import torch import swapae.util as util from swapae.models import BaseModel import swapae.models.networks as networks import swapae.models.networks.loss as loss class SwappingAutoencoderModel(BaseModel): @staticmethod def modify_commandline_options(parser, is_train): BaseModel.modify_commandline_options(parser, is_train) parser.add_argument("--spatial_code_ch", default=8, type=int) parser.add_argument("--global_code_ch", default=2048, type=int) parser.add_argument("--lambda_R1", default=10.0, type=float) parser.add_argument("--lambda_patch_R1", default=1.0, type=float) parser.add_argument("--lambda_L1", default=1.0, type=float) parser.add_argument("--lambda_GAN", default=1.0, type=float) parser.add_argument("--lambda_PatchGAN", default=1.0, type=float) parser.add_argument("--patch_min_scale", default=1 / 8, type=float) parser.add_argument("--patch_max_scale", default=1 / 4, type=float) parser.add_argument("--patch_num_crops", default=8, type=int) parser.add_argument("--patch_use_aggregation", type=util.str2bool, default=True) return parser def initialize(self): self.E = networks.create_network(self.opt, self.opt.netE, "encoder") self.G = networks.create_network(self.opt, self.opt.netG, "generator") if self.opt.lambda_GAN > 0.0: self.D = networks.create_network( self.opt, self.opt.netD, "discriminator") if self.opt.lambda_PatchGAN > 0.0: self.Dpatch = networks.create_network( self.opt, self.opt.netPatchD, "patch_discriminator" ) # Count the iteration count of the discriminator # Used for lazy R1 regularization (c.f. Appendix B of StyleGAN2) self.register_buffer( "num_discriminator_iters", torch.zeros(1, dtype=torch.long) ) self.l1_loss = torch.nn.L1Loss() if (not self.opt.isTrain) or self.opt.continue_train: self.load() if self.opt.num_gpus > 0: self.to("cuda:0") def per_gpu_initialize(self): pass def swap(self, x): """ Swaps (or mixes) the ordering of the minibatch """ shape = x.shape assert shape[0] % 2 == 0, "Minibatch size must be a multiple of 2" new_shape = [shape[0] // 2, 2] + list(shape[1:]) x = x.view(*new_shape) x = torch.flip(x, [1]) return x.view(*shape) def compute_image_discriminator_losses(self, real, rec, mix): if self.opt.lambda_GAN == 0.0: return {} pred_real = self.D(real) pred_rec = self.D(rec) pred_mix = self.D(mix) losses = {} losses["D_real"] = loss.gan_loss( pred_real, should_be_classified_as_real=True ) * self.opt.lambda_GAN losses["D_rec"] = loss.gan_loss( pred_rec, should_be_classified_as_real=False ) * (0.5 * self.opt.lambda_GAN) losses["D_mix"] = loss.gan_loss( pred_mix, should_be_classified_as_real=False ) * (0.5 * self.opt.lambda_GAN) return losses def get_random_crops(self, x, crop_window=None): """ Make random crops. Corresponds to the yellow and blue random crops of Figure 2. """ crops = util.apply_random_crop( x, self.opt.patch_size, (self.opt.patch_min_scale, self.opt.patch_max_scale), num_crops=self.opt.patch_num_crops ) return crops def compute_patch_discriminator_losses(self, real, mix): losses = {} real_feat = self.Dpatch.extract_features( self.get_random_crops(real), aggregate=self.opt.patch_use_aggregation ) target_feat = self.Dpatch.extract_features(self.get_random_crops(real)) mix_feat = self.Dpatch.extract_features(self.get_random_crops(mix)) losses["PatchD_real"] = loss.gan_loss( self.Dpatch.discriminate_features(real_feat, target_feat), should_be_classified_as_real=True, ) * self.opt.lambda_PatchGAN losses["PatchD_mix"] = loss.gan_loss( self.Dpatch.discriminate_features(real_feat, mix_feat), should_be_classified_as_real=False, ) * self.opt.lambda_PatchGAN return losses def compute_discriminator_losses(self, real): self.num_discriminator_iters.add_(1) sp, gl = self.E(real) B = real.size(0) assert B % 2 == 0, "Batch size must be even on each GPU." # To save memory, compute the GAN loss on only # half of the reconstructed images rec = self.G(sp[:B // 2], gl[:B // 2]) mix = self.G(self.swap(sp), gl) losses = self.compute_image_discriminator_losses(real, rec, mix) if self.opt.lambda_PatchGAN > 0.0: patch_losses = self.compute_patch_discriminator_losses(real, mix) losses.update(patch_losses) metrics = {} # no metrics to report for the Discriminator iteration return losses, metrics, sp.detach(), gl.detach() def compute_R1_loss(self, real): losses = {} if self.opt.lambda_R1 > 0.0: real.requires_grad_() pred_real = self.D(real).sum() grad_real, = torch.autograd.grad( outputs=pred_real, inputs=[real], create_graph=True, retain_graph=True, ) grad_real2 = grad_real.pow(2) dims = list(range(1, grad_real2.ndim)) grad_penalty = grad_real2.sum(dims) * (self.opt.lambda_R1 * 0.5) else: grad_penalty = 0.0 if self.opt.lambda_patch_R1 > 0.0: real_crop = self.get_random_crops(real).detach() real_crop.requires_grad_() target_crop = self.get_random_crops(real).detach() target_crop.requires_grad_() real_feat = self.Dpatch.extract_features( real_crop, aggregate=self.opt.patch_use_aggregation) target_feat = self.Dpatch.extract_features(target_crop) pred_real_patch = self.Dpatch.discriminate_features( real_feat, target_feat ).sum() grad_real, grad_target = torch.autograd.grad( outputs=pred_real_patch, inputs=[real_crop, target_crop], create_graph=True, retain_graph=True, ) dims = list(range(1, grad_real.ndim)) grad_crop_penalty = grad_real.pow(2).sum(dims) + \ grad_target.pow(2).sum(dims) grad_crop_penalty *= (0.5 * self.opt.lambda_patch_R1 * 0.5) else: grad_crop_penalty = 0.0 losses["D_R1"] = grad_penalty + grad_crop_penalty return losses def compute_generator_losses(self, real, sp_ma, gl_ma): losses, metrics = {}, {} B = real.size(0) sp, gl = self.E(real) rec = self.G(sp[:B // 2], gl[:B // 2]) # only on B//2 to save memory sp_mix = self.swap(sp) if self.opt.crop_size >= 1024: # another momery-saving trick: reduce #outputs to save memory real = real[B // 2:] gl = gl[B // 2:] sp_mix = sp_mix[B // 2:] mix = self.G(sp_mix, gl) # record the error of the reconstructed images for monitoring purposes metrics["L1_dist"] = self.l1_loss(rec, real[:B // 2]) if self.opt.lambda_L1 > 0.0: losses["G_L1"] = metrics["L1_dist"] * self.opt.lambda_L1 if self.opt.lambda_GAN > 0.0: losses["G_GAN_rec"] = loss.gan_loss( self.D(rec), should_be_classified_as_real=True ) * (self.opt.lambda_GAN * 0.5) losses["G_GAN_mix"] = loss.gan_loss( self.D(mix), should_be_classified_as_real=True ) * (self.opt.lambda_GAN * 1.0) if self.opt.lambda_PatchGAN > 0.0: real_feat = self.Dpatch.extract_features( self.get_random_crops(real), aggregate=self.opt.patch_use_aggregation).detach() mix_feat = self.Dpatch.extract_features(self.get_random_crops(mix)) losses["G_mix"] = loss.gan_loss( self.Dpatch.discriminate_features(real_feat, mix_feat), should_be_classified_as_real=True, ) * self.opt.lambda_PatchGAN return losses, metrics def get_visuals_for_snapshot(self, real): if self.opt.isTrain: # avoid the overhead of generating too many visuals during training real = real[:2] if self.opt.num_gpus > 1 else real[:4] sp, gl = self.E(real) layout = util.resize2d_tensor(util.visualize_spatial_code(sp), real) rec = self.G(sp, gl) mix = self.G(sp, self.swap(gl)) visuals = {"real": real, "layout": layout, "rec": rec, "mix": mix} return visuals def fix_noise(self, sample_image=None): """ The generator architecture is stochastic because of the noise input at each layer (StyleGAN2 architecture). It could lead to flickering of the outputs even when identical inputs are given. Prevent flickering by fixing the noise injection of the generator. """ if sample_image is not None: # The generator should be run at least once, # so that the noise dimensions could be computed sp, gl = self.E(sample_image) self.G(sp, gl) noise_var = self.G.fix_and_gather_noise_parameters() return noise_var def encode(self, image, extract_features=False): return self.E(image, extract_features=extract_features) def decode(self, spatial_code, global_code): return self.G(spatial_code, global_code) def get_parameters_for_mode(self, mode): if mode == "generator": return list(self.G.parameters()) + list(self.E.parameters()) elif mode == "discriminator": Dparams = [] if self.opt.lambda_GAN > 0.0: Dparams += list(self.D.parameters()) if self.opt.lambda_PatchGAN > 0.0: Dparams += list(self.Dpatch.parameters()) return Dparams