def __init__(self, hparams) -> None:
super().__init__()
self.hparams = hparams
self.generator = Generator(
self.hparams.size_h,
self.hparams.size_w,
self.hparams.log_size,
self.hparams.latent,
self.hparams.n_mlp,
channel_multiplier=self.hparams.channel_multiplier)
self.discriminator = Discriminator(
self.hparams.size_h,
self.hparams.size_w,
self.hparams.log_size,
channel_multiplier=self.hparams.channel_multiplier)
self.g_ema = Generator(
self.hparams.size_h,
self.hparams.size_w,
self.hparams.log_size,
self.hparams.latent,
self.hparams.n_mlp,
channel_multiplier=self.hparams.channel_multiplier)
self.img_dim = (3, self.hparams.size_h, self.hparams.size_w)
self.accum = 0.5**(32 / (10 * 1000))
self.mean_path_length = 0
self.ada_aug_p = max(self.hparams.augment_p, 0)
self.use_ada_augment = self.hparams.augment and self.hparams.augment_p == 0
if self.use_ada_augment:
self.ada_augment = AdaptiveAugment(self.hparams.ada_target,
self.hparams.ada_length,
self.hparams.ada_every)
self.g_ema.eval()
accumulate(self.g_ema, self.generator, 0)
self.example_input_array = [torch.zeros(1, self.hparams.latent)]
def train(args, loader, encoder, generator, discriminator, discriminator_z, g1,
vggnet, pwcnet, e_optim, d_optim, dz_optim, g1_optim, e_ema, e_tf,
g1_ema, device):
mmd_eval = functools.partial(mix_rbf_mmd2,
sigma_list=[2.0, 5.0, 10.0, 20.0, 40.0, 80.0])
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
d_loss_val = 0
e_loss_val = 0
rec_loss_val = 0
vgg_loss_val = 0
adv_loss_val = 0
loss_dict = {
"d": torch.tensor(0., device=device),
"real_score": torch.tensor(0., device=device),
"fake_score": torch.tensor(0., device=device),
"r1_d": torch.tensor(0., device=device),
"r1_e": torch.tensor(0., device=device),
"rec": torch.tensor(0., device=device),
}
avg_pix_loss = util.AverageMeter()
avg_vgg_loss = util.AverageMeter()
if args.distributed:
e_module = encoder.module
d_module = discriminator.module
g_module = generator.module
g1_module = g1.module if args.train_latent_mlp else None
else:
e_module = encoder
d_module = discriminator
g_module = generator
g1_module = g1 if args.train_latent_mlp else None
accum = 0.5**(32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 256,
device)
# sample_x = accumulate_batches(loader, args.n_sample).to(device)
sample_x = load_real_samples(args, loader)
requires_grad(generator, False) # always False
generator.eval() # Generator should be ema and in eval mode
# if args.no_ema or e_ema is None:
# e_ema = encoder
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
real_img = next(loader)
real_img = real_img.to(device)
batch = real_img.shape[0]
# Train Encoder
if args.toggle_grads:
requires_grad(encoder, True)
requires_grad(discriminator, False)
pix_loss = vgg_loss = adv_loss = rec_loss = torch.tensor(0.,
device=device)
kld_z = torch.tensor(0., device=device)
mmd_z = torch.tensor(0., device=device)
gan_z = torch.tensor(0., device=device)
etf_z = torch.tensor(0., device=device)
latent_real, logvar = encoder(real_img)
if args.reparameterization:
latent_real = reparameterize(latent_real, logvar)
if args.train_latent_mlp:
fake_img, _ = generator([g1(latent_real)],
input_is_latent=True,
return_latents=False)
else:
fake_img, _ = generator([latent_real],
input_is_latent=False,
return_latents=False)
if args.lambda_adv > 0:
if args.augment:
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
fake_img_aug = fake_img
fake_pred = discriminator(fake_img_aug)
adv_loss = g_nonsaturating_loss(fake_pred)
if args.lambda_pix > 0:
pix_loss = torch.mean((real_img - fake_img)**2)
if args.lambda_vgg > 0:
real_feat = vggnet(real_img)
fake_feat = vggnet(fake_img)
vgg_loss = torch.mean((real_feat - fake_feat)**2)
if args.lambda_kld_z > 0:
z_mean = latent_real.view(batch, -1)
kld_z = -0.5 * torch.sum(1. + logvar - z_mean.pow(2) -
logvar.exp()) / batch
# print(kld_z)
if args.lambda_mmd_z > 0:
z_real = torch.randn(batch, args.latent_full, device=device)
mmd_z = mmd_eval(latent_real, z_real)
# print(mmd_z)
if args.lambda_gan_z > 0:
fake_pred = discriminator_z(latent_real)
gan_z = g_nonsaturating_loss(fake_pred)
# print(gan_z)
if args.use_latent_teacher_forcing and args.lambda_etf > 0:
w_tf, _ = e_tf(real_img)
if args.train_latent_mlp:
w_pred = g1(latent_real)
else:
w_pred = generator.get_latent(latent_real)
etf_z = torch.mean((w_tf - w_pred)**2)
# print(etf_z)
if args.train_on_fake and args.lambda_rec > 0:
z_real = torch.randn(args.batch, args.latent_full, device=device)
if args.train_latent_mlp:
fake_img, _ = generator([g1(z_real)],
input_is_latent=True,
return_latents=False)
else:
fake_img, _ = generator([z_real],
input_is_latent=False,
return_latents=False)
# fake_img, _ = generator([z_real], input_is_latent=False, return_latents=True)
z_fake, z_logvar = encoder(fake_img)
if args.reparameterization:
z_fake = reparameterize(z_fake, z_logvar)
rec_loss = torch.mean((z_real - z_fake)**2)
loss_dict["rec"] = rec_loss
# print(rec_loss)
e_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
e_loss = e_loss + args.lambda_kld_z * kld_z + args.lambda_mmd_z * mmd_z + args.lambda_gan_z * gan_z + args.lambda_etf * etf_z + rec_loss * args.lambda_rec
loss_dict["e"] = e_loss
loss_dict["pix"] = pix_loss
loss_dict["vgg"] = vgg_loss
loss_dict["adv"] = adv_loss
if args.train_latent_mlp and g1 is not None:
g1.zero_grad()
encoder.zero_grad()
e_loss.backward()
e_optim.step()
if args.train_latent_mlp and g1_optim is not None:
g1_optim.step()
# if args.train_on_fake:
# e_regularize = args.e_rec_every > 0 and i % args.e_rec_every == 0
# if e_regularize and args.lambda_rec > 0:
# # noise = mixing_noise(args.batch, args.latent, args.mixing, device)
# # fake_img, latent_fake = generator(noise, input_is_latent=False, return_latents=True)
# z_real = torch.randn(args.batch, args.latent_full, device=device)
# fake_img, w_real = generator([z_real], input_is_latent=False, return_latents=True)
# z_fake, logvar = encoder(fake_img)
# if args.reparameterization:
# z_fake = reparameterize(z_fake, logvar)
# rec_loss = torch.mean((z_real - z_fake) ** 2)
# encoder.zero_grad()
# (rec_loss * args.lambda_rec).backward()
# e_optim.step()
# loss_dict["rec"] = rec_loss
e_regularize = args.e_reg_every > 0 and i % args.e_reg_every == 0
if e_regularize:
# why not regularize on augmented real?
real_img.requires_grad = True
real_pred, logvar = encoder(real_img)
if args.reparameterization:
real_pred = reparameterize(real_pred, logvar)
r1_loss_e = d_r1_loss(real_pred, real_img)
encoder.zero_grad()
(args.r1 / 2 * r1_loss_e * args.e_reg_every +
0 * real_pred.view(-1)[0]).backward()
e_optim.step()
loss_dict["r1_e"] = r1_loss_e
if not args.no_ema and e_ema is not None:
accumulate(e_ema, e_module, accum)
if args.train_latent_mlp:
accumulate(g1_ema, g1_module, accum)
# Train Discriminator
if args.toggle_grads:
requires_grad(encoder, False)
requires_grad(discriminator, True)
if not args.no_update_discriminator and args.lambda_adv > 0:
latent_real, logvar = encoder(real_img)
if args.reparameterization:
latent_real = reparameterize(latent_real, logvar)
if args.train_latent_mlp:
fake_img, _ = generator([g1(latent_real)],
input_is_latent=True,
return_latents=False)
else:
fake_img, _ = generator([latent_real],
input_is_latent=False,
return_latents=False)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_img_aug = fake_img
fake_pred = discriminator(fake_img_aug)
real_pred = discriminator(real_img_aug)
d_loss = d_logistic_loss(real_pred, fake_pred)
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
z_real = torch.randn(batch, args.latent_full, device=device)
fake_pred = discriminator_z(latent_real.detach())
real_pred = discriminator_z(z_real)
d_loss_z = d_logistic_loss(real_pred, fake_pred)
discriminator_z.zero_grad()
d_loss_z.backward()
dz_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = args.d_reg_every > 0 and i % args.d_reg_every == 0
if d_regularize:
# why not regularize on augmented real?
real_img.requires_grad = True
real_pred = discriminator(real_img)
r1_loss_d = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss_d * args.d_reg_every +
0 * real_pred.view(-1)[0]).backward()
# Why 0* ? Answer is here https://github.com/rosinality/stylegan2-pytorch/issues/76
d_optim.step()
loss_dict["r1_d"] = r1_loss_d
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
e_loss_val = loss_reduced["e"].mean().item()
r1_d_val = loss_reduced["r1_d"].mean().item()
r1_e_val = loss_reduced["r1_e"].mean().item()
pix_loss_val = loss_reduced["pix"].mean().item()
vgg_loss_val = loss_reduced["vgg"].mean().item()
adv_loss_val = loss_reduced["adv"].mean().item()
rec_loss_val = loss_reduced["rec"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
avg_pix_loss.update(pix_loss_val, real_img.shape[0])
avg_vgg_loss.update(vgg_loss_val, real_img.shape[0])
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; e: {e_loss_val:.4f}; r1_d: {r1_d_val:.4f}; r1_e: {r1_e_val:.4f}; "
f"pix: {pix_loss_val:.4f}; vgg: {vgg_loss_val:.4f}; adv: {adv_loss_val:.4f}; "
f"rec: {rec_loss_val:.4f}; augment: {ada_aug_p:.4f}"))
if i % args.log_every == 0:
with torch.no_grad():
latent_x, _ = e_ema(sample_x)
if args.train_latent_mlp:
g1_ema.eval()
fake_x, _ = generator([g1_ema(latent_x)],
input_is_latent=True,
return_latents=False)
else:
fake_x, _ = generator([latent_x],
input_is_latent=False,
return_latents=False)
sample_pix_loss = torch.sum((sample_x - fake_x)**2)
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(
f"{i:07d}; pix: {avg_pix_loss.avg}; vgg: {avg_vgg_loss.avg}; "
f"ref: {sample_pix_loss.item()};\n")
if wandb and args.wandb:
wandb.log({
"Encoder": e_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1 D": r1_d_val,
"R1 E": r1_e_val,
"Pix Loss": pix_loss_val,
"VGG Loss": vgg_loss_val,
"Adv Loss": adv_loss_val,
"Rec Loss": rec_loss_val,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
})
if i % args.log_every == 0:
with torch.no_grad():
e_eval = encoder if args.no_ema else e_ema
e_eval.eval()
nrow = int(args.n_sample**0.5)
nchw = list(sample_x.shape)[1:]
latent_real, _ = e_eval(sample_x)
if args.train_latent_mlp:
g1_ema.eval()
fake_img, _ = generator([g1_ema(latent_real)],
input_is_latent=True,
return_latents=False)
else:
fake_img, _ = generator([latent_real],
input_is_latent=False,
return_latents=False)
sample = torch.cat(
(sample_x.reshape(args.n_sample // nrow, nrow, *nchw),
fake_img.reshape(args.n_sample // nrow, nrow, *nchw)),
1)
utils.save_image(
sample.reshape(2 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
e_eval.train()
if i % args.save_every == 0:
e_eval = encoder if args.no_ema else e_ema
torch.save(
{
"e":
e_module.state_dict(),
"d":
d_module.state_dict(),
"g1":
g1_module.state_dict()
if args.train_latent_mlp else None,
"g1_ema":
g1_ema.state_dict() if args.train_latent_mlp else None,
"g_ema":
g_module.state_dict(),
"e_ema":
e_eval.state_dict(),
"e_optim":
e_optim.state_dict(),
"d_optim":
d_optim.state_dict(),
"args":
args,
"ada_aug_p":
ada_aug_p,
"iter":
i,
},
os.path.join(args.log_dir, 'weight',
f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"e":
e_module.state_dict(),
"d":
d_module.state_dict(),
"g1":
g1_module.state_dict()
if args.train_latent_mlp else None,
"g1_ema":
g1_ema.state_dict() if args.train_latent_mlp else None,
"g_ema":
g_module.state_dict(),
"e_ema":
e_eval.state_dict(),
"e_optim":
e_optim.state_dict(),
"d_optim":
d_optim.state_dict(),
"args":
args,
"ada_aug_p":
ada_aug_p,
"iter":
i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, generator, encoder, discriminator, vggnet, g_optim,
e_optim, d_optim, g_ema, e_ema, device):
kwargs_d = {'detach_aux': False}
if args.dataset == 'imagefolder':
loader = sample_data2(loader)
else:
loader = sample_data(loader)
if args.eval_every > 0:
inception = nn.DataParallel(load_patched_inception_v3()).to(device)
inception.eval()
with open(args.inception, "rb") as f:
embeds = pickle.load(f)
real_mean = embeds["mean"]
real_cov = embeds["cov"]
else:
inception = real_mean = real_cov = None
mean_latent = None
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
r1_loss = torch.tensor(0.0, device=device)
g_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
loss_dict = {}
avg_pix_loss = util.AverageMeter()
avg_vgg_loss = util.AverageMeter()
if args.distributed:
g_module = generator.module
e_module = encoder.module
d_module = discriminator.module
else:
g_module = generator
e_module = encoder
d_module = discriminator
accum = 0.5**(32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 256,
device)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
sample_x = load_real_samples(args, loader)
if sample_x.ndim > 4:
sample_x = sample_x[:, 0, ...]
