ai-art

  1import torch
  2import torch.nn as nn
  3
  4from core.taming.utils import hinge_d_loss, vanilla_d_loss, adopt_weight, weights_init
  5
  6from core.taming.modules.discriminator import NLayerDiscriminator
  7
  8from core.taming.modules.losses import LPIPS
  9
 10
 11class DummyLoss(nn.Module):
 12    def __init__(self):
 13        super().__init__()
 14
 15
 16class VQLPIPSWithDiscriminator(nn.Module):
 17    def __init__(self, disc_start, codebook_weight=1.0, pixelloss_weight=1.0,
 18                 disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0,
 19                 perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
 20                 disc_ndf=64, disc_loss="hinge", model_dir=None):
 21        super().__init__()
 22        assert disc_loss in ["hinge", "vanilla"]
 23        self.codebook_weight = codebook_weight
 24        self.pixel_weight = pixelloss_weight
 25        self.perceptual_loss = LPIPS(model_dir=model_dir).eval()
 26        self.perceptual_weight = perceptual_weight
 27
 28        self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
 29                                                 n_layers=disc_num_layers,
 30                                                 use_actnorm=use_actnorm,
 31                                                 ndf=disc_ndf
 32                                                 ).apply(weights_init)
 33        self.discriminator_iter_start = disc_start
 34        if disc_loss == "hinge":
 35            self.disc_loss = hinge_d_loss
 36        elif disc_loss == "vanilla":
 37            self.disc_loss = vanilla_d_loss
 38        else:
 39            raise ValueError(f"Unknown GAN loss '{disc_loss}'.")
 40        print(f"VQLPIPSWithDiscriminator running with {disc_loss} loss.")
 41        self.disc_factor = disc_factor
 42        self.discriminator_weight = disc_weight
 43        self.disc_conditional = disc_conditional
 44
 45    def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
 46        if last_layer is not None:
 47            nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
 48            g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
 49        else:
 50            nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
 51            g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]
 52
 53        d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
 54        d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
 55        d_weight = d_weight * self.discriminator_weight
 56        return d_weight
 57
 58    def forward(self, codebook_loss, inputs, reconstructions, optimizer_idx,
 59                global_step, last_layer=None, cond=None, split="train"):
 60        rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
 61        if self.perceptual_weight > 0:
 62            p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous())
 63            rec_loss = rec_loss + self.perceptual_weight * p_loss
 64        else:
 65            p_loss = torch.tensor([0.0])
 66
 67        nll_loss = rec_loss
 68        nll_loss = torch.mean(nll_loss)
 69
 70        # now the GAN part
 71        if optimizer_idx == 0:
 72            # generator update
 73            if cond is None:
 74                assert not self.disc_conditional
 75                logits_fake = self.discriminator(reconstructions.contiguous())
 76            else:
 77                assert self.disc_conditional
 78                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1))
 79            g_loss = -torch.mean(logits_fake)
 80
 81            try:
 82                d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
 83            except RuntimeError:
 84                assert not self.training
 85                d_weight = torch.tensor(0.0)
 86
 87            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
 88            loss = nll_loss + d_weight * disc_factor * g_loss + self.codebook_weight * codebook_loss.mean()
 89
 90            return loss
 91
 92        if optimizer_idx == 1:
 93            # second pass for discriminator update
 94            if cond is None:
 95                logits_real = self.discriminator(inputs.contiguous().detach())
 96                logits_fake = self.discriminator(reconstructions.contiguous().detach())
 97            else:
 98                logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1))
 99                logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1))
100
101            disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
102            d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
103
104            return d_loss