import torch import swapae.util as util from swapae.models import MultiGPUModelWrapper from swapae.optimizers.base_optimizer import BaseOptimizer class SwappingAutoencoderOptimizer(BaseOptimizer): """ Class for running the optimization of the model parameters. Implements Generator / Discriminator training, R1 gradient penalty, decaying learning rates, and reporting training progress. """ @staticmethod def modify_commandline_options(parser, is_train): parser.add_argument("--lr", default=0.002, type=float) parser.add_argument("--beta1", default=0.0, type=float) parser.add_argument("--beta2", default=0.99, type=float) parser.add_argument( "--R1_once_every", default=16, type=int, help="lazy R1 regularization. R1 loss is computed " "once in 1/R1_freq times", ) return parser def __init__(self, model: MultiGPUModelWrapper): self.opt = model.opt opt = self.opt self.model = model self.train_mode_counter = 0 self.discriminator_iter_counter = 0 self.Gparams = self.model.get_parameters_for_mode("generator") self.Dparams = self.model.get_parameters_for_mode("discriminator") self.optimizer_G = torch.optim.Adam( self.Gparams, lr=opt.lr, betas=(opt.beta1, opt.beta2) ) # c.f. StyleGAN2 (https://arxiv.org/abs/1912.04958) Appendix B c = opt.R1_once_every / (1 + opt.R1_once_every) self.optimizer_D = torch.optim.Adam( self.Dparams, lr=opt.lr * c, betas=(opt.beta1 ** c, opt.beta2 ** c) ) def set_requires_grad(self, params, requires_grad): """ For more efficient optimization, turn on and off recording of gradients for |params|. """ for p in params: p.requires_grad_(requires_grad) def prepare_images(self, data_i): return data_i["real_A"] def toggle_training_mode(self): modes = ["discriminator", "generator"] self.train_mode_counter = (self.train_mode_counter + 1) % len(modes) return modes[self.train_mode_counter] def train_one_step(self, data_i, total_steps_so_far): images_minibatch = self.prepare_images(data_i) if self.toggle_training_mode() == "generator": losses = self.train_discriminator_one_step(images_minibatch) else: losses = self.train_generator_one_step(images_minibatch) return util.to_numpy(losses) def train_generator_one_step(self, images): self.set_requires_grad(self.Dparams, False) self.set_requires_grad(self.Gparams, True) sp_ma, gl_ma = None, None self.optimizer_G.zero_grad() g_losses, g_metrics = self.model( images, sp_ma, gl_ma, command="compute_generator_losses" ) g_loss = sum([v.mean() for v in g_losses.values()]) g_loss.backward() self.optimizer_G.step() g_losses.update(g_metrics) return g_losses def train_discriminator_one_step(self, images): if self.opt.lambda_GAN == 0.0 and self.opt.lambda_PatchGAN == 0.0: return {} self.set_requires_grad(self.Dparams, True) self.set_requires_grad(self.Gparams, False) self.discriminator_iter_counter += 1 self.optimizer_D.zero_grad() d_losses, d_metrics, sp, gl = self.model( images, command="compute_discriminator_losses" ) self.previous_sp = sp.detach() self.previous_gl = gl.detach() d_loss = sum([v.mean() for v in d_losses.values()]) d_loss.backward() self.optimizer_D.step() needs_R1 = self.opt.lambda_R1 > 0.0 or self.opt.lambda_patch_R1 > 0.0 needs_R1_at_current_iter = needs_R1 and \ self.discriminator_iter_counter % self.opt.R1_once_every == 0 if needs_R1_at_current_iter: self.optimizer_D.zero_grad() r1_losses = self.model(images, command="compute_R1_loss") d_losses.update(r1_losses) r1_loss = sum([v.mean() for v in r1_losses.values()]) r1_loss = r1_loss * self.opt.R1_once_every r1_loss.backward() self.optimizer_D.step() d_losses["D_total"] = sum([v.mean() for v in d_losses.values()]) d_losses.update(d_metrics) return d_losses def get_visuals_for_snapshot(self, data_i): images = self.prepare_images(data_i) with torch.no_grad(): return self.model(images, command="get_visuals_for_snapshot") def save(self, total_steps_so_far): self.model.save(total_steps_so_far)