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
real_img = next(loader)
real_img = real_img.to(device)
# Train Discriminator
requires_grad(generator, False)
requires_grad(encoder, False)
requires_grad(discriminator, True)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
real_pred = discriminator(real_img)
fake_pred = discriminator(fake_img)
rec_pred = discriminator(rec_img)
d_loss_real = F.softplus(-real_pred).mean()
d_loss_fake = F.softplus(fake_pred).mean()
d_loss_rec = F.softplus(rec_pred).mean()
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
loss_dict["rec_score"] = rec_pred.mean()
d_loss = d_loss_real + d_loss_fake + d_loss_rec
loss_dict["d"] = d_loss
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
d_regularize = args.d_reg_every > 0 and i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
real_pred = discriminator(real_img)
r1_loss = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every +
0 * real_pred[0]).backward()
d_optim.step()
loss_dict["r1"] = r1_loss
# # Train Encoder and Generator
# requires_grad(generator, True)
# requires_grad(encoder, True)
# requires_grad(discriminator, False)
# pix_loss = vgg_loss = adv_loss = torch.tensor(0., device=device)
# noise = mixing_noise(args.batch, args.latent, args.mixing, device)
# fake_img, _ = generator(noise)
# latent_real, _ = encoder(real_img)
# rec_img, _ = generator([latent_real], input_is_latent=True)
# fake_pred = discriminator(fake_img)
# rec_pred = discriminator(rec_img)
# g_loss_fake = g_nonsaturating_loss(fake_pred)
# g_loss_rec = g_nonsaturating_loss(rec_pred)
# adv_loss = g_loss_fake + g_loss_rec
# if args.lambda_pix > 0:
# if args.pix_loss == 'l2':
# pix_loss = torch.mean((rec_img - real_img) ** 2)
# else:
# pix_loss = F.l1_loss(rec_img, real_img)
# if args.lambda_vgg > 0:
# vgg_loss = torch.mean((vggnet(real_img) - vggnet(rec_img)) ** 2)
# e_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
# loss_dict["e"] = e_loss
# encoder.zero_grad()
# generator.zero_grad()
# e_loss.backward()
# e_optim.step()
# g_optim.step()
# Train Encoder
requires_grad(generator, False)
requires_grad(encoder, True)
requires_grad(discriminator, False)
pix_loss = vgg_loss = adv_loss = torch.tensor(0., device=device)
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
rec_pred = discriminator(rec_img)
g_loss_rec = g_nonsaturating_loss(rec_pred)
adv_loss = g_loss_rec
if args.lambda_pix > 0:
if args.pix_loss == 'l2':
pix_loss = torch.mean((rec_img - real_img)**2)
else:
pix_loss = F.l1_loss(rec_img, real_img)
if args.lambda_vgg > 0:
vgg_loss = torch.mean((vggnet(real_img) - vggnet(rec_img))**2)
e_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
loss_dict["e"] = e_loss
encoder.zero_grad()
e_loss.backward()
e_optim.step()
# Train Generator
requires_grad(generator, True)
requires_grad(encoder, False)
requires_grad(discriminator, False)
pix_loss = vgg_loss = adv_loss = torch.tensor(0., device=device)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
fake_pred = discriminator(fake_img)
rec_pred = discriminator(rec_img)
g_loss_fake = g_nonsaturating_loss(fake_pred)
g_loss_rec = g_nonsaturating_loss(rec_pred)
adv_loss = g_loss_fake + g_loss_rec
if args.lambda_pix > 0:
if args.pix_loss == 'l2':
pix_loss = torch.mean((rec_img - real_img)**2)
else:
pix_loss = F.l1_loss(rec_img, real_img)
if args.lambda_vgg > 0:
vgg_loss = torch.mean((vggnet(real_img) - vggnet(rec_img))**2)
g_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
loss_dict["g"] = g_loss
generator.zero_grad()
g_loss.backward()
g_optim.step()
g_regularize = args.g_reg_every > 0 and i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing,
device)
fake_img, latents = generator(noise, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
mean_path_length_avg = (reduce_sum(mean_path_length).item() /
get_world_size())
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
with torch.no_grad():
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
if args.pix_loss == 'l2':
pix_loss = torch.mean((rec_img - real_img)**2)
else:
pix_loss = F.l1_loss(rec_img, real_img)
vgg_loss = torch.mean((vggnet(real_img) - vggnet(rec_img))**2)
pix_loss_val = pix_loss.mean().item()
vgg_loss_val = vgg_loss.mean().item()
accumulate(e_ema, e_module, accum)
accumulate(g_ema, g_module, accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
e_loss_val = loss_reduced["e"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
avg_pix_loss.update(pix_loss_val, real_img.shape[0])
avg_vgg_loss.update(vgg_loss_val, real_img.shape[0])
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; e: {e_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}"))
if i % args.log_every == 0:
with torch.no_grad():
latent_x, _ = e_ema(sample_x)
fake_x, _ = generator([latent_x],
input_is_latent=True,
return_latents=False)
sample_pix_loss = torch.sum((sample_x - fake_x)**2)
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(
f"{i:07d}; pix: {avg_pix_loss.avg}; vgg: {avg_vgg_loss.avg}; "
f"ref: {sample_pix_loss.item()};\n")
if args.eval_every > 0 and i % args.eval_every == 0:
with torch.no_grad():
g_ema.eval()
if args.truncation < 1:
mean_latent = g_ema.mean_latent(4096)
features = extract_feature_from_samples(
g_ema, inception, args.truncation, mean_latent, 64,
args.n_sample_fid, args.device).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
print("fid:", fid)
with open(os.path.join(args.log_dir, 'log_fid.txt'),
'a+') as f:
f.write(f"{i:07d}: fid: {float(fid):.4f}\n")
if wandb and args.wandb:
wandb.log({
"Generator": g_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1": r1_val,
"Path Length Regularization": path_loss_val,
"Mean Path Length": mean_path_length,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
"Path Length": path_length_val,
})
if i % args.log_every == 0:
with torch.no_grad():
# Fixed fake samples
g_ema.eval()
sample, _ = g_ema([sample_z])
utils.save_image(
sample,
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-sample.png"),
nrow=int(args.n_sample**0.5),
normalize=True,
value_range=(-1, 1),
)
# Reconstruction samples
e_ema.eval()
nrow = int(args.n_sample**0.5)
nchw = list(sample_x.shape)[1:]
latent_real, _ = e_ema(sample_x)
fake_img, _ = g_ema([latent_real],
input_is_latent=True,
return_latents=False)
sample = torch.cat(
(sample_x.reshape(args.n_sample // nrow, nrow, *nchw),
fake_img.reshape(args.n_sample // nrow, nrow, *nchw)),
1)
utils.save_image(
sample.reshape(2 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-recon.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
if i % args.save_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight',
f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, generator, discriminator, g_optim, d_optim, g_ema, device):
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar, initial=args.start_iter, dynamic_ncols=True, smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
r1_loss = torch.tensor(0.0, device=device)
g_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
loss_dict = {}
if args.distributed:
g_module = generator.module
d_module = discriminator.module
else:
g_module = generator
d_module = discriminator
accum = 0.5 ** (32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 8, device)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
real_img = next(loader)
real_img = real_img.to(device)
requires_grad(generator, False)
requires_grad(discriminator, True)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_pred = discriminator(fake_img)
real_pred = discriminator(real_img_aug)
d_loss = d_logistic_loss(real_pred, fake_pred)
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
else:
real_img_aug = real_img
real_pred = discriminator(real_img_aug)
r1_loss = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every + 0 * real_pred[0]).backward()
d_optim.step()
loss_dict["r1"] = r1_loss
requires_grad(generator, True)
requires_grad(discriminator, False)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = discriminator(fake_img)
g_loss = g_nonsaturating_loss(fake_pred)
loss_dict["g"] = g_loss
generator.zero_grad()
g_loss.backward()
g_optim.step()
g_regularize = i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing, device)
fake_img, latents = generator(noise, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length
)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
mean_path_length_avg = (
reduce_sum(mean_path_length).item() / get_world_size()
)
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
accumulate(g_ema, g_module, accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
g_loss_val = loss_reduced["g"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
if get_rank() == 0:
pbar.set_description(
(
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}"
)
)
if wandb and args.wandb:
wandb.log(
{
"Generator": g_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1": r1_val,
"Path Length Regularization": path_loss_val,
"Mean Path Length": mean_path_length,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
"Path Length": path_length_val,
}
)
if i % 100 == 0:
with torch.no_grad():
g_ema.eval()
sample, _ = g_ema([sample_z])
utils.save_image(
sample,
f"sample/{str(i).zfill(6)}.png",
nrow=int(args.n_sample ** 0.5),
normalize=True,
range=(-1, 1),
)
if i % 10000 == 0:
torch.save(
{
"g": g_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
},
f"checkpoint/{str(i).zfill(6)}.pt",
)
def train(
args,
loader,
encoder,
generator,
discriminator,
discriminator3d, # video disctiminator
posterior,
prior,
factor, # a learnable matrix
vggnet,
e_optim,
d_optim,
dv_optim,
q_optim, # q for posterior
p_optim, # p for prior
f_optim, # f for factor
e_ema,
device
):
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar, initial=args.start_iter, dynamic_ncols=True, smoothing=0.01)
d_loss_val = 0
e_loss_val = 0
rec_loss_val = 0
vgg_loss_val = 0
adv_loss_val = 0
loss_dict = {"d": torch.tensor(0., device=device),
"real_score": torch.tensor(0., device=device),
"fake_score": torch.tensor(0., device=device),
"r1_d": torch.tensor(0., device=device),
"r1_e": torch.tensor(0., device=device),
"rec": torch.tensor(0., device=device),}
if args.distributed:
e_module = encoder.module
d_module = discriminator.module
g_module = generator.module
else:
e_module = encoder
d_module = discriminator
g_module = generator
accum = 0.5 ** (32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
latent_full = args.latent_full
factor_dim_full = args.factor_dim_full
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 256, device)
sample_x = accumulate_batches(loader, args.n_sample).to(device)
utils.save_image(
sample_x.view(-1, *list(sample_x.shape)[2:]),
os.path.join(args.log_dir, 'sample', f"real-img.png"),
nrow=sample_x.shape[1],
normalize=True,
value_range=(-1, 1),
)
util.save_video(
sample_x[0],
os.path.join(args.log_dir, 'sample', f"real-vid.mp4")
)
requires_grad(generator, False) # always False
generator.eval() # Generator should be ema and in eval mode
if args.no_update_encoder:
encoder = e_ema if e_ema is not None else encoder
requires_grad(encoder, False)
encoder.eval()
from models.networks_3d import GANLoss
criterionGAN = GANLoss()
# criterionL1 = nn.L1Loss()
# if args.no_ema or e_ema is None:
# e_ema = encoder
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
data = next(loader)
real_seq = data['frames']
real_seq = real_seq.to(device) # [N, T, C, H, W]
shape = list(real_seq.shape)
N, T = shape[:2]
# Train Encoder with frame-level objectives
if args.toggle_grads:
if not args.no_update_encoder:
requires_grad(encoder, True)
requires_grad(discriminator, False)
pix_loss = vgg_loss = adv_loss = rec_loss = vid_loss = l1y_loss = torch.tensor(0., device=device)
# TODO: real_seq -> encoder -> posterior -> generator -> fake_seq
# f: [N, latent_full]; y: [N, T, D]
fake_img, fake_seq, y_post = reconstruct_sequence(args, real_seq, encoder, generator, factor, posterior, i, ret_y=True)
# if args.debug == 'no_lstm':
# real_lat = encoder(real_seq.view(-1, *shape[2:]))
# fake_img, _ = generator([real_lat], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
# elif args.debug == 'decomp':
# real_lat = encoder(real_seq.view(-1, *shape[2:])) # [N*T, latent_full]
# f_post = real_lat[::T, ...]
# z_post = real_lat.view(N, T, -1) - f_post.unsqueeze(1)
# if args.use_multi_head:
# y_post = []
# for z, w in zip(torch.split(z_post, 512, 2), factor.weight):
# y_post.append(torch.mm(z.view(N*T, -1), w).view(N, T, -1))
# y_post = torch.cat(y_post, 2)
# else:
# y_post = torch.mm(z_post.view(N*T, -1), factor.weight[0]).view(N, T, -1)
# z_post_hat = factor(y_post)
# f_expand = f_post.unsqueeze(1).expand(-1, T, -1)
# w_post = f_expand + z_post_hat
# fake_img, _ = generator([w_post.view(N*T, latent_full)], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
# else:
# real_lat = encoder(real_seq.view(-1, *shape[2:]))
# # single head: f_post [N, latent_full]; y_post [N, T, D]
# # multi head: f_post [N, n_latent, latent]; y_post [N, T, n_latent, d]
# f_post, y_post = posterior(real_lat.view(N, T, latent_full))
# z_post = factor(y_post)
# f_expand = f_post.unsqueeze(1).expand(-1, T, -1)
# w_post = f_expand + z_post # shape [N, T, latent_full]
# fake_img, _ = generator([w_post.view(N*T, latent_full)], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
# TODO: sample frames
real_img = real_seq.view(N*T, *shape[2:])
# fake_img = fake_seq.view(N*T, *shape[2:])
if args.lambda_adv > 0:
if args.augment:
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
fake_img_aug = fake_img
fake_pred = discriminator(fake_img_aug)
adv_loss = g_nonsaturating_loss(fake_pred)
# TODO: do we always put pix and vgg loss for all frames?
if args.lambda_pix > 0:
pix_loss = torch.mean((real_img - fake_img) ** 2)
if args.lambda_vgg > 0:
real_feat = vggnet(real_img)
fake_feat = vggnet(fake_img)
vgg_loss = torch.mean((real_feat - fake_feat) ** 2)
# Train Encoder with video-level objectives
# TODO: video adversarial loss
if args.lambda_vid > 0:
fake_pred = discriminator3d(flip_video(fake_seq.transpose(1, 2)))
vid_loss = criterionGAN(fake_pred, True)
if args.lambda_l1y > 0:
# l1y_loss = criterionL1(y_post)
l1y_loss = torch.mean(torch.abs(y_post))
e_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
e_loss = e_loss + args.lambda_vid * vid_loss + args.lambda_l1y * l1y_loss
loss_dict["e"] = e_loss
loss_dict["pix"] = pix_loss
loss_dict["vgg"] = vgg_loss
loss_dict["adv"] = adv_loss
if not args.no_update_encoder:
encoder.zero_grad()
posterior.zero_grad()
e_loss.backward()
q_optim.step()
if not args.no_update_encoder:
e_optim.step()
