hma / magvit2 /modules /losses /vqperceptual.py
LeroyWaa's picture
draft
246c106
raw
history blame
12.8 kB
"""
Modified Open-MAGVIT2 code to use VQConfig.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from magvit2.config import VQConfig
from magvit2.modules.losses.lpips import LPIPS
from magvit2.modules.discriminator.model import NLayerDiscriminator, weights_init
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 _sigmoid_cross_entropy_with_logits(labels, logits):
"""
non-saturating loss
"""
zeros = torch.zeros_like(logits, dtype=logits.dtype)
condition = (logits >= zeros)
relu_logits = torch.where(condition, logits, zeros)
neg_abs_logits = torch.where(condition, -logits, logits)
return relu_logits - logits * labels + torch.log1p(torch.exp(neg_abs_logits))
def non_saturate_gen_loss(logits_fake):
"""
logits_fake: [B 1 H W]
"""
B, _, _, _ = logits_fake.shape
logits_fake = logits_fake.reshape(B, -1)
logits_fake = torch.mean(logits_fake, dim=-1)
gen_loss = torch.mean(_sigmoid_cross_entropy_with_logits(
labels = torch.ones_like(logits_fake), logits=logits_fake
))
return gen_loss
def non_saturate_discriminator_loss(logits_real, logits_fake):
B, _, _, _ = logits_fake.shape
logits_real = logits_fake.reshape(B, -1)
logits_fake = logits_fake.reshape(B, -1)
logits_fake = logits_fake.mean(dim=-1)
logits_real = logits_real.mean(dim=-1)
real_loss = _sigmoid_cross_entropy_with_logits(
labels=torch.ones_like(logits_real), logits=logits_real)
fake_loss = _sigmoid_cross_entropy_with_logits(
labels= torch.zeros_like(logits_fake), logits=logits_fake
)
discr_loss = real_loss.mean() + fake_loss.mean()
return discr_loss
class LeCAM_EMA(object):
def __init__(self, init=0., decay=0.999):
self.logits_real_ema = init
self.logits_fake_ema = init
self.decay = decay
def update(self, logits_real, logits_fake):
self.logits_real_ema = self.logits_real_ema * self.decay + torch.mean(logits_real).item() * (1- self.decay)
self.logits_fake_ema = self.logits_fake_ema * self.decay + torch.mean(logits_fake).item() * (1 - self.decay)
def lecam_reg(real_pred, fake_pred, lecam_ema):
reg = torch.mean(F.relu(real_pred - lecam_ema.logits_fake_ema).pow(2)) + \
torch.mean(F.relu(lecam_ema.logits_real_ema - fake_pred).pow(2))
return reg
class VQLPIPSWithDiscriminator(nn.Module):
# def __init__(self, disc_start, codebook_weight=1.0, pixelloss_weight=1.0,
# disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0,
# commit_weight = 0.25, codebook_enlarge_ratio=3, codebook_enlarge_steps=2000,
# perceptual_weight=1.0, use_actnorm=False, disc_conditional=False,
# disc_ndf=64, disc_loss="hinge", gen_loss_weight=None, lecam_loss_weight=None):
def __init__(self, config: VQConfig):
super().__init__()
assert config.disc_loss in ["hinge", "vanilla", "non_saturate"]
self.codebook_weight = config.codebook_weight
self.pixel_weight = config.pixelloss_weight
self.perceptual_loss = LPIPS().eval()
self.perceptual_weight = config.perceptual_weight
self.commit_weight = config.commit_weight
self.codebook_enlarge_ratio = config.codebook_enlarge_ratio
self.codebook_enlarge_steps = config.codebook_enlarge_steps
self.gen_loss_weight = config.gen_loss_weight
self.lecam_loss_weight = config.lecam_loss_weight
if self.lecam_loss_weight is not None:
self.lecam_ema = LeCAM_EMA()
self.discriminator = NLayerDiscriminator(
input_nc=config.disc_in_channels,
n_layers=config.disc_num_layers,
use_actnorm=config.use_actnorm,
ndf=config.disc_ndf
).apply(weights_init)
self.discriminator_iter_start = config.disc_start
self.disc_loss = {
"hinge": hinge_d_loss,
"vanilla": vanilla_d_loss,
"non_saturate": non_saturate_discriminator_loss,
}[config.disc_loss]
print(f"VQLPIPSWithDiscriminator running with {config.disc_loss} loss.")
