|
import torch |
|
from torch import nn |
|
import torch.nn.functional as F |
|
from .lpips import LPIPS |
|
from einops import rearrange |
|
from .discriminator import NLayerDiscriminator, weights_init, NLayerDiscriminator3D |
|
|
|
|
|
def hinge_d_loss(logits_real, logits_fake): |
|
loss_real = torch.mean(F.relu(1.0 - logits_real)) |
|
loss_fake = torch.mean(F.relu(1.0 + 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 hinge_d_loss_with_exemplar_weights(logits_real, logits_fake, weights): |
|
assert weights.shape[0] == logits_real.shape[0] == logits_fake.shape[0] |
|
loss_real = torch.mean(F.relu(1.0 - logits_real), dim=[1, 2, 3]) |
|
loss_fake = torch.mean(F.relu(1.0 + logits_fake), dim=[1, 2, 3]) |
|
loss_real = (weights * loss_real).sum() / weights.sum() |
|
loss_fake = (weights * loss_fake).sum() / weights.sum() |
|
d_loss = 0.5 * (loss_real + loss_fake) |
|
return d_loss |
|
|
|
|
|
def adopt_weight(weight, global_step, threshold=0, value=0.0): |
|
if global_step < threshold: |
|
weight = value |
|
return weight |
|
|
|
|
|
def measure_perplexity(predicted_indices, n_embed): |
|
|
|
|
|
encodings = F.one_hot(predicted_indices, n_embed).float().reshape(-1, n_embed) |
|
avg_probs = encodings.mean(0) |
|
perplexity = (-(avg_probs * torch.log(avg_probs + 1e-10)).sum()).exp() |
|
cluster_use = torch.sum(avg_probs > 0) |
|
return perplexity, cluster_use |
|
|
|
|
|
def l1(x, y): |
|
return torch.abs(x - y) |
|
|
|
|
|
def l2(x, y): |
|
return torch.pow((x - y), 2) |
|
|
|
|
|
class LPIPSWithDiscriminator(nn.Module): |
|
def __init__( |
|
self, |
|
disc_start, |
|
logvar_init=0.0, |
|
kl_weight=1.0, |
|
pixelloss_weight=1.0, |
|
perceptual_weight=1.0, |
|
|
|
disc_num_layers=3, |
|
disc_in_channels=3, |
|
disc_factor=1.0, |
|
disc_weight=1.0, |
|
use_actnorm=False, |
|
disc_conditional=False, |
|
disc_loss="hinge", |
|
): |
|
|
|
super().__init__() |
|
assert disc_loss in ["hinge", "vanilla"] |
|
self.kl_weight = kl_weight |
|
self.pixel_weight = pixelloss_weight |
|
self.perceptual_loss = LPIPS().eval() |
|
self.perceptual_weight = perceptual_weight |
|
self.logvar = nn.Parameter(torch.ones(size=()) * logvar_init) |
|
|
|
self.discriminator = NLayerDiscriminator( |
|
input_nc=disc_in_channels, n_layers=disc_num_layers, use_actnorm=use_actnorm |
|
).apply(weights_init) |
|
self.discriminator_iter_start = disc_start |
|
self.disc_loss = hinge_d_loss if disc_loss == "hinge" else vanilla_d_loss |
|
self.disc_factor = disc_factor |
|
self.discriminator_weight = disc_weight |
|
self.disc_conditional = 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, |
|
inputs, |
|
reconstructions, |
|
posteriors, |
|
optimizer_idx, |
|
global_step, |
|
split="train", |
|
weights=None, |
|
last_layer=None, |
|
cond=None, |
|
): |
|
inputs = rearrange(inputs, "b c t h w -> (b t) c h w").contiguous() |
|
reconstructions = rearrange( |
|
reconstructions, "b c t h w -> (b t) c h w" |
|
).contiguous() |
|
rec_loss = torch.abs(inputs - reconstructions) |
|
if self.perceptual_weight > 0: |
|
p_loss = self.perceptual_loss(inputs, reconstructions) |
|
rec_loss = rec_loss + self.perceptual_weight * p_loss |
|
nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar |
|
weighted_nll_loss = nll_loss |
|
if weights is not None: |
|
weighted_nll_loss = weights * nll_loss |
|
weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0] |
|
nll_loss = torch.sum(nll_loss) / nll_loss.shape[0] |
|
kl_loss = posteriors.kl() |
|