# if args.train_on_fake:
# e_regularize = args.e_rec_every > 0 and i % args.e_rec_every == 0
# if e_regularize and args.lambda_rec > 0:
# noise = mixing_noise(args.batch, args.latent, args.mixing, device)
# fake_img, latent_fake = generator(noise, input_is_latent=False, return_latents=True)
# latent_pred = encoder(fake_img)
# if latent_pred.ndim < 3:
# latent_pred = latent_pred.unsqueeze(1).repeat(1, latent_fake.size(1), 1)
# rec_loss = torch.mean((latent_fake - latent_pred) ** 2)
# encoder.zero_grad()
# (rec_loss * args.lambda_rec).backward()
# e_optim.step()
# loss_dict["rec"] = rec_loss
# e_regularize = args.e_reg_every > 0 and i % args.e_reg_every == 0
# if e_regularize:
# # why not regularize on augmented real?
# real_img.requires_grad = True
# real_pred = encoder(real_img)
# r1_loss_e = d_r1_loss(real_pred, real_img)
# encoder.zero_grad()
# (args.r1 / 2 * r1_loss_e * args.e_reg_every + 0 * real_pred.view(-1)[0]).backward()
# e_optim.step()
# loss_dict["r1_e"] = r1_loss_e
if not args.no_update_encoder:
if not args.no_ema and e_ema is not None:
accumulate(e_ema, e_module, accum)
# Train Discriminator
if args.toggle_grads:
requires_grad(encoder, False)
requires_grad(discriminator, True)
fake_img, fake_seq = reconstruct_sequence(args, real_seq, encoder, generator, factor, posterior)
# if args.debug == 'no_lstm':
# real_lat = encoder(real_seq.view(-1, *shape[2:]))
# fake_img, _ = generator([real_lat], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
# elif args.debug == 'decomp':
# real_lat = encoder(real_seq.view(-1, *shape[2:])) # [N*T, latent_full]
# f_post = real_lat[::T, ...]
# z_post = real_lat.view(N, T, -1) - f_post.unsqueeze(1)
# if args.use_multi_head:
# y_post = []
# for z, w in zip(torch.split(z_post, 512, 2), factor.weight):
# y_post.append(torch.mm(z.view(N*T, -1), w).view(N, T, -1))
# y_post = torch.cat(y_post, 2)
# else:
# y_post = torch.mm(z_post.view(N*T, -1), factor.weight[0]).view(N, T, -1)
# z_post_hat = factor(y_post)
# f_expand = f_post.unsqueeze(1).expand(-1, T, -1)
# w_post = f_expand + z_post_hat
# fake_img, _ = generator([w_post.view(N*T, latent_full)], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
# elif args.debug == 'coef':
# real_lat = encoder(real_seq.view(-1, *shape[2:])) # [N*T, latent_full]
# f_post = real_lat[::T, ...]
# z_post_hat = real_lat.view(N, T, -1) - f_post.unsqueeze(1)
# y_post = torch.mm(z_post_hat.view(N*T, -1), factor.weight).view(N, T, -1)
# z_post = factor(y_post)
# f_expand = f_post.unsqueeze(1).expand(-1, T, -1)
# w_post = f_expand + z_post
# fake_img, _ = generator([w_post.view(N*T, latent_full)], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
# else:
# real_lat = encoder(real_seq.view(-1, *shape[2:]))
# f_post, y_post = posterior(real_lat.view(N, T, latent_full))
# z_post = factor(y_post)
# f_expand = f_post.unsqueeze(1).expand(-1, T, -1)
# w_post = f_expand + z_post # shape [N, T, latent_full]
# fake_img, _ = generator([w_post.view(N*T, latent_full)], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
# fake_img = fake_seq.view(N*T, *shape[2:])
if not args.no_update_discriminator:
if args.lambda_adv > 0:
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_img_aug = fake_img
fake_pred = discriminator(fake_img_aug)
real_pred = discriminator(real_img_aug)
d_loss = d_logistic_loss(real_pred, fake_pred)
# Train video discriminator
if args.lambda_vid > 0:
pred_real = discriminator3d(flip_video(real_seq.transpose(1, 2)))
pred_fake = discriminator3d(flip_video(fake_seq.transpose(1, 2)))
dv_loss_real = criterionGAN(pred_real, True)
dv_loss_fake = criterionGAN(pred_fake, False)
dv_loss = 0.5 * (dv_loss_real + dv_loss_fake)
d_loss = d_loss + dv_loss
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
if args.lambda_adv > 0:
discriminator.zero_grad()
if args.lambda_vid > 0:
discriminator3d.zero_grad()
d_loss.backward()
if args.lambda_adv > 0:
d_optim.step()
if args.lambda_vid > 0:
dv_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = args.d_reg_every > 0 and i % args.d_reg_every == 0
if d_regularize:
# why not regularize on augmented real?
real_img.requires_grad = True
real_pred = discriminator(real_img)
r1_loss_d = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss_d * args.d_reg_every + 0 * real_pred.view(-1)[0]).backward()
# Why 0* ? Answer is here https://github.com/rosinality/stylegan2-pytorch/issues/76
d_optim.step()
loss_dict["r1_d"] = r1_loss_d
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
e_loss_val = loss_reduced["e"].mean().item()
r1_d_val = loss_reduced["r1_d"].mean().item()
r1_e_val = loss_reduced["r1_e"].mean().item()
pix_loss_val = loss_reduced["pix"].mean().item()
vgg_loss_val = loss_reduced["vgg"].mean().item()
adv_loss_val = loss_reduced["adv"].mean().item()
rec_loss_val = loss_reduced["rec"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
if get_rank() == 0:
pbar.set_description(
(
f"d: {d_loss_val:.4f}; e: {e_loss_val:.4f}; r1_d: {r1_d_val:.4f}; r1_e: {r1_e_val:.4f}; "
f"pix: {pix_loss_val:.4f}; vgg: {vgg_loss_val:.4f}; adv: {adv_loss_val:.4f}; "
f"rec: {rec_loss_val:.4f}; augment: {ada_aug_p:.4f}"
)
)
if wandb and args.wandb:
wandb.log(
{
"Encoder": e_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1 D": r1_d_val,
"R1 E": r1_e_val,
"Pix Loss": pix_loss_val,
"VGG Loss": vgg_loss_val,
"Adv Loss": adv_loss_val,
"Rec Loss": rec_loss_val,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
}
)
if i % args.log_every == 0:
with torch.no_grad():
e_eval = encoder if args.no_ema else e_ema
e_eval.eval()
posterior.eval()
# N = sample_x.shape[0]
fake_img, fake_seq = reconstruct_sequence(args, sample_x, e_eval, generator, factor, posterior)
# if args.debug == 'no_lstm':
# real_lat = encoder(sample_x.view(-1, *shape[2:]))
# fake_img, _ = generator([real_lat], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
# elif args.debug == 'decomp':
# real_lat = encoder(sample_x.view(-1, *shape[2:])) # [N*T, latent_full]
# f_post = real_lat[::T, ...]
# z_post = real_lat.view(N, T, -1) - f_post.unsqueeze(1)
# if args.use_multi_head:
# y_post = []
# for z, w in zip(torch.split(z_post, 512, 2), factor.weight):
# y_post.append(torch.mm(z.view(N*T, -1), w).view(N, T, -1))
# y_post = torch.cat(y_post, 2)
# else:
# y_post = torch.mm(z_post.view(N*T, -1), factor.weight[0]).view(N, T, -1)
# z_post_hat = factor(y_post)
# f_expand = f_post.unsqueeze(1).expand(-1, T, -1)
# w_post = f_expand + z_post_hat
# fake_img, _ = generator([w_post.view(N*T, latent_full)], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
# else:
# x_lat = encoder(sample_x.view(-1, *shape[2:]))
# f_post, y_post = posterior(x_lat.view(N, T, latent_full))
# z_post = factor(y_post)
# f_expand = f_post.unsqueeze(1).expand(-1, T, -1)
# w_post = f_expand + z_post
# fake_img, _ = generator([w_post.view(N*T, latent_full)], input_is_latent=True, return_latents=False)
# fake_seq = fake_img.view(N, T, *shape[2:])
utils.save_image(
torch.cat((sample_x, fake_seq), 1).view(-1, *shape[2:]),
os.path.join(args.log_dir, 'sample', f"{str(i).zfill(6)}-img_recon.png"),
nrow=T,
normalize=True,
value_range=(-1, 1),
)
util.save_video(
fake_seq[random.randint(0, args.n_sample-1)],
os.path.join(args.log_dir, 'sample', f"{str(i).zfill(6)}-vid_recon.mp4")
)
fake_img, fake_seq = swap_sequence(args, sample_x, e_eval, generator, factor, posterior)
utils.save_image(
torch.cat((sample_x, fake_seq), 1).view(-1, *shape[2:]),
os.path.join(args.log_dir, 'sample', f"{str(i).zfill(6)}-img_swap.png"),
nrow=T,
normalize=True,
value_range=(-1, 1),
)
e_eval.train()
posterior.train()
if i % args.save_every == 0:
e_eval = encoder if args.no_ema else e_ema
torch.save(
{
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_module.state_dict(),
"e_ema": e_eval.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"{str(i).zfill(6)}.pt"),
)
if not args.debug and i % args.save_latest_every == 0:
torch.save(
{
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_module.state_dict(),
"e_ema": e_eval.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, generator, discriminator, g_optim, d_optim, g_ema, device):
inception = real_mean = real_cov = mean_latent = None
if args.eval_every > 0:
inception = nn.DataParallel(load_patched_inception_v3()).to(device)
inception.eval()
with open(args.inception, "rb") as f:
embeds = pickle.load(f)
real_mean = embeds["mean"]
real_cov = embeds["cov"]
if get_rank() == 0:
if args.eval_every > 0:
with open(os.path.join(args.log_dir, 'log_fid.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
if args.log_every > 0:
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar, initial=args.start_iter, dynamic_ncols=True, smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
r1_loss = torch.tensor(0.0, device=device)
g_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
loss_dict = {}
if args.distributed:
g_module = generator.module
d_module = discriminator.module
else:
g_module = generator
d_module = discriminator
# accum = 0.5 ** (32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, args.ada_every, device)
args.n_sheets = int(np.ceil(args.n_classes / args.n_class_per_sheet))
args.n_sample_per_sheet = args.n_sample_per_class * args.n_class_per_sheet
args.n_sample = args.n_sample_per_sheet * args.n_sheets
sample_z = torch.randn(args.n_sample, args.latent, device=device)
sample_y = torch.arange(args.n_classes).repeat(args.n_sample_per_class, 1).t().reshape(-1).to(device)
if args.n_sample > args.n_sample_per_class * args.n_classes:
sample_y1 = make_fake_label(args.n_sample - args.n_sample_per_class * args.n_classes, args.n_classes, device)
sample_y = torch.cat([sample_y, sample_y1], 0)
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
# Train Discriminator
requires_grad(generator, False)
requires_grad(discriminator, True)
for step_index in range(args.n_step_d):
real_img, real_labels = next(loader)
real_img, real_labels = real_img.to(device), real_labels.to(device)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_labels = make_fake_label(args.batch, args.n_classes, device)
fake_img, _ = generator(noise, fake_labels)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_pred = discriminator(fake_img, fake_labels)
real_pred = discriminator(real_img_aug, real_labels)
d_loss = d_logistic_loss(real_pred, fake_pred)
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
else:
real_img_aug = real_img
real_pred = discriminator(real_img_aug, real_labels)
r1_loss = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every + 0 * real_pred[0]).backward()
d_optim.step()
loss_dict["r1"] = r1_loss
# Train Generator
requires_grad(generator, True)
requires_grad(discriminator, False)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_labels = make_fake_label(args.batch, args.n_classes, device)
fake_img, _ = generator(noise, fake_labels)
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = discriminator(fake_img, fake_labels)
g_loss = g_nonsaturating_loss(fake_pred)
loss_dict["g"] = g_loss
generator.zero_grad()
g_loss.backward()
g_optim.step()
g_regularize = args.g_reg_every > 0 and i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing, device)
fake_labels = make_fake_label(args.batch, args.n_classes, device)
fake_img, latents = generator(noise, fake_labels, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length
)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
mean_path_length_avg = (
reduce_sum(mean_path_length).item() / get_world_size()
)
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
# Update G_ema
# G_ema = G * (1-ema_beta) + G_ema * ema_beta
ema_nimg = args.ema_kimg * 1000
if args.ema_rampup is not None:
ema_nimg = min(ema_nimg, i * args.batch * args.ema_rampup)
accum = 0.5 ** (args.batch / max(ema_nimg, 1e-8))
accumulate(g_ema, g_module, accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
g_loss_val = loss_reduced["g"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
if get_rank() == 0:
pbar.set_description(
(
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}"
)
)
if wandb and args.wandb:
wandb.log(
{
"Generator": g_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1": r1_val,
"Path Length Regularization": path_loss_val,
"Mean Path Length": mean_path_length,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
"Path Length": path_length_val,
}
)
if i % args.log_every == 0:
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(
(
f"{i:07d}; "
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f};\n"
)
)
if i % args.log_every == 0:
with torch.no_grad():
g_ema.eval()
for sheet_index in range(args.n_sheets):
sample_z_sheet = sample_z[sheet_index*args.n_sample_per_sheet:(sheet_index+1)*args.n_sample_per_sheet]
sample_y_sheet = sample_y[sheet_index*args.n_sample_per_sheet:(sheet_index+1)*args.n_sample_per_sheet]
sample, _ = g_ema([sample_z_sheet], sample_y_sheet)
utils.save_image(
sample,
os.path.join(args.log_dir, 'sample', f"{str(i).zfill(6)}_{sheet_index}.png"),
nrow=args.n_sample_per_class,
normalize=True,
value_range=(-1, 1),
)
if args.eval_every > 0 and i % args.eval_every == 0:
with torch.no_grad():
g_ema.eval()
if args.truncation < 1:
mean_latent = g_ema.mean_latent(4096)
features = extract_feature_from_samples(
g_ema, inception, args.truncation, mean_latent, 64, args.n_sample_fid, args.device,
n_classes=args.n_classes,
).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid = calc_fid(sample_mean, sample_cov, real_mean, real_cov)
# print("fid:", fid)
with open(os.path.join(args.log_dir, 'log_fid.txt'), 'a+') as f:
f.write(f"{i:07d}; fid: {float(fid):.4f};\n")
if i % args.save_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, loader2, generator, encoder, discriminator, vggnet,
g_optim, e_optim, d_optim, g_ema, e_ema, device):
inception = real_mean = real_cov = mean_latent = None
if args.eval_every > 0:
inception = nn.DataParallel(load_patched_inception_v3()).to(device)
inception.eval()
with open(args.inception, "rb") as f:
embeds = pickle.load(f)
real_mean = embeds["mean"]
real_cov = embeds["cov"]
if get_rank() == 0:
if args.eval_every > 0:
with open(os.path.join(args.log_dir, 'log_fid.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
if args.log_every > 0:
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
r1_loss = torch.tensor(0.0, device=device)
g_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
d_loss_val = r1_val = real_score_val = recx_score_val = 0
loss_dict = {
"d": torch.tensor(0.0, device=device),
"r1": torch.tensor(0.0, device=device)
}
avg_pix_loss = util.AverageMeter()
avg_vgg_loss = util.AverageMeter()
if args.distributed:
g_module = generator.module
e_module = encoder.module
d_module = discriminator.module
else:
g_module = generator
e_module = encoder
d_module = discriminator
d_weight = torch.tensor(1.0, device=device)
last_layer = None
if args.use_adaptive_weight:
if args.distributed:
last_layer = generator.module.get_last_layer()
else:
last_layer = generator.get_last_layer()
# accum = 0.5 ** (32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
r_t_dict = {'real': 0, 'recx': 0} # r_t stat
g_scale = 1
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length,
args.ada_every, device)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
sample_x = load_real_samples(args, loader)
if sample_x.ndim > 4:
sample_x = sample_x[:, 0, ...]
