff --git a/models/roma_unsb_model.py b/models/roma_unsb_model.py index ec2fe8d..2c275a7 100644 --- a/models/roma_unsb_model.py +++ b/models/roma_unsb_model.py @@ -284,10 +284,10 @@ class RomaUnsbModel(BaseModel): for nce_layer in self.nce_layers: self.criterionNCE.append(PatchNCELoss(opt).to(self.device)) self.criterionIdt = torch.nn.L1Loss().to(self.device) - self.optimizer_G1 = torch.optim.Adam(self.netG.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2)) - self.optimizer_D1 = torch.optim.Adam(self.netD.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2)) - self.optimizer_E1 = torch.optim.Adam(self.netE.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2)) - self.optimizers = [self.optimizer_G1, self.optimizer_D1, self.optimizer_E1] + self.optimizer_G = torch.optim.Adam(self.netG.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2)) + self.optimizer_D = torch.optim.Adam(self.netD.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2)) + self.optimizer_E = torch.optim.Adam(self.netE.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2)) + self.optimizers = [self.optimizer_G, self.optimizer_D, self.optimizer_E] self.cao = ContentAwareOptimization(opt.lambda_inc, opt.eta_ratio) #损失函数 self.ctn = ContentAwareTemporalNorm() #生成的伪光流 @@ -483,28 +483,28 @@ class RomaUnsbModel(BaseModel): # [[1,576,768],[1,576,768],[1,576,768]] # [3,576,768] - # 生成图像的梯度 - fake_gradient = torch.autograd.grad(self.mutil_fake_B0_tokens.sum(), self.mutil_fake_B0_tokens, create_graph=True)[0] + ## 生成图像的梯度 + #fake_gradient = torch.autograd.grad(self.mutil_fake_B0_tokens.sum(), self.mutil_fake_B0_tokens, create_graph=True)[0] + # + ## 梯度图 + #self.weight_fake = self.cao.generate_weight_map(fake_gradient) + # + ## 生成图像的CTN光流图 + #self.f_content = self.ctn(self.weight_fake) + # + ## 变换后的图片 + #self.warped_real_A_noisy2 = warp(self.real_A_noisy, self.f_content) + #self.warped_fake_B0 = warp(self.fake_B0,self.f_content) + # + ## 经过第二次生成器 + #self.warped_fake_B0_2 = self.netG(self.warped_real_A_noisy2, self.time, z_in) - # 梯度图 - self.weight_fake = self.cao.generate_weight_map(fake_gradient) - - # 生成图像的CTN光流图 - self.f_content = self.ctn(self.weight_fake) - - # 变换后的图片 - self.warped_real_A_noisy2 = warp(self.real_A_noisy, self.f_content) - self.warped_fake_B0 = warp(self.fake_B0,self.f_content) - - # 经过第二次生成器 - self.warped_fake_B0_2 = self.netG(self.warped_real_A_noisy2, self.time, z_in) - - warped_fake_B0_2=self.warped_fake_B0_2 - warped_fake_B0=self.warped_fake_B0 - self.warped_fake_B0_2_resize = self.resize(warped_fake_B0_2) - self.warped_fake_B0_resize = self.resize(warped_fake_B0) - self.mutil_warped_fake_B0_tokens = self.netPreViT(self.warped_fake_B0_resize, self.atten_layers, get_tokens=True) - self.mutil_fake_B0_2_tokens = self.netPreViT(self.warped_fake_B0_2_resize, self.atten_layers, get_tokens=True) + #warped_fake_B0_2=self.warped_fake_B0_2 + #warped_fake_B0=self.warped_fake_B0 + #self.warped_fake_B0_2_resize = self.resize(warped_fake_B0_2) + #self.warped_fake_B0_resize = self.resize(warped_fake_B0) + #self.mutil_warped_fake_B0_tokens = self.netPreViT(self.warped_fake_B0_resize, self.atten_layers, get_tokens=True) + #self.mutil_fake_B0_2_tokens = self.netPreViT(self.warped_fake_B0_2_resize, self.atten_layers, get_tokens=True) def compute_D_loss(self): @@ -563,9 +563,10 @@ class RomaUnsbModel(BaseModel): else: loss_global = 0.0 - if self.opt.lambda_ctn > 0.0: - wapped_fake_B = warp(self.fake_B, self.f_content) # use updated self.f_content - self.l2_loss = F.mse_loss(self.fake_B_2, wapped_fake_B) # complete the loss calculation + self.l2_loss = 0.0 + #if self.opt.lambda_ctn > 0.0: + # wapped_fake_B = warp(self.fake_B, self.f_content) # use updated self.f_content + # self.l2_loss = F.mse_loss(self.fake_B_2, wapped_fake_B) # complete the loss calculation self.loss_G = self.loss_G_GAN + self.opt.lambda_SB * self.loss_SB + self.opt.lambda_ctn * self.l2_loss + loss_global * self.opt.lambda_global return self.loss_G