basicsr/models/stylegan2_model.py

import cv2
import math
import numpy as np
import random
import torch
from collections import OrderedDict
from os import path as osp

from basicsr.archs import build_network
from basicsr.losses import build_loss
from basicsr.losses.losses import g_path_regularize, r1_penalty
from basicsr.utils import imwrite, tensor2img
from basicsr.utils.registry import MODEL_REGISTRY
from .base_model import BaseModel


@MODEL_REGISTRY.register()
class StyleGAN2Model(BaseModel):
    """StyleGAN2 model."""

    def __init__(self, opt):
        super(StyleGAN2Model, self).__init__(opt)

        # define network net_g
        self.net_g = build_network(opt['network_g'])
        self.net_g = self.model_to_device(self.net_g)
        self.print_network(self.net_g)
        # load pretrained model
        load_path = self.opt['path'].get('pretrain_network_g', None)
        if load_path is not None:
            param_key = self.opt['path'].get('param_key_g', 'params')
            self.load_network(self.net_g, load_path, self.opt['path'].get('strict_load_g', True), param_key)

        # latent dimension: self.num_style_feat
        self.num_style_feat = opt['network_g']['num_style_feat']
        num_val_samples = self.opt['val'].get('num_val_samples', 16)
        self.fixed_sample = torch.randn(num_val_samples, self.num_style_feat, device=self.device)

        if self.is_train:
            self.init_training_settings()

    def init_training_settings(self):
        train_opt = self.opt['train']

        # define network net_d
        self.net_d = build_network(self.opt['network_d'])
        self.net_d = self.model_to_device(self.net_d)
        self.print_network(self.net_d)

        # load pretrained model
        load_path = self.opt['path'].get('pretrain_network_d', None)
        if load_path is not None:
            self.load_network(self.net_d, load_path, self.opt['path'].get('strict_load_d', True))

        # define network net_g with Exponential Moving Average (EMA)
        # net_g_ema only used for testing on one GPU and saving, do not need to
        # wrap with DistributedDataParallel
        self.net_g_ema = build_network(self.opt['network_g']).to(self.device)
        # load pretrained model
        load_path = self.opt['path'].get('pretrain_network_g', None)
        if load_path is not None:
            self.load_network(self.net_g_ema, load_path, self.opt['path'].get('strict_load_g', True), 'params_ema')
        else:
            self.model_ema(0)  # copy net_g weight

        self.net_g.train()
        self.net_d.train()
        self.net_g_ema.eval()

        # define losses
        # gan loss (wgan)
        self.cri_gan = build_loss(train_opt['gan_opt']).to(self.device)
        # regularization weights
        self.r1_reg_weight = train_opt['r1_reg_weight']  # for discriminator
        self.path_reg_weight = train_opt['path_reg_weight']  # for generator

        self.net_g_reg_every = train_opt['net_g_reg_every']
        self.net_d_reg_every = train_opt['net_d_reg_every']
        self.mixing_prob = train_opt['mixing_prob']

        self.mean_path_length = 0

        # set up optimizers and schedulers
        self.setup_optimizers()
        self.setup_schedulers()

    def setup_optimizers(self):
        train_opt = self.opt['train']
        # optimizer g
        net_g_reg_ratio = self.net_g_reg_every / (self.net_g_reg_every + 1)
        if self.opt['network_g']['type'] == 'StyleGAN2GeneratorC':
            normal_params = []
            style_mlp_params = []
            modulation_conv_params = []
            for name, param in self.net_g.named_parameters():
                if 'modulation' in name:
                    normal_params.append(param)
                elif 'style_mlp' in name:
                    style_mlp_params.append(param)
                elif 'modulated_conv' in name:
                    modulation_conv_params.append(param)
                else:
                    normal_params.append(param)
            optim_params_g = [
                {  # add normal params first
                    'params': normal_params,
                    'lr': train_opt['optim_g']['lr']
                },
                {
                    'params': style_mlp_params,
                    'lr': train_opt['optim_g']['lr'] * 0.01
                },
                {
                    'params': modulation_conv_params,
                    'lr': train_opt['optim_g']['lr'] / 3
                }
            ]
        else:
            normal_params = []
            for name, param in self.net_g.named_parameters():
                normal_params.append(param)
            optim_params_g = [{  # add normal params first
                'params': normal_params,
                'lr': train_opt['optim_g']['lr']
            }]

        optim_type = train_opt['optim_g'].pop('type')
        lr = train_opt['optim_g']['lr'] * net_g_reg_ratio
        betas = (0**net_g_reg_ratio, 0.99**net_g_reg_ratio)
        self.optimizer_g = self.get_optimizer(optim_type, optim_params_g, lr, betas=betas)
        self.optimizers.append(self.optimizer_g)