n_step_max = max(args.n_step_d, args.n_step_e)
requires_grad(g_ema, False)
requires_grad(e_ema, False)
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
if args.debug: util.seed_everything(i)
real_imgs = [next(loader).to(device) for _ in range(n_step_max)]
# Train Discriminator
if args.lambda_adv > 0:
requires_grad(generator, False)
requires_grad(encoder, False)
requires_grad(discriminator, True)
for step_index in range(args.n_step_d):
real_img = real_imgs[step_index]
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
rec_img_aug, _ = augment(rec_img, ada_aug_p)
else:
real_img_aug = real_img
rec_img_aug = rec_img
real_pred = discriminator(real_img_aug)
rec_pred = discriminator(rec_img_aug)
d_loss_real = F.softplus(-real_pred).mean()
d_loss_rec = F.softplus(rec_pred).mean()
loss_dict["real_score"] = real_pred.mean()
loss_dict["recx_score"] = rec_pred.mean()
d_loss = d_loss_real + d_loss_rec * args.lambda_rec_d
loss_dict["d"] = d_loss
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
# Compute batchwise r_t
r_t_dict['real'] = torch.sign(real_pred).sum().item() / args.batch
d_regularize = i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
else:
real_img_aug = real_img
real_pred = discriminator(real_img_aug)
r1_loss = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every +
0 * real_pred[0]).backward()
d_optim.step()
loss_dict["r1"] = r1_loss
r_t_dict['recx'] = torch.sign(rec_pred).sum().item() / args.batch
# Train AutoEncoder
requires_grad(encoder, True)
requires_grad(generator, True)
requires_grad(discriminator, False)
if args.debug: util.seed_everything(i)
pix_loss = vgg_loss = adv_loss = torch.tensor(0., device=device)
for step_index in range(args.n_step_e):
real_img = real_imgs[step_index]
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
if args.lambda_pix > 0:
if args.pix_loss == 'l2':
pix_loss = torch.mean((rec_img - real_img)**2)
elif args.pix_loss == 'l1':
pix_loss = F.l1_loss(rec_img, real_img)
if args.lambda_vgg > 0:
vgg_loss = torch.mean((vggnet(real_img) - vggnet(rec_img))**2)
if args.lambda_adv > 0:
if args.augment:
rec_img_aug, _ = augment(rec_img, ada_aug_p)
else:
rec_img_aug = rec_img
rec_pred = discriminator(rec_img_aug)
adv_loss = g_nonsaturating_loss(rec_pred)
if args.use_adaptive_weight and i >= args.disc_iter_start:
nll_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg
g_loss = adv_loss * args.lambda_adv
d_weight = calculate_adaptive_weight(nll_loss,
g_loss,
last_layer=last_layer)
ae_loss = (pix_loss * args.lambda_pix +
vgg_loss * args.lambda_vgg +
d_weight * adv_loss * args.lambda_adv)
loss_dict["ae"] = ae_loss
loss_dict["pix"] = pix_loss
loss_dict["vgg"] = vgg_loss
loss_dict["adv"] = adv_loss
encoder.zero_grad()
generator.zero_grad()
ae_loss.backward()
e_optim.step()
if args.g_decay is not None:
scale_grad(generator, g_scale)
g_scale *= args.g_decay
g_optim.step()
g_regularize = args.g_reg_every > 0 and i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing,
device)
fake_img, latents = generator(noise, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
mean_path_length_avg = (reduce_sum(mean_path_length).item() /
get_world_size())
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
# Update EMA
ema_nimg = args.ema_kimg * 1000
if args.ema_rampup is not None:
ema_nimg = min(ema_nimg, i * args.batch * args.ema_rampup)
accum = 0.5**(args.batch / max(ema_nimg, 1e-8))
accumulate(g_ema, g_module, 0 if args.no_ema_g else accum)
accumulate(e_ema, e_module, 0 if args.no_ema_e else accum)
loss_reduced = reduce_loss_dict(loss_dict)
ae_loss_val = loss_reduced["ae"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
pix_loss_val = loss_reduced["pix"].mean().item()
vgg_loss_val = loss_reduced["vgg"].mean().item()
adv_loss_val = loss_reduced["adv"].mean().item()
if args.lambda_adv > 0:
d_loss_val = loss_reduced["d"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
recx_score_val = loss_reduced["recx_score"].mean().item()
avg_pix_loss.update(pix_loss_val, real_img.shape[0])
avg_vgg_loss.update(vgg_loss_val, real_img.shape[0])
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; ae: {ae_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}; "
f"d_weight: {d_weight.item():.4f}; "
f"pix: {pix_loss_val:.4f}; vgg: {vgg_loss_val:.4f}; adv: {adv_loss_val:.4f}"
))
if i % args.log_every == 0:
with torch.no_grad():
g_ema.eval()
e_ema.eval()
nrow = int(args.n_sample**0.5)
nchw = list(sample_x.shape)[1:]
# Reconstruction of real images
latent_x, _ = e_ema(sample_x)
rec_real, _ = g_ema([latent_x], input_is_latent=True)
sample = torch.cat(
(sample_x.reshape(args.n_sample // nrow, nrow, *nchw),
rec_real.reshape(args.n_sample // nrow, nrow, *nchw)),
1)
utils.save_image(
sample.reshape(2 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-recon.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
ref_pix_loss = torch.sum(torch.abs(sample_x - rec_real))
ref_vgg_loss = torch.mean(
(vggnet(sample_x) -
vggnet(rec_real))**2) if vggnet is not None else 0
# Fixed fake samples and reconstructions
sample, _ = g_ema([sample_z])
utils.save_image(
sample,
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-sample.png"),
nrow=int(args.n_sample**0.5),
normalize=True,
value_range=(-1, 1),
)
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write((
f"{i:07d}; "
f"d: {d_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean_path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}; {'; '.join([f'{k}: {r_t_dict[k]:.4f}' for k in r_t_dict])}; "
f"real_score: {real_score_val:.4f}; recx_score: {recx_score_val:.4f}; "
f"pix: {avg_pix_loss.avg:.4f}; vgg: {avg_vgg_loss.avg:.4f}; "
f"ref_pix: {ref_pix_loss.item():.4f}; ref_vgg: {ref_vgg_loss.item():.4f}; "
f"d_weight: {d_weight.item():.4f}; "
f"\n"))
if wandb and args.wandb:
wandb.log({
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1": r1_val,
"Path Length Regularization": path_loss_val,
"Mean Path Length": mean_path_length,
"Real Score": real_score_val,
"Path Length": path_length_val,
})
if args.eval_every > 0 and i % args.eval_every == 0:
with torch.no_grad():
fid_sa = fid_re = fid_sr = 0
g_ema.eval()
e_ema.eval()
if args.truncation < 1:
mean_latent = g_ema.mean_latent(4096)
# Real reconstruction FID
if 'fid_recon' in args.which_metric:
features = extract_feature_from_reconstruction(
e_ema,
g_ema,
inception,
args.truncation,
mean_latent,
loader2,
args.device,
mode='recon',
).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid_re = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
with open(os.path.join(args.log_dir, 'log_fid.txt'),
'a+') as f:
f.write(f"{i:07d}; rec_real: {float(fid_re):.4f};\n")
if i % args.save_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight',
f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, encoder, generator, discriminator, vggnet, pwcnet,
e_optim, d_optim, e_ema, device):
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
d_loss_val = 0
e_loss_val = 0
rec_loss_val = 0
vgg_loss_val = 0
adv_loss_val = 0
loss_dict = {
"d": torch.tensor(0., device=device),
"real_score": torch.tensor(0., device=device),
"fake_score": torch.tensor(0., device=device),
"r1_d": torch.tensor(0., device=device),
"r1_e": torch.tensor(0., device=device),
"rec": torch.tensor(0., device=device),
}
avg_pix_loss = util.AverageMeter()
avg_vgg_loss = util.AverageMeter()
if args.distributed:
e_module = encoder.module
d_module = discriminator.module
g_module = generator.module
else:
e_module = encoder
d_module = discriminator
g_module = generator
accum = 0.5**(32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 256,
device)
sample_x = accumulate_batches(loader, args.n_sample).to(device)
requires_grad(generator, False) # always False
generator.eval() # Generator should be ema and in eval mode
# if args.no_ema or e_ema is None:
# e_ema = encoder
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
real_img1, real_img2 = next(loader)
real_img1 = real_img1.to(device)
real_img2 = real_img2.to(device)
# Train Encoder
if args.toggle_grads:
requires_grad(encoder, True)
requires_grad(discriminator, False)
pix_loss = vgg_loss = adv_loss = rec_loss = torch.tensor(0.,
device=device)
latent_real1 = encoder(real_img1)
fake_img1, _ = generator([latent_real1],
input_is_latent=True,
return_latents=False)
latent_real2 = encoder(real_img2)
fake_img2, _ = generator([latent_real2],
input_is_latent=True,
return_latents=False)
if args.lambda_adv > 0:
# if args.augment:
# fake_img_aug1, _ = augment(fake_img1, ada_aug_p)
# else:
# fake_img_aug1 = fake_img1
fake_img_pair = torch.cat((fake_img1, fake_img2), 1)
fake_pred = discriminator(fake_img_pair)
adv_loss = g_nonsaturating_loss(fake_pred)
if args.lambda_pix > 0:
pix_loss = torch.mean((real_img1 - fake_img1)**2)
if args.reconstruct_pair:
pix_loss += torch.mean((real_img2 - fake_img2)**2)
if args.lambda_vgg > 0:
real_feat = vggnet(real_img1)
fake_feat = vggnet(fake_img1)
vgg_loss = torch.mean((real_feat - fake_feat)**2)
if args.reconstruct_pair:
real_feat = vggnet(real_img2)
fake_feat = vggnet(fake_img2)
vgg_loss += torch.mean((real_feat - fake_feat)**2)
e_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
loss_dict["e"] = e_loss
loss_dict["pix"] = pix_loss
loss_dict["vgg"] = vgg_loss
loss_dict["adv"] = adv_loss
encoder.zero_grad()
e_loss.backward()
e_optim.step()
# if args.train_on_fake:
# e_regularize = args.e_rec_every > 0 and i % args.e_rec_every == 0
# if e_regularize and args.lambda_rec > 0:
# noise = mixing_noise(args.batch, args.latent, args.mixing, device)
# fake_img, latent_fake = generator(noise, input_is_latent=False, return_latents=True)
# latent_pred = encoder(fake_img)
# if latent_pred.ndim < 3:
# latent_pred = latent_pred.unsqueeze(1).repeat(1, latent_fake.size(1), 1)
# rec_loss = torch.mean((latent_fake - latent_pred) ** 2)
# encoder.zero_grad()
# (rec_loss * args.lambda_rec).backward()
# e_optim.step()
# loss_dict["rec"] = rec_loss
e_regularize = args.e_reg_every > 0 and i % args.e_reg_every == 0
if e_regularize:
# why not regularize on augmented real?
real_img_pair = torch.cat((real_img1, real_img2), 1)
real_img_pair.requires_grad = True
real_pred = encoder(real_img_pair)
r1_loss_e = d_r1_loss(real_pred, real_img_pair)
encoder.zero_grad()
(args.r1 / 2 * r1_loss_e * args.e_reg_every +
0 * real_pred.view(-1)[0]).backward()
e_optim.step()
loss_dict["r1_e"] = r1_loss_e
if not args.no_ema and e_ema is not None:
accumulate(e_ema, e_module, accum)
# Train Discriminator
if args.toggle_grads:
requires_grad(encoder, False)
requires_grad(discriminator, True)
if not args.no_update_discriminator and args.lambda_adv > 0:
latent_real1 = encoder(real_img1)
fake_img1, _ = generator([latent_real1],
input_is_latent=True,
return_latents=False)
latent_real2 = encoder(real_img2)
fake_img2, _ = generator([latent_real2],
input_is_latent=True,
return_latents=False)
# if args.augment:
# real_img_aug, _ = augment(real_img, ada_aug_p)
# fake_img_aug, _ = augment(fake_img, ada_aug_p)
# else:
# real_img_aug = real_img
# fake_img_aug = fake_img
fake_img_pair = torch.cat((fake_img1, fake_img2), 1)
real_img_pair = torch.cat((real_img1, real_img2), 1)
fake_pred = discriminator(fake_img_pair)
real_pred = discriminator(real_img_pair)
d_loss = d_logistic_loss(real_pred, fake_pred)
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
# if args.augment and args.augment_p == 0:
# ada_aug_p = ada_augment.tune(real_pred)
# r_t_stat = ada_augment.r_t_stat
d_regularize = args.d_reg_every > 0 and i % args.d_reg_every == 0
if d_regularize:
# why not regularize on augmented real?
real_img_pair = torch.cat((real_img1, real_img2), 1)
real_img_pair.requires_grad = True
real_pred = discriminator(real_img_pair)
r1_loss_d = d_r1_loss(real_pred, real_img_pair)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss_d * args.d_reg_every +
0 * real_pred.view(-1)[0]).backward()
# Why 0* ? Answer is here https://github.com/rosinality/stylegan2-pytorch/issues/76
d_optim.step()
loss_dict["r1_d"] = r1_loss_d
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
e_loss_val = loss_reduced["e"].mean().item()
r1_d_val = loss_reduced["r1_d"].mean().item()
r1_e_val = loss_reduced["r1_e"].mean().item()
pix_loss_val = loss_reduced["pix"].mean().item()
vgg_loss_val = loss_reduced["vgg"].mean().item()
adv_loss_val = loss_reduced["adv"].mean().item()
rec_loss_val = loss_reduced["rec"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
avg_pix_loss.update(pix_loss_val, real_img1.shape[0])
avg_vgg_loss.update(vgg_loss_val, real_img1.shape[0])
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; e: {e_loss_val:.4f}; r1_d: {r1_d_val:.4f}; r1_e: {r1_e_val:.4f}; "
f"pix: {pix_loss_val:.4f}; vgg: {vgg_loss_val:.4f}; adv: {adv_loss_val:.4f}; "
f"rec: {rec_loss_val:.4f}; augment: {ada_aug_p:.4f}"))
if i % args.log_every == 0:
with torch.no_grad():
latent_x = e_ema(sample_x)
fake_x, _ = generator([latent_x],
input_is_latent=True,
return_latents=False)
sample_pix_loss = torch.sum((sample_x - fake_x)**2)
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(
f"{i:07d}; pix: {avg_pix_loss.avg}; vgg: {avg_vgg_loss.avg}; "
f"ref: {sample_pix_loss.item()};\n")
if wandb and args.wandb:
wandb.log({
"Encoder": e_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1 D": r1_d_val,
"R1 E": r1_e_val,
"Pix Loss": pix_loss_val,
"VGG Loss": vgg_loss_val,
"Adv Loss": adv_loss_val,
"Rec Loss": rec_loss_val,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
})
if i % args.log_every == 0:
with torch.no_grad():
e_eval = encoder if args.no_ema else e_ema
e_eval.eval()
nrow = int(args.n_sample**0.5)
nchw = list(sample_x.shape)[1:]
latent_real = e_eval(sample_x)
fake_img, _ = generator([latent_real],
input_is_latent=True,
return_latents=False)
sample = torch.cat(
(sample_x.reshape(args.n_sample // nrow, nrow, *nchw),
fake_img.reshape(args.n_sample // nrow, nrow, *nchw)),
1)
utils.save_image(
sample.reshape(2 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
e_eval.train()
if i % args.save_every == 0:
e_eval = encoder if args.no_ema else e_ema
torch.save(
{
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_module.state_dict(),
"e_ema": e_eval.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight',
f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_module.state_dict(),
"e_ema": e_eval.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, generator, discriminator, g_optim, d_optim, g_ema, device):
loader = sample_data(loader, datatype="imagefolder")
# inception related:
if (get_rank() == 0):
from calc_inception import load_patched_inception_v3
inception = load_patched_inception_v3().to(device)
inception.eval()
if args.eval_every > 0:
with open(os.path.join(args.log_dir, 'log_fid.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-' * 50}\n")
if args.log_every > 0:
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-' * 50}\n")
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar, initial=args.start_iter, dynamic_ncols=True, smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
r1_loss = torch.tensor(0.0, device=device)
g_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
loss_dict = {}
if args.distributed:
g_module = generator.module
d_module = discriminator.module
else:
g_module = generator
d_module = discriminator
accum = 0.5 ** (32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 8, device)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
real_img = next(loader)
real_img = real_img.to(device)
requires_grad(generator, False)
requires_grad(discriminator, True)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_pred = discriminator(fake_img)
real_pred = discriminator(real_img_aug)
d_loss = d_logistic_loss(real_pred, fake_pred)
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
else:
real_img_aug = real_img
real_pred = discriminator(real_img_aug)
r1_loss = d_r1_loss(real_pred, real_img,args)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every + 0 * real_pred[0]).backward()
d_optim.step()
loss_dict["r1"] = r1_loss
requires_grad(generator, True)
requires_grad(discriminator, False)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = discriminator(fake_img)
g_loss = g_nonsaturating_loss(fake_pred)
loss_dict["g"] = g_loss
generator.zero_grad()
g_loss.backward()
g_optim.step()
#g_reg starts
g_regularize = False
if args.useG_reg==True:
# print("I entered g_reg")
g_regularize = i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing, device)
fake_img, latents = generator(noise, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length
)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
mean_path_length_avg = (
reduce_sum(mean_path_length).item() / get_world_size()
)
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
accumulate(g_ema, g_module, accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
g_loss_val = loss_reduced["g"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
if get_rank() == 0:
pbar.set_description(
(
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}"
)
)
# inception related:
if args.eval_every > 0 and i % args.eval_every == 0:
real_mean = real_cov = mean_latent = None
with open(args.inception, "rb") as f:
embeds = pickle.load(f)
real_mean = embeds["mean"]
real_cov = embeds["cov"]
# print("yahooo!\n")
with torch.no_grad():
g_ema.eval()
if args.truncation < 1:
mean_latent = g_ema.mean_latent(4096)
# print("I am fine sir!\n")
features = extract_feature_from_samples(
g_ema, inception, args.truncation, mean_latent, 64, args.n_sample_fid, device
).numpy()
# print("I am normal sir!")
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid = calc_fid(sample_mean, sample_cov, real_mean, real_cov)
with open(os.path.join(args.log_dir, 'log_fid.txt'), 'a+') as f:
f.write(f"{i:07d}; fid: {float(fid):.4f};\n")
# print("alright hurray \n")
if i % args.log_every == 0:
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(
(
f"{i:07d}; "
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f};\n"
)
)
if i % args.log_every == 0:
with torch.no_grad():
g_ema.eval()
sample, _ = g_ema([sample_z])
utils.save_image(
sample,
os.path.join(args.log_dir, 'sample', f"{str(i).zfill(6)}.png"),
nrow=int(args.n_sample ** 0.5),
normalize=True,
range=(-1, 1),
)
if i % args.save_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, loader2, generator, encoder, discriminator,
discriminator2, vggnet, g_optim, e_optim, d_optim, d2_optim, g_ema,
e_ema, device):
inception = real_mean = real_cov = mean_latent = None
if args.eval_every > 0:
inception = nn.DataParallel(load_patched_inception_v3()).to(device)
inception.eval()
with open(args.inception, "rb") as f:
embeds = pickle.load(f)
real_mean = embeds["mean"]
real_cov = embeds["cov"]
if get_rank() == 0:
if args.eval_every > 0:
with open(os.path.join(args.log_dir, 'log_fid.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
if args.log_every > 0:
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
r1_loss = torch.tensor(0.0, device=device)
g_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
loss_dict = {
'recx_score': torch.tensor(0.0, device=device),
'ae_fake': torch.tensor(0.0, device=device),
'ae_real': torch.tensor(0.0, device=device),
'pix': torch.tensor(0.0, device=device),
'vgg': torch.tensor(0.0, device=device),
}
avg_pix_loss = util.AverageMeter()
avg_vgg_loss = util.AverageMeter()
if args.distributed:
g_module = generator.module
e_module = encoder.module
d_module = discriminator.module
else:
g_module = generator
e_module = encoder
d_module = discriminator
d2_module = None
if discriminator2 is not None:
if args.distributed:
d2_module = discriminator2.module
else:
d2_module = discriminator2
# When joint training enabled, d_weight balances reconstruction loss and adversarial loss on
# recontructed real images. This does not balance the overall AE loss and GAN loss.
d_weight = torch.tensor(1.0, device=device)
last_layer = None
if args.use_adaptive_weight:
if args.distributed:
last_layer = generator.module.get_last_layer()
else:
last_layer = generator.get_last_layer()
g_scale = 1
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
r_t_dict = {'real': 0, 'fake': 0, 'recx': 0} # r_t stat
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length,
args.ada_every, device)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
sample_x = load_real_samples(args, loader)
if sample_x.ndim > 4:
sample_x = sample_x[:, 0, ...]
input_is_latent = args.latent_space != 'z' # Encode in z space?
n_step_max = max(args.n_step_d, args.n_step_e)
requires_grad(g_ema, False)
requires_grad(e_ema, False)
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
if args.debug: util.seed_everything(i)
real_imgs = [next(loader).to(device) for _ in range(n_step_max)]
# Train Discriminator and Encoder
requires_grad(generator, False)
requires_grad(encoder, True)
requires_grad(discriminator, True)
requires_grad(discriminator2, True)
for step_index in range(args.n_step_d):
real_img = real_imgs[step_index]
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_img_aug = fake_img
real_pred = discriminator(encoder(real_img_aug)[0])
fake_pred = discriminator(encoder(fake_img_aug)[0])
d_loss_real = F.softplus(-real_pred).mean()
d_loss_fake = F.softplus(fake_pred).mean()
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
d_loss_rec = 0.
if args.lambda_rec_d > 0:
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real],
input_is_latent=input_is_latent)
if args.augment:
rec_img, _ = augment(rec_img, ada_aug_p)
rec_pred = discriminator(encoder(rec_img)[0])
d_loss_rec = F.softplus(rec_pred).mean()
loss_dict["recx_score"] = rec_pred.mean()
r_t_dict['recx'] = torch.sign(
rec_pred).sum().item() / args.batch
d_loss = d_loss_real + d_loss_fake * args.lambda_fake_d + d_loss_rec * args.lambda_rec_d
loss_dict["d"] = d_loss
discriminator.zero_grad()
encoder.zero_grad()
d_loss.backward()
d_optim.step()
e_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
r_t_dict['real'] = torch.sign(real_pred).sum().item() / args.batch
r_t_dict['fake'] = torch.sign(fake_pred).sum().item() / args.batch
d_regularize = args.d_reg_every > 0 and i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
else:
real_img_aug = real_img
real_pred = discriminator(encoder(real_img_aug)[0])
r1_loss = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
encoder.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every +
0 * real_pred[0]).backward()
d_optim.step()
e_optim.step()
loss_dict["r1"] = r1_loss
# Train Generator
requires_grad(generator, True)
requires_grad(encoder, False)
requires_grad(discriminator, False)
requires_grad(discriminator2, False)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
fake_img_aug = fake_img
fake_pred = discriminator(encoder(fake_img_aug)[0])
g_loss_fake = g_nonsaturating_loss(fake_pred)
loss_dict["g"] = g_loss_fake
generator.zero_grad()
(g_loss_fake * args.lambda_fake_g).backward()
g_optim.step()
g_regularize = args.g_reg_every > 0 and i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing,
device)
fake_img, latents = generator(noise, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
mean_path_length_avg = (reduce_sum(mean_path_length).item() /
get_world_size())
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
# Train Encoder (and Generator)
joint = (not args.no_joint) and (g_scale > 1e-6)
# Train AE on fake samples (latent reconstruction)
if args.lambda_rec_w + (args.lambda_pix_fake + args.lambda_vgg_fake +
args.lambda_adv_fake) > 0:
requires_grad(encoder, True)
requires_grad(generator, joint)
requires_grad(discriminator, False)
requires_grad(discriminator2, False)
for step_index in range(args.n_step_e):
# mixing_prob = 0 if args.which_latent == 'w_tied' else args.mixing
# noise = mixing_noise(args.batch, args.latent, mixing_prob, device)
# fake_img, latent_fake = generator(noise, return_latents=True, detach_style=not args.no_detach_style)
# if args.which_latent == 'w_tied':
# latent_fake = latent_fake[:,0,:]
# else:
# latent_fake = latent_fake.view(args.batch, -1)
# latent_pred, _ = encoder(fake_img)
# ae_loss_fake = torch.mean((latent_pred - latent_fake.detach()) ** 2)
ae_loss_fake = 0
mixing_prob = 0 if args.which_latent == 'w_tied' else args.mixing
if args.lambda_rec_w > 0:
noise = mixing_noise(args.batch, args.latent, mixing_prob,
device)
fake_img, latent_fake = generator(
noise,
return_latents=True,
detach_style=not args.no_detach_style)
if args.which_latent == 'w_tied':
latent_fake = latent_fake[:, 0, :]
else:
latent_fake = latent_fake.view(args.batch, -1)
latent_pred, _ = encoder(fake_img)
ae_loss_fake = torch.mean(
(latent_pred - latent_fake.detach())**2)
if args.lambda_pix_fake + args.lambda_vgg_fake + args.lambda_adv_fake > 0:
pix_loss = vgg_loss = adv_loss = torch.tensor(
0., device=device)
noise = mixing_noise(args.batch, args.latent, mixing_prob,
device)
fake_img, _ = generator(noise, detach_style=False)
fake_img = fake_img.detach()
latent_pred, _ = encoder(fake_img)
rec_img, _ = generator([latent_pred],
input_is_latent=input_is_latent)
if args.lambda_pix_fake > 0:
if args.pix_loss == 'l2':
pix_loss = torch.mean((rec_img - fake_img)**2)
elif args.pix_loss == 'l1':
pix_loss = F.l1_loss(rec_img, fake_img)
if args.lambda_vgg_fake > 0:
vgg_loss = torch.mean(
(vggnet(fake_img) - vggnet(rec_img))**2)
ae_loss_fake = (ae_loss_fake +
pix_loss * args.lambda_pix_fake +
vgg_loss * args.lambda_vgg_fake)
loss_dict["ae_fake"] = ae_loss_fake
if joint:
encoder.zero_grad()
generator.zero_grad()
(ae_loss_fake * args.lambda_rec_w).backward()
e_optim.step()
if args.g_decay is not None:
scale_grad(generator, g_scale)