kl_loss = torch.sum(kl_loss) / kl_loss.shape[0] |
|
|
|
|
|
if optimizer_idx == 0: |
|
|
|
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 = -torch.mean(logits_fake) |
|
|
|
if self.disc_factor > 0.0: |
|
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(0.0) |
|
|
|
disc_factor = adopt_weight( |
|
self.disc_factor, global_step, threshold=self.discriminator_iter_start |
|
) |
|
loss = ( |
|
weighted_nll_loss |
|
+ self.kl_weight * kl_loss |
|
+ d_weight * disc_factor * g_loss |
|
) |
|
log = { |
|
"{}/total_loss".format(split): loss.clone().detach().mean(), |
|
"{}/logvar".format(split): self.logvar.detach(), |
|
"{}/kl_loss".format(split): kl_loss.detach().mean(), |
|
"{}/nll_loss".format(split): nll_loss.detach().mean(), |
|
"{}/rec_loss".format(split): rec_loss.detach().mean(), |
|
"{}/d_weight".format(split): d_weight.detach(), |
|
"{}/disc_factor".format(split): torch.tensor(disc_factor), |
|
"{}/g_loss".format(split): g_loss.detach().mean(), |
|
} |
|
return loss, log |
|
|
|
if optimizer_idx == 1: |
|
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 |
|
) |
|
d_loss = disc_factor * self.disc_loss(logits_real, logits_fake) |
|
|
|
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(), |
|
} |
|
return d_loss, log |
|
|
|
|
|
class LPIPSWithDiscriminator3D(nn.Module): |
|
def __init__( |
|
self, |
|
disc_start, |
|
logvar_init=0.0, |
|
kl_weight=1.0, |
|
pixelloss_weight=1.0, |
|
perceptual_weight=1.0, |
|
|
|
disc_num_layers=3, |
|
disc_in_channels=3, |
|
disc_factor=1.0, |
|
disc_weight=1.0, |
|
use_actnorm=False, |
|
disc_conditional=False, |
|
disc_loss="hinge", |
|
): |
|
|
|
super().__init__() |
|
assert disc_loss in ["hinge", "vanilla"] |
|
self.kl_weight = kl_weight |
|
self.pixel_weight = pixelloss_weight |
|
self.perceptual_loss = LPIPS().eval() |
|
self.perceptual_weight = perceptual_weight |
|
self.logvar = nn.Parameter(torch.ones(size=()) * logvar_init) |
|
|
|
self.discriminator = NLayerDiscriminator3D( |
|
input_nc=disc_in_channels, n_layers=disc_num_layers, use_actnorm=use_actnorm |
|
).apply(weights_init) |
|
self.discriminator_iter_start = disc_start |
|
self.disc_loss = hinge_d_loss if disc_loss == "hinge" else vanilla_d_loss |
|
self.disc_factor = disc_factor |
|
self.discriminator_weight = disc_weight |
|
self.disc_conditional = 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, |
|
inputs, |
|
reconstructions, |
|
posteriors, |
|
optimizer_idx, |
|
global_step, |
|
split="train", |
|
weights=None, |
|
last_layer=None, |
|
cond=None, |
|
): |
|
t = inputs.shape[2] |
|
inputs = rearrange(inputs, "b c t h w -> (b t) c h w").contiguous() |
|
reconstructions = rearrange( |
|
reconstructions, "b c t h w -> (b t) c h w" |
|
).contiguous() |
|
rec_loss = torch.abs(inputs - reconstructions) |
|
if self.perceptual_weight > 0: |
|
p_loss = self.perceptual_loss(inputs, reconstructions) |
|
rec_loss = rec_loss + self.perceptual_weight * p_loss |
|
nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar |
|
weighted_nll_loss = nll_loss |
|
if weights is not None: |
|
weighted_nll_loss = weights * nll_loss |
|
weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0] |
|
nll_loss = torch.sum(nll_loss) / nll_loss.shape[0] |
|
kl_loss = posteriors.kl() |
|
kl_loss = torch.sum(kl_loss) / kl_loss.shape[0] |
|
inputs = rearrange(inputs, "(b t) c h w -> b c t h w", t=t).contiguous() |
|
reconstructions = rearrange( |
|
reconstructions, "(b t) c h w -> b c t h w", t=t |