        # optimizer d
        net_d_reg_ratio = self.net_d_reg_every / (self.net_d_reg_every + 1)
        if self.opt['network_d']['type'] == 'StyleGAN2DiscriminatorC':
            normal_params = []
            linear_params = []
            for name, param in self.net_d.named_parameters():
                if 'final_linear' in name:
                    linear_params.append(param)
                else:
                    normal_params.append(param)
            optim_params_d = [
                {  # add normal params first
                    'params': normal_params,
                    'lr': train_opt['optim_d']['lr']
                },
                {
                    'params': linear_params,
                    'lr': train_opt['optim_d']['lr'] * (1 / math.sqrt(512))
                }
            ]
        else:
            normal_params = []
            for name, param in self.net_d.named_parameters():
                normal_params.append(param)
            optim_params_d = [{  # add normal params first
                'params': normal_params,
                'lr': train_opt['optim_d']['lr']
            }]

        optim_type = train_opt['optim_d'].pop('type')
        lr = train_opt['optim_d']['lr'] * net_d_reg_ratio
        betas = (0**net_d_reg_ratio, 0.99**net_d_reg_ratio)
        self.optimizer_d = self.get_optimizer(optim_type, optim_params_d, lr, betas=betas)
        self.optimizers.append(self.optimizer_d)

    def feed_data(self, data):
        self.real_img = data['gt'].to(self.device)

    def make_noise(self, batch, num_noise):
        if num_noise == 1:
            noises = torch.randn(batch, self.num_style_feat, device=self.device)
        else:
            noises = torch.randn(num_noise, batch, self.num_style_feat, device=self.device).unbind(0)
        return noises

    def mixing_noise(self, batch, prob):
        if random.random() < prob:
            return self.make_noise(batch, 2)
        else:
            return [self.make_noise(batch, 1)]

    def optimize_parameters(self, current_iter):
        loss_dict = OrderedDict()

        # optimize net_d
        for p in self.net_d.parameters():
            p.requires_grad = True
        self.optimizer_d.zero_grad()

        batch = self.real_img.size(0)
        noise = self.mixing_noise(batch, self.mixing_prob)
        fake_img, _ = self.net_g(noise)
        fake_pred = self.net_d(fake_img.detach())

        real_pred = self.net_d(self.real_img)
        # wgan loss with softplus (logistic loss) for discriminator
        l_d = self.cri_gan(real_pred, True, is_disc=True) + self.cri_gan(fake_pred, False, is_disc=True)
        loss_dict['l_d'] = l_d
        # In wgan, real_score should be positive and fake_score should be
        # negative
        loss_dict['real_score'] = real_pred.detach().mean()
        loss_dict['fake_score'] = fake_pred.detach().mean()
        l_d.backward()

        if current_iter % self.net_d_reg_every == 0:
            self.real_img.requires_grad = True
            real_pred = self.net_d(self.real_img)
            l_d_r1 = r1_penalty(real_pred, self.real_img)
            l_d_r1 = (self.r1_reg_weight / 2 * l_d_r1 * self.net_d_reg_every + 0 * real_pred[0])
            # TODO: why do we need to add 0 * real_pred, otherwise, a runtime
            # error will arise: RuntimeError: Expected to have finished
            # reduction in the prior iteration before starting a new one.
            # This error indicates that your module has parameters that were
            # not used in producing loss.
            loss_dict['l_d_r1'] = l_d_r1.detach().mean()
            l_d_r1.backward()

        self.optimizer_d.step()

        # optimize net_g
        for p in self.net_d.parameters():
            p.requires_grad = False
        self.optimizer_g.zero_grad()

        noise = self.mixing_noise(batch, self.mixing_prob)
        fake_img, _ = self.net_g(noise)
        fake_pred = self.net_d(fake_img)

        # wgan loss with softplus (non-saturating loss) for generator
        l_g = self.cri_gan(fake_pred, True, is_disc=False)
        loss_dict['l_g'] = l_g
        l_g.backward()

        if current_iter % self.net_g_reg_every == 0:
            path_batch_size = max(1, batch // self.opt['train']['path_batch_shrink'])
            noise = self.mixing_noise(path_batch_size, self.mixing_prob)
            fake_img, latents = self.net_g(noise, return_latents=True)
            l_g_path, path_lengths, self.mean_path_length = g_path_regularize(fake_img, latents, self.mean_path_length)

            l_g_path = (self.path_reg_weight * self.net_g_reg_every * l_g_path + 0 * fake_img[0, 0, 0, 0])
            # TODO:  why do we need to add 0 * fake_img[0, 0, 0, 0]
            l_g_path.backward()
            loss_dict['l_g_path'] = l_g_path.detach().mean()
            loss_dict['path_length'] = path_lengths

        self.optimizer_g.step()

        self.log_dict = self.reduce_loss_dict(loss_dict)

        # EMA
        self.model_ema(decay=0.5**(32 / (10 * 1000)))

    def test(self):
        with torch.no_grad():
            self.net_g_ema.eval()
            self.output, _ = self.net_g_ema([self.fixed_sample])

    def dist_validation(self, dataloader, current_iter, tb_logger, save_img):
        if self.opt['rank'] == 0:
            self.nondist_validation(dataloader, current_iter, tb_logger, save_img)

    def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):
        assert dataloader is None, 'Validation dataloader should be None.'
        self.test()
        result = tensor2img(self.output, min_max=(-1, 1))
        if self.opt['is_train']:
            save_img_path = osp.join(self.opt['path']['visualization'], 'train', f'train_{current_iter}.png')
        else:
            save_img_path = osp.join(self.opt['path']['visualization'], 'test', f'test_{self.opt["name"]}.png')
        imwrite(result, save_img_path)
        # add sample images to tb_logger
        result = (result / 255.).astype(np.float32)
        result = cv2.cvtColor(result, cv2.COLOR_BGR2RGB)
        if tb_logger is not None:
            tb_logger.add_image('samples', result, global_step=current_iter, dataformats='HWC')

    def save(self, epoch, current_iter):
        self.save_network([self.net_g, self.net_g_ema], 'net_g', current_iter, param_key=['params', 'params_ema'])
        self.save_network(self.net_d, 'net_d', current_iter)
        self.save_training_state(epoch, current_iter)