# Do NOT update F (or generator.style). Grad should be zero when style
# is detached in generator, but we explicitly zero it, just in case.
if not args.no_detach_style:
generator.style.zero_grad()
g_optim.step()
else:
encoder.zero_grad()
(ae_loss_fake * args.lambda_rec_w).backward()
e_optim.step()
# Train AE on real samples (image reconstruction)
if args.lambda_pix + args.lambda_vgg + args.lambda_adv > 0:
requires_grad(encoder, True)
requires_grad(generator, joint)
requires_grad(discriminator, False)
requires_grad(discriminator2, False)
pix_loss = vgg_loss = adv_loss = torch.tensor(0., device=device)
for step_index in range(args.n_step_e):
real_img = real_imgs[step_index]
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real],
input_is_latent=input_is_latent)
if args.lambda_pix > 0:
if args.pix_loss == 'l2':
pix_loss = torch.mean((rec_img - real_img)**2)
elif args.pix_loss == 'l1':
pix_loss = F.l1_loss(rec_img, real_img)
if args.lambda_vgg > 0:
vgg_loss = torch.mean(
(vggnet(real_img) - vggnet(rec_img))**2)
if args.lambda_adv > 0:
if args.augment:
rec_img_aug, _ = augment(rec_img, ada_aug_p)
else:
rec_img_aug = rec_img
rec_pred = discriminator(encoder(rec_img_aug)[0])
adv_loss = g_nonsaturating_loss(rec_pred)
if args.use_adaptive_weight and i >= args.disc_iter_start:
nll_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg
g_loss = adv_loss * args.lambda_adv
d_weight = calculate_adaptive_weight(nll_loss,
g_loss,
last_layer=last_layer)
ae_loss_real = (pix_loss * args.lambda_pix +
vgg_loss * args.lambda_vgg +
d_weight * adv_loss * args.lambda_adv)
loss_dict["ae_real"] = ae_loss_real
loss_dict["pix"] = pix_loss
loss_dict["vgg"] = vgg_loss
loss_dict["adv"] = adv_loss
if joint:
encoder.zero_grad()
generator.zero_grad()
ae_loss_real.backward()
e_optim.step()
if args.g_decay is not None:
scale_grad(generator, g_scale)
g_optim.step()
else:
encoder.zero_grad()
ae_loss_real.backward()
e_optim.step()
if args.g_decay is not None:
g_scale *= args.g_decay
# Update EMA
ema_nimg = args.ema_kimg * 1000
if args.ema_rampup is not None:
ema_nimg = min(ema_nimg, i * args.batch * args.ema_rampup)
accum = 0.5**(args.batch / max(ema_nimg, 1e-8))
accumulate(g_ema, g_module, 0 if args.no_ema_g else accum)
accumulate(e_ema, e_module, 0 if args.no_ema_e else accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
g_loss_val = loss_reduced["g"].mean().item()
ae_real_val = loss_reduced["ae_real"].mean().item()
ae_fake_val = loss_reduced["ae_fake"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
recx_score_val = loss_reduced["recx_score"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
pix_loss_val = loss_reduced["pix"].mean().item()
vgg_loss_val = loss_reduced["vgg"].mean().item()
avg_pix_loss.update(pix_loss_val, real_img.shape[0])
avg_vgg_loss.update(vgg_loss_val, real_img.shape[0])
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; r1: {r1_val:.4f}; "
f"ae_fake: {ae_fake_val:.4f}; ae_real: {ae_real_val:.4f}; "
f"g: {g_loss_val:.4f}; path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}; "
f"d_weight: {d_weight.item():.4f}; "))
if i % args.log_every == 0:
with torch.no_grad():
g_ema.eval()
e_ema.eval()
nrow = int(args.n_sample**0.5)
nchw = list(sample_x.shape)[1:]
# Reconstruction of real images
latent_x, _ = e_ema(sample_x)
rec_real, _ = g_ema([latent_x],
input_is_latent=input_is_latent)
sample = torch.cat(
(sample_x.reshape(args.n_sample // nrow, nrow, *nchw),
rec_real.reshape(args.n_sample // nrow, nrow, *nchw)),
1)
utils.save_image(
sample.reshape(2 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-recon.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
ref_pix_loss = torch.sum(torch.abs(sample_x - rec_real))
ref_vgg_loss = torch.mean(
(vggnet(sample_x) -
vggnet(rec_real))**2) if vggnet is not None else 0
# Fixed fake samples and reconstructions
sample_gz, _ = g_ema([sample_z])
latent_gz, _ = e_ema(sample_gz)
rec_fake, _ = g_ema([latent_gz],
input_is_latent=input_is_latent)
sample = torch.cat(
(sample_gz.reshape(args.n_sample // nrow, nrow, *nchw),
rec_fake.reshape(args.n_sample // nrow, nrow, *nchw)),
1)
utils.save_image(
sample.reshape(2 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-sample.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write((
f"{i:07d}; "
f"d: {d_loss_val:.4f}; r1: {r1_val:.4f}; "
f"ae_fake: {ae_fake_val:.4f}; ae_real: {ae_real_val:.4f}; "
f"g: {g_loss_val:.4f}; path: {path_loss_val:.4f}; mean_path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}; {'; '.join([f'{k}: {r_t_dict[k]:.4f}' for k in r_t_dict])}; "
f"real_score: {real_score_val:.4f}; fake_score: {fake_score_val:.4f}; recx_score: {recx_score_val:.4f}; "
f"pix: {avg_pix_loss.avg:.4f}; vgg: {avg_vgg_loss.avg:.4f}; "
f"ref_pix: {ref_pix_loss.item():.4f}; ref_vgg: {ref_vgg_loss.item():.4f}; "
f"d_weight: {d_weight.item():.4f}; "
f"\n"))
if wandb and args.wandb:
wandb.log({
"Generator": g_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1": r1_val,
"Path Length Regularization": path_loss_val,
"Mean Path Length": mean_path_length,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
"Path Length": path_length_val,
})
if args.eval_every > 0 and i % args.eval_every == 0:
with torch.no_grad():
fid_sa = fid_re = fid_sr = 0
g_ema.eval()
e_ema.eval()
if args.truncation < 1:
mean_latent = g_ema.mean_latent(4096)
# Sample FID
if 'fid_sample' in args.which_metric:
features = extract_feature_from_samples(
g_ema, inception, args.truncation, mean_latent, 64,
args.n_sample_fid, args.device).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid_sa = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
# Sample reconstruction FID
if 'fid_sample_recon' in args.which_metric:
features = extract_feature_from_samples(
g_ema,
inception,
args.truncation,
mean_latent,
64,
args.n_sample_fid,
args.device,
mode='recon',
encoder=e_ema,
input_is_latent=input_is_latent,
).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid_sr = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
# Real reconstruction FID
if 'fid_recon' in args.which_metric:
features = extract_feature_from_reconstruction(
e_ema,
g_ema,
inception,
args.truncation,
mean_latent,
loader2,
args.device,
input_is_latent=input_is_latent,
mode='recon',
).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid_re = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
with open(os.path.join(args.log_dir, 'log_fid.txt'),
'a+') as f:
f.write(
f"{i:07d}; sample: {float(fid_sa):.4f}; rec_fake: {float(fid_sr):.4f}; rec_real: {float(fid_re):.4f};\n"
)
if i % args.save_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"d2":
d2_module.state_dict() if args.decouple_d else None,
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"d2_optim":
d2_optim.state_dict() if args.decouple_d else None,
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight',
f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"d2":
d2_module.state_dict() if args.decouple_d else None,
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"d2_optim":
d2_optim.state_dict() if args.decouple_d else None,
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, generator, encoder, discriminator, discriminator2,
vggnet, g_optim, e_optim, d_optim, d2_optim, g_ema, e_ema, device):
# kwargs_d = {'detach_aux': args.detach_d_aux_head}
if args.dataset == 'imagefolder':
loader = sample_data2(loader)
else:
loader = sample_data(loader)
if args.eval_every > 0:
inception = nn.DataParallel(load_patched_inception_v3()).to(device)
inception.eval()
with open(args.inception, "rb") as f:
embeds = pickle.load(f)
real_mean = embeds["mean"]
real_cov = embeds["cov"]
else:
inception = real_mean = real_cov = None
mean_latent = None
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
r1_loss = torch.tensor(0.0, device=device)
g_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
loss_dict = {}
avg_pix_loss = util.AverageMeter()
avg_vgg_loss = util.AverageMeter()
if args.distributed:
g_module = generator.module
e_module = encoder.module
d_module = discriminator.module
else:
g_module = generator
e_module = encoder
d_module = discriminator
d2_module = None
if discriminator2 is not None:
if args.distributed:
d2_module = discriminator2.module
else:
d2_module = discriminator2
accum = 0.5**(32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 256,
device)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
sample_x = load_real_samples(args, loader)
sample_x1 = sample_x[:, 0, ...]
sample_x2 = sample_x[:, -1, ...]
sample_idx = torch.randperm(args.n_sample)
n_step_max = max(args.n_step_d, args.n_step_e)
requires_grad(g_ema, False)
requires_grad(e_ema, False)
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
frames = [get_batch(loader, device) for _ in range(n_step_max)]
# Train Discriminator
requires_grad(generator, False)
requires_grad(encoder, False)
requires_grad(discriminator, True)
for step_index in range(args.n_step_d):
frames1, frames2 = frames[step_index]
real_img = frames1
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
if args.use_ema:
g_ema.eval()
fake_img, _ = g_ema(noise)
else:
fake_img, _ = generator(noise)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_pred = discriminator(fake_img)
real_pred = discriminator(real_img_aug)
d_loss_fake = F.softplus(fake_pred).mean()
d_loss_real = F.softplus(-real_pred).mean()
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
d_loss_rec = 0.
if args.lambda_rec_d > 0 and not args.decouple_d: # Do not train D on x_rec if decouple_d
if args.use_ema:
e_ema.eval()
g_ema.eval()
latent_real, _ = e_ema(real_img)
rec_img, _ = g_ema([latent_real], input_is_latent=True)
else:
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
rec_pred = discriminator(rec_img)
d_loss_rec = F.softplus(rec_pred).mean()
loss_dict["rec_score"] = rec_pred.mean()
d_loss_cross = 0.
if args.lambda_cross_d > 0 and not args.decouple_d:
if args.use_ema:
e_ema.eval()
w1, _ = e_ema(frames1)
w2, _ = e_ema(frames2)
else:
w1, _ = encoder(frames1)
w2, _ = encoder(frames2)
dw = w2 - w1
dw_shuffle = dw[torch.randperm(args.batch), ...]
if args.use_ema:
g_ema.eval()
cross_img, _ = g_ema([w1 + dw_shuffle],
input_is_latent=True)
else:
cross_img, _ = generator([w1 + dw_shuffle],
input_is_latent=True)
cross_pred = discriminator(cross_img)
d_loss_cross = F.softplus(cross_pred).mean()
d_loss_fake_cross = 0.
if args.lambda_fake_cross_d > 0:
if args.use_ema:
e_ema.eval()
w1, _ = e_ema(frames1)
w2, _ = e_ema(frames2)
else:
w1, _ = encoder(frames1)
w2, _ = encoder(frames2)
dw = w2 - w1
noise = mixing_noise(args.batch, args.latent, args.mixing,
device)
if args.use_ema:
g_ema.eval()
style = g_ema.get_styles(noise).view(args.batch, -1)
else:
style = generator.get_styles(noise).view(args.batch, -1)
if dw.shape[1] < style.shape[1]: # W space
dw = dw.repeat(1, args.n_latent)
if args.use_ema:
cross_img, _ = g_ema([style + dw], input_is_latent=True)
else:
cross_img, _ = generator([style + dw],
input_is_latent=True)
fake_cross_pred = discriminator(cross_img)
d_loss_fake_cross = F.softplus(fake_cross_pred).mean()
d_loss = (d_loss_real + d_loss_fake +
d_loss_fake_cross * args.lambda_fake_cross_d +
d_loss_rec * args.lambda_rec_d +
d_loss_cross * args.lambda_cross_d)
loss_dict["d"] = d_loss
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
real_pred = discriminator(real_img)
r1_loss = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every +
0 * real_pred[0]).backward()
d_optim.step()
loss_dict["r1"] = r1_loss
# Train Discriminator2
if args.decouple_d and discriminator2 is not None:
requires_grad(generator, False)
requires_grad(encoder, False)
requires_grad(discriminator2, True)
for step_index in range(
args.n_step_e): # n_step_d2 is same as n_step_e
frames1, frames2 = frames[step_index]
real_img = frames1
if args.use_ema:
e_ema.eval()
g_ema.eval()
latent_real, _ = e_ema(real_img)
rec_img, _ = g_ema([latent_real], input_is_latent=True)
else:
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
rec_pred = discriminator2(rec_img)
d2_loss_rec = F.softplus(rec_pred).mean()
real_pred1 = discriminator2(frames1)
d2_loss_real = F.softplus(-real_pred1).mean()
if args.use_frames2_d:
real_pred2 = discriminator2(frames2)
d2_loss_real += F.softplus(-real_pred2).mean()
if args.use_ema:
e_ema.eval()
w1, _ = e_ema(frames1)
w2, _ = e_ema(frames2)
else:
w1, _ = encoder(frames1)
w2, _ = encoder(frames2)
dw = w2 - w1
dw_shuffle = dw[torch.randperm(args.batch), ...]
cross_img, _ = generator([w1 + dw_shuffle],
input_is_latent=True)
cross_pred = discriminator2(cross_img)
d2_loss_cross = F.softplus(cross_pred).mean()
d2_loss = d2_loss_real + d2_loss_rec + d2_loss_cross
loss_dict["d2"] = d2_loss
loss_dict["rec_score"] = rec_pred.mean()
loss_dict["cross_score"] = cross_pred.mean()
discriminator2.zero_grad()
d2_loss.backward()
d2_optim.step()
d_regularize = i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
real_pred = discriminator2(real_img)
r1_loss = d_r1_loss(real_pred, real_img)
discriminator2.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every +
0 * real_pred[0]).backward()
d2_optim.step()
# Train Encoder
requires_grad(encoder, True)
requires_grad(generator, args.train_ge)
requires_grad(discriminator, False)
if discriminator2 is not None:
requires_grad(discriminator2, False)
pix_loss = vgg_loss = adv_loss = torch.tensor(0., device=device)
for step_index in range(args.n_step_e):
frames1, frames2 = frames[step_index]
real_img = frames1
latent_real, _ = encoder(real_img)
if args.use_ema:
g_ema.eval()
rec_img, _ = g_ema([latent_real], input_is_latent=True)
else:
rec_img, _ = generator([latent_real], input_is_latent=True)
if args.lambda_adv > 0:
if not args.decouple_d:
rec_pred = discriminator(rec_img)
else:
rec_pred = discriminator2(rec_img)
adv_loss = g_nonsaturating_loss(rec_pred)
if args.lambda_pix > 0:
if args.pix_loss == 'l2':
pix_loss = torch.mean((rec_img - real_img)**2)
else:
pix_loss = F.l1_loss(rec_img, real_img)
if args.lambda_vgg > 0:
vgg_loss = torch.mean((vggnet(real_img) - vggnet(rec_img))**2)
e_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
loss_dict["e"] = e_loss
loss_dict["pix"] = pix_loss
loss_dict["vgg"] = vgg_loss
loss_dict["adv"] = adv_loss
if args.train_ge:
encoder.zero_grad()
generator.zero_grad()
e_loss.backward()
e_optim.step()
g_optim.step()
else:
encoder.zero_grad()
e_loss.backward()
e_optim.step()
# Train Generator
requires_grad(generator, True)
requires_grad(discriminator, False)
if discriminator2 is not None:
requires_grad(discriminator2, False)
frames1, frames2 = frames[0]
real_img = frames1
g_loss_fake = 0.
if args.lambda_fake_g > 0:
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = discriminator(fake_img)
g_loss_fake = g_nonsaturating_loss(fake_pred)
g_loss_rec = 0.
if args.lambda_rec_g > 0:
if args.use_ema:
e_ema.eval()
latent_real, _ = e_ema(real_img)
else:
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
if not args.decouple_d:
rec_pred = discriminator(rec_img)
else:
rec_pred = discriminator2(rec_img)
g_loss_rec = g_nonsaturating_loss(rec_pred)
g_loss_cross = 0.
if args.lambda_cross_g > 0:
if args.use_ema:
e_ema.eval()
w1, _ = e_ema(frames1)
w2, _ = e_ema(frames2)
else:
w1, _ = encoder(frames1)
w2, _ = encoder(frames2)
dw = w2 - w1
dw_shuffle = dw[torch.randperm(args.batch), ...]
cross_img, _ = generator([w1 + dw_shuffle], input_is_latent=True)
if not args.decouple_d:
cross_pred = discriminator(cross_img)
else:
cross_pred = discriminator2(cross_img)
g_loss_cross = g_nonsaturating_loss(cross_pred)
g_loss_fake_cross = 0.