self.disc_factor = config.disc_factor
self.discriminator_weight = config.disc_weight
self.disc_conditional = config.disc_conditional
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, codebook_loss, loss_break, inputs, reconstructions, optimizer_idx,
global_step, last_layer=None, cond=None, split="train"):
# now the GAN part
if optimizer_idx == 0:
### This code was previously outside this if statement, but seemed unnecessary? - Kevin
rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous())
nll_loss = rec_loss.clone()
if self.perceptual_weight > 0:
p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous())
nll_loss = nll_loss + self.perceptual_weight * p_loss
else:
p_loss = torch.tensor([0.0])
nll_loss = torch.mean(nll_loss)
###
# generator update
if cond is None:
assert not self.disc_conditional
logits_fake = self.discriminator(reconstructions.contiguous())
else:
assert self.disc_conditional
logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1))
g_loss = non_saturate_gen_loss(logits_fake)
if self.gen_loss_weight is None:
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)
else:
d_weight = torch.tensor(self.gen_loss_weight)
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
if not self.training:
real_g_loss = disc_factor * g_loss
g_loss = d_weight * disc_factor * g_loss
scale_codebook_loss = self.codebook_weight * codebook_loss #entropy_loss
if self.codebook_enlarge_ratio > 0:
scale_codebook_loss = self.codebook_enlarge_ratio * (max(0, 1 - global_step / self.codebook_enlarge_steps)) * scale_codebook_loss + scale_codebook_loss
loss = nll_loss + g_loss + scale_codebook_loss + loss_break.commitment * self.commit_weight
if disc_factor == 0:
log = {"{}/total_loss".format(split): loss.clone().detach(),
"{}/per_sample_entropy".format(split): loss_break.per_sample_entropy.detach(),
"{}/codebook_entropy".format(split): loss_break.codebook_entropy.detach(),
"{}/commit_loss".format(split): loss_break.commitment.detach(),
"{}/nll_loss".format(split): nll_loss.detach(),
"{}/reconstruct_loss".format(split): rec_loss.detach().mean(),
"{}/perceptual_loss".format(split): p_loss.detach().mean(),
"{}/d_weight".format(split): torch.tensor(0.0),
"{}/disc_factor".format(split): torch.tensor(0.0),
"{}/g_loss".format(split): torch.tensor(0.0),
}
else:
if self.training:
log = {"{}/total_loss".format(split): loss.clone().detach(),
"{}/per_sample_entropy".format(split): loss_break.per_sample_entropy.detach(),
"{}/codebook_entropy".format(split): loss_break.codebook_entropy.detach(),
"{}/commit_loss".format(split): loss_break.commitment.detach(),
"{}/entropy_loss".format(split): codebook_loss.detach(),
"{}/nll_loss".format(split): nll_loss.detach(),
"{}/reconstruct_loss".format(split): rec_loss.detach().mean(),
"{}/perceptual_loss".format(split): p_loss.detach().mean(),
"{}/d_weight".format(split): d_weight,
"{}/disc_factor".format(split): torch.tensor(disc_factor),
"{}/g_loss".format(split): g_loss.detach(),
}
else:
# validation only monitor the reconstruct_loss and p_loss
log = {
"{}/reconstruct_loss".format(split): rec_loss.detach().mean(),
"{}/perceptual_loss".format(split): p_loss.detach().mean(),
"{}/g_loss".format(split): real_g_loss.detach(),
}
return loss, log
if optimizer_idx == 1:
# second pass for discriminator update
if cond is None:
logits_real = self.discriminator(inputs.contiguous().detach())
logits_fake = self.discriminator(reconstructions.contiguous().detach())
else:
logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1))
logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1))
disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start)
#---------------------------------------------------------------------------------------
# Non-Saturate Loss is the Format of GAN Training, for D Loss, We still adopt Hinge Loss
#---------------------------------------------------------------------------------------
if self.lecam_loss_weight is not None:
self.lecam_ema.update(logits_real, logits_fake)
lecam_loss = lecam_reg(logits_real, logits_fake, self.lecam_ema)
non_saturate_d_loss = self.disc_loss(logits_real, logits_fake)
d_loss = disc_factor * (lecam_loss * self.lecam_loss_weight + non_saturate_d_loss)
else:
non_saturate_d_loss = self.disc_loss(logits_real, logits_fake)
d_loss = disc_factor * non_saturate_d_loss
# d_loss = disc_factor *
if disc_factor == 0:
log = {"{}/disc_loss".format(split): torch.tensor(0.0),
"{}/logits_real".format(split): torch.tensor(0.0),
"{}/logits_fake".format(split): torch.tensor(0.0),
"{}/disc_factor".format(split): torch.tensor(disc_factor),
"{}/lecam_loss".format(split): lecam_loss.detach(),
"{}/non_saturated_d_loss".format(split): non_saturate_d_loss.detach(),
}
else:
log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(),
"{}/logits_real".format(split): logits_real.detach().mean(),
"{}/logits_fake".format(split): logits_fake.detach().mean(),
"{}/disc_factor".format(split): torch.tensor(disc_factor),
"{}/lecam_loss".format(split): lecam_loss.detach(),
"{}/non_saturated_d_loss".format(split): non_saturate_d_loss.detach(),
}
return d_loss, log