|
).contiguous() |
|
|
|
if optimizer_idx == 0: |
|
|
|
if cond is None: |
|
assert not self.disc_conditional |
|
logits_fake = self.discriminator(reconstructions) |
|
else: |
|
assert self.disc_conditional |
|
logits_fake = self.discriminator( |
|
torch.cat((reconstructions, cond), dim=1) |
|
) |
|
g_loss = -torch.mean(logits_fake) |
|
|
|
if self.disc_factor > 0.0: |
|
try: |
|
d_weight = self.calculate_adaptive_weight( |
|
nll_loss, g_loss, last_layer=last_layer |
|
) |
|
except RuntimeError as e: |
|
assert not self.training, print(e) |
|
d_weight = torch.tensor(0.0) |
|
else: |
|
d_weight = torch.tensor(0.0) |
|
|
|
disc_factor = adopt_weight( |
|
self.disc_factor, global_step, threshold=self.discriminator_iter_start |
|
) |
|
loss = ( |
|
weighted_nll_loss |
|
+ self.kl_weight * kl_loss |
|
+ d_weight * disc_factor * g_loss |
|
) |
|
log = { |
|
"{}/total_loss".format(split): loss.clone().detach().mean(), |
|
"{}/logvar".format(split): self.logvar.detach(), |
|
"{}/kl_loss".format(split): kl_loss.detach().mean(), |
|
"{}/nll_loss".format(split): nll_loss.detach().mean(), |
|
"{}/rec_loss".format(split): rec_loss.detach().mean(), |
|
"{}/d_weight".format(split): d_weight.detach(), |
|
"{}/disc_factor".format(split): torch.tensor(disc_factor), |
|
"{}/g_loss".format(split): g_loss.detach().mean(), |
|
} |
|
return loss, log |
|
|
|
if optimizer_idx == 1: |
|
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 |
|
) |
|
d_loss = disc_factor * self.disc_loss(logits_real, logits_fake) |
|
|
|
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(), |
|
} |
|
return d_loss, log |
|
|
|
|
|
class SimpleLPIPS(nn.Module): |
|
def __init__( |
|
self, |
|
logvar_init=0.0, |
|
kl_weight=1.0, |
|
pixelloss_weight=1.0, |
|
perceptual_weight=1.0, |
|
disc_loss="hinge", |
|
): |
|
|
|
super().__init__() |
|
assert disc_loss in ["hinge", "vanilla"] |
|
self.kl_weight = kl_weight |
|
self.pixel_weight = pixelloss_weight |
|
self.perceptual_loss = LPIPS().eval() |
|
self.perceptual_weight = perceptual_weight |
|
self.logvar = nn.Parameter(torch.ones(size=()) * logvar_init) |
|
|
|
def forward( |
|
self, |
|
inputs, |
|
reconstructions, |
|
posteriors, |
|
split="train", |
|
weights=None, |
|
): |
|
inputs = rearrange(inputs, "b c t h w -> (b t) c h w").contiguous() |
|
reconstructions = rearrange( |
|
reconstructions, "b c t h w -> (b t) c h w" |
|
).contiguous() |
|
rec_loss = torch.abs(inputs - reconstructions) |
|
if self.perceptual_weight > 0: |
|
p_loss = self.perceptual_loss(inputs, reconstructions) |
|
rec_loss = rec_loss + self.perceptual_weight * p_loss |
|
nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar |
|
weighted_nll_loss = nll_loss |
|
if weights is not None: |
|
weighted_nll_loss = weights * nll_loss |
|
weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0] |
|
nll_loss = torch.sum(nll_loss) / nll_loss.shape[0] |
|
kl_loss = posteriors.kl() |
|
kl_loss = torch.sum(kl_loss) / kl_loss.shape[0] |
|
loss = weighted_nll_loss + self.kl_weight * kl_loss |
|
log = { |
|
"{}/total_loss".format(split): loss.clone().detach().mean(), |
|
"{}/logvar".format(split): self.logvar.detach(), |
|
"{}/kl_loss".format(split): kl_loss.detach().mean(), |
|
"{}/nll_loss".format(split): nll_loss.detach().mean(), |
|
"{}/rec_loss".format(split): rec_loss.detach().mean(), |
|
} |
|
if self.perceptual_weight > 0: |
|
log.update({"{}/p_loss".format(split): p_loss.detach().mean()}) |
|
return loss, log |
|
|