if args.lambda_fake_cross_g > 0:
if args.use_ema:
e_ema.eval()
w1, _ = e_ema(frames1)
w2, _ = e_ema(frames2)
else:
w1, _ = encoder(frames1)
w2, _ = encoder(frames2)
dw = w2 - w1
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
style = generator.get_styles(noise).view(args.batch, -1)
if dw.shape[1] < style.shape[1]: # W space
dw = dw.repeat(1, args.n_latent)
cross_img, _ = generator([style + dw], input_is_latent=True)
fake_cross_pred = discriminator(cross_img)
g_loss_fake_cross = g_nonsaturating_loss(fake_cross_pred)
g_loss = (g_loss_fake * args.lambda_fake_g +
g_loss_rec * args.lambda_rec_g +
g_loss_cross * args.lambda_cross_g +
g_loss_fake_cross * args.lambda_fake_cross_g)
loss_dict["g"] = g_loss
generator.zero_grad()
g_loss.backward()
g_optim.step()
g_regularize = args.g_reg_every > 0 and i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing,
device)
fake_img, latents = generator(noise, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
mean_path_length_avg = (reduce_sum(mean_path_length).item() /
get_world_size())
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
accumulate(e_ema, e_module, accum)
accumulate(g_ema, g_module, accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
g_loss_val = loss_reduced["g"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
pix_loss_val = loss_reduced["pix"].mean().item()
vgg_loss_val = loss_reduced["vgg"].mean().item()
adv_loss_val = loss_reduced["adv"].mean().item()
avg_pix_loss.update(pix_loss_val, real_img.shape[0])
avg_vgg_loss.update(vgg_loss_val, real_img.shape[0])
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}; "
f"pix: {pix_loss_val:.4f}; vgg: {vgg_loss_val:.4f}; adv: {adv_loss_val:.4f}"
))
if i % args.log_every == 0:
with torch.no_grad():
latent_x, _ = e_ema(sample_x)
fake_x, _ = generator([latent_x],
input_is_latent=True,
return_latents=False)
sample_pix_loss = torch.sum((sample_x - fake_x)**2)
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(
f"{i:07d}; pix: {avg_pix_loss.avg}; vgg: {avg_vgg_loss.avg}; "
f"ref: {sample_pix_loss.item()};\n")
if args.eval_every > 0 and i % args.eval_every == 0:
with torch.no_grad():
g_ema.eval()
if args.truncation < 1:
mean_latent = g_ema.mean_latent(4096)
features = extract_feature_from_samples(
g_ema, inception, args.truncation, mean_latent, 64,
args.n_sample_fid, args.device).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
print("fid:", fid)
with open(os.path.join(args.log_dir, 'log_fid.txt'),
'a+') as f:
f.write(f"{i:07d}; fid: {float(fid):.4f};\n")
if wandb and args.wandb:
wandb.log({
"Generator": g_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1": r1_val,
"Path Length Regularization": path_loss_val,
"Mean Path Length": mean_path_length,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
"Path Length": path_length_val,
})
if i % args.log_every == 0:
with torch.no_grad():
# Fixed fake samples
g_ema.eval()
sample, _ = g_ema([sample_z])
utils.save_image(
sample,
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-sample.png"),
nrow=int(args.n_sample**0.5),
normalize=True,
value_range=(-1, 1),
)
# Reconstruction samples
e_ema.eval()
nrow = int(args.n_sample**0.5)
nchw = list(sample_x.shape)[1:]
latent_real, _ = e_ema(sample_x)
fake_img, _ = g_ema([latent_real],
input_is_latent=True,
return_latents=False)
sample = torch.cat(
(sample_x.reshape(args.n_sample // nrow, nrow, *nchw),
fake_img.reshape(args.n_sample // nrow, nrow, *nchw)),
1)
utils.save_image(
sample.reshape(2 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-recon.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
if i % args.save_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"d2":
d2_module.state_dict() if args.decouple_d else None,
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"d2_optim":
d2_optim.state_dict() if args.decouple_d else None,
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight',
f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"d2":
d2_module.state_dict() if args.decouple_d else None,
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"d2_optim":
d2_optim.state_dict() if args.decouple_d else None,
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, generator, discriminator, g_optim, d_optim, g_ema,
device):
loader = sample_data(loader)
print("Save checkpoint very %d iter." % args.save_iter)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
mean_path_length = 0
with torch.cuda.amp.autocast():
r1_loss = torch.tensor(0.0, device=device)
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
loss_dict = {}
if args.distributed:
g_module = generator.module
d_module = discriminator.module
else:
g_module = generator
d_module = discriminator
accum = 0.5**(32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 256,
device)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
for idx in pbar:
i = idx + args.start_iter
print("iter: %d" % i)
if i > args.iter:
print("Done!")
break
##############################################################################
# Casts operations to mixed precision
# https://pytorch.org/docs/stable/notes/amp_examples.html
with torch.cuda.amp.autocast():
real_img = next(loader)
real_img = real_img.to(device)
requires_grad(generator, False)
requires_grad(discriminator, True)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_pred = discriminator(fake_img)
real_pred = discriminator(real_img_aug)
d_loss = d_logistic_loss(real_pred, fake_pred)
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
discriminator.zero_grad()
# Scales loss. Calls backward() on scaled loss to create scaled gradients.
# Backward passes under autocast are not recommended.
# Backward ops run in the same dtype autocast chose for corresponding forward ops.
scaler.scale(d_loss).backward()
# scaler.step() first unscales the gradients of the optimizer's assigned params.
# If these gradients do not contain infs or NaNs, optimizer.step() is then called,
# otherwise, optimizer.step() is skipped.
scaler.step(d_optim)
scaler.update()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = i % args.d_reg_every == 0
# Runs the forward pass with autocasting.
if d_regularize:
with torch.cuda.amp.autocast():
real_img.requires_grad = True
real_pred = discriminator(real_img)
r1_loss = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
scaler.scale((args.r1 / 2 * r1_loss * args.d_reg_every +
0 * real_pred[0])).backward()
scaler.step(d_optim)
scaler.update()
loss_dict["r1"] = r1_loss
# Runs the forward pass with autocasting.
with torch.cuda.amp.autocast():
requires_grad(generator, True)
requires_grad(discriminator, False)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = discriminator(fake_img)
g_loss = g_nonsaturating_loss(fake_pred)
loss_dict["g"] = g_loss
generator.zero_grad()
scaler.scale(g_loss).backward()
#g_loss.backward()
scaler.step(g_optim)
scaler.update()
#g_optim.step()
g_regularize = i % args.g_reg_every == 0
if g_regularize:
with torch.cuda.amp.autocast():
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing,
device)
fake_img, latents = generator(noise, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
scaler.scale(weighted_path_loss).backward()
#weighted_path_loss.backward()
scaler.step(g_optim)
#g_optim.step()
# Updates the scale for next iteration
scaler.update()
mean_path_length_avg = (reduce_sum(mean_path_length).item() /
get_world_size())
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
accumulate(g_ema, g_module, accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
g_loss_val = loss_reduced["g"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}"))
if wandb and args.wandb:
wandb.log({
"Generator": g_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1": r1_val,
"Path Length Regularization": path_loss_val,
"Mean Path Length": mean_path_length,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
"Path Length": path_length_val,
})
if i % 100 == 0:
with torch.no_grad():
g_ema.eval()
sample, _ = g_ema([sample_z])
utils.save_image(
sample,
f"sample/{str(i).zfill(6)}.png",
nrow=int(args.n_sample**0.5),
normalize=True,
range=(-1, 1),
)
if i % args.save_iter == 0:
torch.save(
{
"g": g_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
},
f"checkpoint/{str(i).zfill(6)}.pt",
)
def train(args, loader, encoder, generator, discriminator, discriminator_w,
vggnet, pwcnet, e_optim, g_optim, g1_optim, d_optim, dw_optim,
e_ema, g_ema, device):
loader = sample_data(loader)
args.toggle_grads = True
args.augment = False
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar, initial=args.start_iter, dynamic_ncols=True, smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
e_loss_val = 0
rec_loss_val = 0
vgg_loss_val = 0
adv_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
loss_dict = {"d": torch.tensor(0., device=device),
"real_score": torch.tensor(0., device=device),
"fake_score": torch.tensor(0., device=device),
"hybrid_score": torch.tensor(0., device=device),
"r1_d": torch.tensor(0., device=device),
"rec": torch.tensor(0., device=device),}
avg_pix_loss = util.AverageMeter()
avg_vgg_loss = util.AverageMeter()
if args.distributed:
e_module = encoder.module
d_module = discriminator.module
g_module = generator.module
else:
e_module = encoder
d_module = discriminator
g_module = generator
accum = 0.5 ** (32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 256, device)
# sample_x = accumulate_batches(loader, args.n_sample).to(device)
sample_x = load_real_samples(args, loader)
sample_x1 = sample_x[:,0,...]
sample_x2 = sample_x[:,-1,...]
sample_idx = torch.randperm(args.n_sample)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
for idx in pbar:
i = idx + args.start_iter
if get_rank() == 0:
if i % args.log_every == 0:
with torch.no_grad():
e_eval = e_ema
e_eval.eval()
g_ema.eval()
nrow = int(args.n_sample ** 0.5)
nchw = list(sample_x1.shape)[1:]
# Recon
latent_real, _ = e_eval(sample_x1)
fake_img, _ = g_ema([latent_real], input_is_latent=True, return_latents=False)
sample = torch.cat((sample_x1.reshape(args.n_sample//nrow, nrow, *nchw),
fake_img.reshape(args.n_sample//nrow, nrow, *nchw)), 1)
utils.save_image(
sample.reshape(2*args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample', f"{str(i).zfill(6)}-recon.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
# Cross
w1, _ = e_eval(sample_x1)
w2, _ = e_eval(sample_x2)
delta_w = w2 - w1
delta_w = delta_w[sample_idx,...]
fake_img, _ = g_ema([w1 + delta_w], input_is_latent=True, return_latents=False)
sample = torch.cat((sample_x2.reshape(args.n_sample//nrow, nrow, *nchw),
fake_img.reshape(args.n_sample//nrow, nrow, *nchw)), 1)
utils.save_image(
sample.reshape(2*args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample', f"{str(i).zfill(6)}-cross.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
# Sample
sample, _ = g_ema([sample_z])
utils.save_image(
sample,
os.path.join(args.log_dir, 'sample', f"{str(i).zfill(6)}-sample.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
e_eval.train()
if i > args.iter:
print("Done!")
break
frames = next(loader) # [N, T, C, H, W]
batch = frames.shape[0]
frames1 = frames[:,0,...]
selected_indices = torch.sort(torch.multinomial(torch.ones(batch, args.nframe_num-1), 2)+1, 1)[0]
frames2 = frames[range(batch),selected_indices[:,0],...]
frames3 = frames[range(batch),selected_indices[:,1],...]
frames1 = frames1.to(device)
frames2 = frames2.to(device)
frames3 = frames3.to(device)
# Train Discriminator
if args.toggle_grads:
requires_grad(encoder, False)
requires_grad(generator, False)
requires_grad(discriminator, True)
requires_grad(discriminator_w, True)
real_img, fake_img, _, _, _ = cross_reconstruction(encoder, generator, frames1, frames2, frames3, args.cond_disc)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_img_aug = fake_img
fake_pred = discriminator(fake_img_aug)
real_pred = discriminator(real_img_aug)
d_loss = d_logistic_loss(real_pred, fake_pred)
if args.lambda_gan > 0:
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = discriminator(fake_img)
fake_loss = F.softplus(fake_pred)
d_loss += fake_loss.mean() * args.lambda_gan
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
# loss_dict["fake_score"] = fake_pred.mean()
fake_pred1, fake_pred2 = fake_pred.chunk(2, dim=0)
loss_dict["fake_score"] = fake_pred1.mean()
loss_dict["hybrid_score"] = fake_pred2.mean()
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = args.d_reg_every > 0 and i % args.d_reg_every == 0
if d_regularize:
# why not regularize on augmented real?
real_img.requires_grad = True
real_pred = discriminator(real_img)
r1_loss_d = d_r1_loss(real_pred, real_img)
d_optim.zero_grad()
(args.r1 / 2 * r1_loss_d * args.d_reg_every + 0 * real_pred.view(-1)[0]).backward()
# Why 0* ? Answer is here https://github.com/rosinality/stylegan2-pytorch/issues/76
d_optim.step()
loss_dict["r1_d"] = r1_loss_d
# Train Discriminator_W
if args.learned_prior and args.lambda_gan_w > 0:
noise = mixing_noise(args.batch, args.latent, 0, device)
fake_w = generator.get_latent(noise[0])
real_w, _ = encoder(frames1)
fake_pred = discriminator_w(fake_w)
real_pred = discriminator_w(real_w)
d_loss_w = d_logistic_loss(real_pred, fake_pred)
dw_optim.zero_grad()
d_loss_w.backward()
dw_optim.step()
# Train Encoder and Generator
if args.toggle_grads:
requires_grad(encoder, True)
requires_grad(generator, True)
requires_grad(discriminator, False)
requires_grad(discriminator_w, False)
pix_loss = vgg_loss = adv_loss = rec_loss = torch.tensor(0., device=device)
_, fake_img, x_real, x_recon, x_cross = cross_reconstruction(encoder, generator, frames1, frames2, frames3, args.cond_disc)
if args.lambda_adv > 0:
if args.augment:
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
fake_img_aug = fake_img
fake_pred = discriminator(fake_img_aug)
adv_loss = g_nonsaturating_loss(fake_pred)
if args.lambda_pix > 0:
if args.pix_loss == 'l2':
pix_loss = torch.mean((x_recon - x_real) ** 2)
else:
pix_loss = F.l1_loss(x_recon, x_real)
if args.lambda_vgg > 0:
real_feat = vggnet(x_real)
fake_feat = vggnet(x_recon) if not args.vgg_on_cross else vggnet(x_cross)
vgg_loss = torch.mean((fake_feat - real_feat) ** 2)
e_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
if args.lambda_gan > 0 and not args.no_sim_opt:
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = discriminator(fake_img)
g_loss = g_nonsaturating_loss(fake_pred)
e_loss += g_loss * args.lambda_gan
loss_dict["e"] = e_loss
loss_dict["pix"] = pix_loss
loss_dict["vgg"] = vgg_loss
loss_dict["adv"] = adv_loss
e_optim.zero_grad()
g_optim.zero_grad()
e_loss.backward()
e_optim.step()
g_optim.step()
if args.learned_prior:
g_loss_w = 0.
