Spaces:
Runtime error
Runtime error
File size: 11,374 Bytes
1b2a9b1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 |
"""VQGAN Loss
- Adapted from https://github.com/CompVis/taming-transformers
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from .discriminator import NLayerDiscriminator, weights_init
from .blocks import LossCriterion, LossCriterionMask
class DummyLoss(nn.Module):
def __init__(self):
super().__init__()
def adopt_weight(weight, global_step, threshold=0, value=0.):
if global_step < threshold:
weight = value
return weight
def hinge_d_loss(logits_real, logits_fake):
loss_real = torch.mean(F.relu(1. - logits_real))
loss_fake = torch.mean(F.relu(1. + logits_fake))
d_loss = 0.5 * (loss_real + loss_fake)
return d_loss
def vanilla_d_loss(logits_real, logits_fake):
d_loss = 0.5 * (
torch.mean(torch.nn.functional.softplus(-logits_real)) +
torch.mean(torch.nn.functional.softplus(logits_fake)))
return d_loss
def fft_loss(pred, tgt):
return ((torch.fft.fftn(pred, dim=(-2,-1)) - torch.fft.fftn(tgt, dim=(-2,-1)))).abs().mean()
class LPIPSWithDiscriminator(nn.Module):
def __init__(self, disc_start, model_path, pixelloss_weight=1.0,
disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=0.8,
perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
disc_ndf=64, disc_loss="hinge", rec_loss="FFT",
style_layers = [], content_layers = ['r41']):
super().__init__()
assert disc_loss in ["hinge", "vanilla"]
self.pixel_weight = pixelloss_weight
self.perceptual_loss = LossCriterion(style_layers, content_layers,
0, perceptual_weight,
model_path = model_path)
self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
n_layers=disc_num_layers,
use_actnorm=use_actnorm,
ndf=disc_ndf
).apply(weights_init)
self.discriminator_iter_start = disc_start
if disc_loss == "hinge":
self.disc_loss = hinge_d_loss
elif disc_loss == "vanilla":
self.disc_loss = vanilla_d_loss
else:
raise ValueError(f"Unknown GAN loss '{disc_loss}'.")
print(f"VQLPIPSWithDiscriminator running with {disc_loss} and {rec_loss} loss.")
self.disc_factor = disc_factor
self.discriminator_weight = disc_weight
self.disc_conditional = disc_conditional
self.rec_loss = rec_loss
self.perceptual_weight = perceptual_weight
def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
if last_layer is not None:
nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
else:
nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]
d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
d_weight = d_weight * self.discriminator_weight
return d_weight
def forward(self, inputs, reconstructions, optimizer_idx,
global_step, last_layer=None, cond=None, split="train"):
if self.rec_loss == "L1":
rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous()).mean()
elif self.rec_loss == "MSE":
rec_loss = F.mse_loss(reconstructions, inputs)
elif self.rec_loss == "FFT":
rec_loss = fft_loss(inputs, reconstructions)
elif self.rec_loss is None:
rec_loss = 0
else:
raise ValueError("Unkown reconstruction loss, choices are [FFT, L1]")
if self.perceptual_weight > 0:
loss_dict = self.perceptual_loss(reconstructions, inputs, style = False)
p_loss = loss_dict['content']
rec_loss = rec_loss + p_loss
else:
p_loss = torch.zeros(1).cuda()
nll_loss = rec_loss
# adversarial loss for both branches
if optimizer_idx == 0:
log = {}
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
# generator update
if disc_factor > 0:
logits_fake = self.discriminator(reconstructions.contiguous())
g_loss = -torch.mean(logits_fake)
try:
d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
except RuntimeError:
#assert not self.training
d_weight = torch.tensor(0.0)
loss = nll_loss + d_weight * disc_factor * g_loss
log["d_weight"] = d_weight.detach()
log["disc_factor"] = torch.tensor(disc_factor)
log["g_loss"] = g_loss.detach().mean()
else:
loss = nll_loss
log["total_loss"] = loss.clone().detach().mean()
log["nll_loss"] = nll_loss.detach().mean()
log["rec_loss"] = rec_loss.detach().mean()
log["p_loss"] = p_loss.detach().mean()
return loss, log
if optimizer_idx == 1:
# second pass for discriminator update
logits_real = self.discriminator(inputs.contiguous().detach())
logits_fake = self.discriminator(reconstructions.contiguous().detach())
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
log = {"disc_loss": d_loss.clone().detach().mean(),
"logits_real": logits_real.detach().mean(),
"logits_fake": logits_fake.detach().mean()
}
return d_loss, log
class LPIPSWithDiscriminatorMask(nn.Module):
def __init__(self, disc_start, model_path, pixelloss_weight=1.0,
disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=0.8,
perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
disc_ndf=64, disc_loss="hinge", rec_loss="FFT",
style_layers = [], content_layers = ['r41']):
super().__init__()
assert disc_loss in ["hinge", "vanilla"]
self.pixel_weight = pixelloss_weight
self.perceptual_loss = LossCriterionMask(style_layers, content_layers,
0.2, perceptual_weight,
model_path = model_path)
self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels,
n_layers=disc_num_layers,
use_actnorm=use_actnorm,
ndf=disc_ndf
).apply(weights_init)
self.discriminator_iter_start = disc_start
if disc_loss == "hinge":
self.disc_loss = hinge_d_loss
elif disc_loss == "vanilla":
self.disc_loss = vanilla_d_loss
else:
raise ValueError(f"Unknown GAN loss '{disc_loss}'.")
print(f"VQLPIPSWithDiscriminator running with {disc_loss} and {rec_loss} loss.")
self.disc_factor = disc_factor
self.discriminator_weight = disc_weight
self.disc_conditional = disc_conditional
self.rec_loss = rec_loss
self.perceptual_weight = perceptual_weight
def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None):
if last_layer is not None:
nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0]
g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0]
else:
nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0]
g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0]
d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4)
d_weight = torch.clamp(d_weight, 0.0, 1e4).detach()
d_weight = d_weight * self.discriminator_weight
return d_weight
def forward(self, inputs, reconstructions, optimizer_idx,
global_step, mask, last_layer=None, cond=None, split="train"):
if self.rec_loss == "L1":
rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous()).mean()
elif self.rec_loss == "MSE":
rec_loss = F.mse_loss(reconstructions, inputs)
elif self.rec_loss == "FFT":
rec_loss = fft_loss(inputs, reconstructions)
elif self.rec_loss is None:
rec_loss = 0
else:
raise ValueError("Unkown reconstruction loss, choices are [FFT, L1]")
if self.perceptual_weight > 0:
loss_dict = self.perceptual_loss(reconstructions, inputs, mask, style = True)
p_loss = loss_dict['content']
s_loss = loss_dict['style']
rec_loss = rec_loss + p_loss + s_loss
else:
p_loss = torch.zeros(1).cuda()
nll_loss = rec_loss
# adversarial loss for both branches
if optimizer_idx == 0:
# generator update
logits_fake = self.discriminator(reconstructions.contiguous())
g_loss = -torch.mean(logits_fake)
try:
d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer)
except RuntimeError:
#assert not self.training
d_weight = torch.tensor(0.0)
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
loss = nll_loss + d_weight * disc_factor * g_loss
log = {"total_loss": loss.clone().detach().mean(),
"nll_loss": nll_loss.detach().mean(),
"rec_loss": rec_loss.detach().mean(),
"p_loss": p_loss.detach().mean(),
"s_loss": s_loss,
"d_weight": d_weight.detach(),
"disc_factor": torch.tensor(disc_factor),
"g_loss": g_loss.detach().mean(),
}
return loss, log
if optimizer_idx == 1:
# second pass for discriminator update
logits_real = self.discriminator(inputs.contiguous().detach())
logits_fake = self.discriminator(reconstructions.contiguous().detach())
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
d_loss = disc_factor * self.disc_loss(logits_real, logits_fake)
log = {"disc_loss": d_loss.clone().detach().mean(),
"logits_real": logits_real.detach().mean(),
"logits_fake": logits_fake.detach().mean()
}
return d_loss, log |