if args.lambda_gan_w > 0:
noise = mixing_noise(args.batch, args.latent, 0, device)
fake_w = generator.get_latent(noise[0])
fake_pred = discriminator_w(fake_w)
g_loss_w += g_nonsaturating_loss(fake_pred) * args.lambda_gan_w
if args.lambda_adv_w > 0:
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
fake_pred = discriminator(fake_img)
g_loss_w += g_nonsaturating_loss(fake_pred) * args.lambda_adv_w
g1_optim.zero_grad()
g_loss_w.backward()
g1_optim.step()
if args.lambda_gan > 0 and args.no_sim_opt:
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = discriminator(fake_img)
g_loss = g_nonsaturating_loss(fake_pred) * args.lambda_gan
generator.zero_grad()
g_loss.backward()
g_optim.step()
g_regularize = args.lambda_gan > 0 and args.g_reg_every > 0 and i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing, device)
fake_img, latents = generator(noise, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length
)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
# mean_path_length_avg = (
# reduce_sum(mean_path_length).item() / get_world_size()
# )
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
accumulate(e_ema, e_module, accum)
accumulate(g_ema, g_module, accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
e_loss_val = loss_reduced["e"].mean().item()
r1_d_val = loss_reduced["r1_d"].mean().item()
pix_loss_val = loss_reduced["pix"].mean().item()
vgg_loss_val = loss_reduced["vgg"].mean().item()
adv_loss_val = loss_reduced["adv"].mean().item()
rec_loss_val = loss_reduced["rec"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
hybrid_score_val = loss_reduced["hybrid_score"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
# path_length_val = loss_reduced["path_length"].mean().item()
avg_pix_loss.update(pix_loss_val, real_img.shape[0])
avg_vgg_loss.update(vgg_loss_val, real_img.shape[0])
if get_rank() == 0:
pbar.set_description(
(
f"d: {d_loss_val:.4f}; e: {e_loss_val:.4f}; r1_d: {r1_d_val:.4f}; "
f"pix: {pix_loss_val:.4f}; vgg: {vgg_loss_val:.4f}; adv: {adv_loss_val:.4f}; "
f"path: {path_loss_val:.4f}; augment: {ada_aug_p:.4f}"
)
)
if i % args.log_every == 0:
with torch.no_grad():
latent_x, _ = e_ema(sample_x1)
fake_x, _ = generator([latent_x], input_is_latent=True, return_latents=False)
sample_pix_loss = torch.sum((sample_x1 - fake_x) ** 2)
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(f"{i:07d}; pix: {avg_pix_loss.avg}; vgg: {avg_vgg_loss.avg}; "
f"ref: {sample_pix_loss.item()};\n")
if wandb and args.wandb:
wandb.log(
{
"Encoder": e_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1 D": r1_d_val,
"Pix Loss": pix_loss_val,
"VGG Loss": vgg_loss_val,
"Adv Loss": adv_loss_val,
"Rec Loss": rec_loss_val,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
"Hybrid Score": hybrid_score_val,
}
)
if i % args.save_every == 0:
e_eval = e_ema
torch.save(
{
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g": g_module.state_dict(),
"g_ema": g_module.state_dict(),
"e_ema": e_eval.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"g_optim": g_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g": g_module.state_dict(),
"g_ema": g_module.state_dict(),
"e_ema": e_eval.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"g_optim": g_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, loader2, encoder, generator, discriminator, vggnet,
pwcnet, e_optim, d_optim, e_ema, pca_state, device):
inception = real_mean = real_cov = mean_latent = None
if args.eval_every > 0:
inception = nn.DataParallel(load_patched_inception_v3()).to(device)
inception.eval()
with open(args.inception, "rb") as f:
embeds = pickle.load(f)
real_mean = embeds["mean"]
real_cov = embeds["cov"]
if get_rank() == 0:
if args.eval_every > 0:
with open(os.path.join(args.log_dir, 'log_fid.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
if args.log_every > 0:
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
d_loss_val = 0
e_loss_val = 0
rec_loss_val = 0
vgg_loss_val = 0
adv_loss_val = 0
loss_dict = {
"d": torch.tensor(0., device=device),
"real_score": torch.tensor(0., device=device),
"fake_score": torch.tensor(0., device=device),
"r1_d": torch.tensor(0., device=device),
"r1_e": torch.tensor(0., device=device),
"rec": torch.tensor(0., device=device),
}
avg_pix_loss = util.AverageMeter()
avg_vgg_loss = util.AverageMeter()
if args.distributed:
e_module = encoder.module
d_module = discriminator.module
g_module = generator.module
else:
e_module = encoder
d_module = discriminator
g_module = generator
# accum = 0.5 ** (32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length,
args.ada_every, device)
# sample_x = accumulate_batches(loader, args.n_sample).to(device)
sample_x = load_real_samples(args, loader)
if sample_x.ndim > 4:
sample_x = sample_x[:, 0, ...]
input_is_latent = args.latent_space != 'z' # Encode in z space?
requires_grad(generator, False) # always False
generator.eval() # Generator should be ema and in eval mode
g_ema = generator
# if args.no_ema or e_ema is None:
# e_ema = encoder
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
real_img = next(loader)
real_img = real_img.to(device)
# Train Encoder
if args.toggle_grads:
requires_grad(encoder, True)
requires_grad(discriminator, False)
pix_loss = vgg_loss = adv_loss = rec_loss = torch.tensor(0.,
device=device)
latent_real, _ = encoder(real_img)
fake_img, _ = generator([latent_real], input_is_latent=input_is_latent)
if args.lambda_adv > 0:
if args.augment:
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
fake_img_aug = fake_img
fake_pred = discriminator(fake_img_aug)
adv_loss = g_nonsaturating_loss(fake_pred)
if args.lambda_pix > 0:
if args.pix_loss == 'l2':
pix_loss = torch.mean((fake_img - real_img)**2)
else:
pix_loss = F.l1_loss(fake_img, real_img)
if args.lambda_vgg > 0:
real_feat = vggnet(real_img)
fake_feat = vggnet(fake_img)
vgg_loss = torch.mean((real_feat - fake_feat)**2)
e_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
loss_dict["e"] = e_loss
loss_dict["pix"] = pix_loss
loss_dict["vgg"] = vgg_loss
loss_dict["adv"] = adv_loss
encoder.zero_grad()
e_loss.backward()
e_optim.step()
if args.train_on_fake:
e_regularize = args.e_rec_every > 0 and i % args.e_rec_every == 0
if e_regularize and args.lambda_rec > 0:
noise = mixing_noise(args.batch, args.latent, args.mixing,
device)
fake_img, latent_fake = generator(
noise,
input_is_latent=input_is_latent,
return_latents=True)
latent_pred, _ = encoder(fake_img)
if latent_pred.ndim < 3:
latent_pred = latent_pred.unsqueeze(1).repeat(
1, latent_fake.size(1), 1)
rec_loss = torch.mean((latent_fake - latent_pred)**2)
encoder.zero_grad()
(rec_loss * args.lambda_rec).backward()
e_optim.step()
loss_dict["rec"] = rec_loss
# e_regularize = args.e_reg_every > 0 and i % args.e_reg_every == 0
# if e_regularize:
# # why not regularize on augmented real?
# real_img.requires_grad = True
# real_pred, _ = encoder(real_img)
# r1_loss_e = d_r1_loss(real_pred, real_img)
# encoder.zero_grad()
# (args.r1 / 2 * r1_loss_e * args.e_reg_every + 0 * real_pred.view(-1)[0]).backward()
# e_optim.step()
# loss_dict["r1_e"] = r1_loss_e
if not args.no_ema and e_ema is not None:
ema_nimg = args.ema_kimg * 1000
if args.ema_rampup is not None:
ema_nimg = min(ema_nimg, i * args.batch * args.ema_rampup)
accum = 0.5**(args.batch / max(ema_nimg, 1e-8))
accumulate(e_ema, e_module, accum)
# Train Discriminator
if args.toggle_grads:
requires_grad(encoder, False)
requires_grad(discriminator, True)
if not args.no_update_discriminator and args.lambda_adv > 0:
latent_real, _ = encoder(real_img)
fake_img, _ = generator([latent_real],
input_is_latent=input_is_latent)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img_aug, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_img_aug = fake_img
fake_pred = discriminator(fake_img_aug)
real_pred = discriminator(real_img_aug)
d_loss = d_logistic_loss(real_pred, fake_pred)
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = args.d_reg_every > 0 and i % args.d_reg_every == 0
if d_regularize:
# why not regularize on augmented real?
real_img.requires_grad = True
real_pred = discriminator(real_img)
r1_loss_d = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss_d * args.d_reg_every +
0 * real_pred.view(-1)[0]).backward()
# Why 0* ? Answer is here https://github.com/rosinality/stylegan2-pytorch/issues/76
d_optim.step()
loss_dict["r1_d"] = r1_loss_d
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
e_loss_val = loss_reduced["e"].mean().item()
r1_d_val = loss_reduced["r1_d"].mean().item()
r1_e_val = loss_reduced["r1_e"].mean().item()
pix_loss_val = loss_reduced["pix"].mean().item()
vgg_loss_val = loss_reduced["vgg"].mean().item()
adv_loss_val = loss_reduced["adv"].mean().item()
rec_loss_val = loss_reduced["rec"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
avg_pix_loss.update(pix_loss_val, real_img.shape[0])
avg_vgg_loss.update(vgg_loss_val, real_img.shape[0])
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; e: {e_loss_val:.4f}; r1_d: {r1_d_val:.4f}; r1_e: {r1_e_val:.4f}; "
f"pix: {pix_loss_val:.4f}; vgg: {vgg_loss_val:.4f}; adv: {adv_loss_val:.4f}; "
f"rec: {rec_loss_val:.4f}; augment: {ada_aug_p:.4f}"))
if i % args.log_every == 0:
with torch.no_grad():
latent_x, _ = e_ema(sample_x)
fake_x, _ = g_ema([latent_x],
input_is_latent=input_is_latent)
sample_pix_loss = torch.sum((sample_x - fake_x)**2)
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(
f"{i:07d}; pix: {avg_pix_loss.avg}; vgg: {avg_vgg_loss.avg}; "
f"ref: {sample_pix_loss.item()};\n")
if args.eval_every > 0 and i % args.eval_every == 0:
with torch.no_grad():
g_ema.eval()
e_ema.eval()
# Recon
features = extract_feature_from_reconstruction(
e_ema,
g_ema,
inception,
args.truncation,
mean_latent,
loader2,
args.device,
input_is_latent=input_is_latent,
mode='recon',
).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid_re = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
# print("Recon FID:", fid_re)
with open(os.path.join(args.log_dir, 'log_fid.txt'),
'a+') as f:
f.write(f"{i:07d}; recon fid: {float(fid_re):.4f};\n")
if wandb and args.wandb:
wandb.log({
"Encoder": e_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1 D": r1_d_val,
"R1 E": r1_e_val,
"Pix Loss": pix_loss_val,
"VGG Loss": vgg_loss_val,
"Adv Loss": adv_loss_val,
"Rec Loss": rec_loss_val,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
})
if i % args.log_every == 0:
with torch.no_grad():
e_eval = encoder if args.no_ema else e_ema
e_eval.eval()
nrow = int(args.n_sample**0.5)
nchw = list(sample_x.shape)[1:]
latent_real, _ = e_eval(sample_x)
fake_img, _ = generator([latent_real],
input_is_latent=input_is_latent)
sample = torch.cat(
(sample_x.reshape(args.n_sample // nrow, nrow, *nchw),
fake_img.reshape(args.n_sample // nrow, nrow, *nchw)),
1)
utils.save_image(
sample.reshape(2 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
e_eval.train()
if i % args.save_every == 0:
e_eval = encoder if args.no_ema else e_ema
torch.save(
{
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_module.state_dict(),
"e_ema": e_eval.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight',
f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"g_ema": g_module.state_dict(),
"e_ema": e_eval.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, loader2, generator, encoder, discriminator,
discriminator2, vggnet, g_optim, e_optim, d_optim, d2_optim, g_ema,
e_ema, device):
inception = real_mean = real_cov = mean_latent = None
if args.eval_every > 0:
inception = nn.DataParallel(load_patched_inception_v3()).to(device)
inception.eval()
with open(args.inception, "rb") as f:
embeds = pickle.load(f)
real_mean = embeds["mean"]
real_cov = embeds["cov"]
if get_rank() == 0:
if args.eval_every > 0:
with open(os.path.join(args.log_dir, 'log_fid.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
if args.log_every > 0:
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write(f"Name: {getattr(args, 'name', 'NA')}\n{'-'*50}\n")
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
r1_loss = torch.tensor(0.0, device=device)
g_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
loss_dict = {}
avg_pix_loss = util.AverageMeter()
avg_vgg_loss = util.AverageMeter()
if args.distributed:
g_module = generator.module
e_module = encoder.module
d_module = discriminator.module
else:
g_module = generator
e_module = encoder
d_module = discriminator
d2_module = None
if discriminator2 is not None:
if args.distributed:
d2_module = discriminator2.module
else:
d2_module = discriminator2
# accum = 0.5 ** (32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
r_t_dict = {'real': 0, 'fake': 0, 'recx': 0} # r_t stat
real_diff = fake_diff = count = 0
g_scale = 1
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length,
args.ada_every, device)
if args.decouple_d and args.augment:
ada_aug_p2 = args.augment_p if args.augment_p > 0 else 0.0
# r_t_stat2 = 0
if args.augment_p == 0:
ada_augment2 = AdaptiveAugment(args.ada_target, args.ada_length,
args.ada_every, device)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
sample_x = load_real_samples(args, loader)
sample_x1 = sample_x2 = sample_idx = fid_batch_idx = None
if sample_x.ndim > 4:
sample_x1 = sample_x[:, 0, ...]
sample_x2 = sample_x[:, -1, ...]
sample_x = sample_x[:, 0, ...]
n_step_max = max(args.n_step_d, args.n_step_e)
requires_grad(g_ema, False)
requires_grad(e_ema, False)
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
real_imgs = [next(loader).to(device) for _ in range(n_step_max)]
# Train Discriminator
requires_grad(generator, False)
requires_grad(encoder, False)
requires_grad(discriminator, True)
for step_index in range(args.n_step_d):
real_img = real_imgs[step_index]
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
if args.use_ema:
g_ema.eval()
fake_img, _ = g_ema(noise)
else:
fake_img, _ = generator(noise)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_pred = discriminator(fake_img)
real_pred = discriminator(real_img_aug)
d_loss_fake = F.softplus(fake_pred).mean()
d_loss_real = F.softplus(-real_pred).mean()
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
d_loss_rec = 0.
if args.lambda_rec_d > 0 and not args.decouple_d: # Do not train D on x_rec if decouple_d
if args.use_ema:
e_ema.eval()
g_ema.eval()
latent_real, _ = e_ema(real_img)
rec_img, _ = g_ema([latent_real], input_is_latent=True)
else:
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
if args.augment:
rec_img, _ = augment(rec_img, ada_aug_p)
rec_pred = discriminator(rec_img)
d_loss_rec = F.softplus(rec_pred).mean()
loss_dict["recx_score"] = rec_pred.mean()
d_loss = d_loss_real + d_loss_fake * args.lambda_fake_d + d_loss_rec * args.lambda_rec_d
loss_dict["d"] = d_loss
discriminator.zero_grad()
d_loss.backward()
d_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
# Compute batchwise r_t
r_t_dict['real'] = torch.sign(real_pred).sum().item() / args.batch
r_t_dict['fake'] = torch.sign(fake_pred).sum().item() / args.batch
with torch.no_grad():
real_diff += torch.mean(real_pred - rec_pred).item()
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
x_fake, _ = generator(noise)
x_recf, _ = generator([encoder(x_fake)[0]], input_is_latent=True)
recf_pred = discriminator(x_recf)
fake_pred = discriminator(x_fake)
fake_diff += torch.mean(fake_pred - recf_pred).item()
count += 1
d_regularize = i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
else:
real_img_aug = real_img
real_pred = discriminator(real_img_aug)
r1_loss = d_r1_loss(real_pred, real_img)
discriminator.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every +
0 * real_pred[0]).backward()
d_optim.step()
loss_dict["r1"] = r1_loss
# Train Discriminator2
if args.decouple_d and discriminator2 is not None:
requires_grad(generator, False)
requires_grad(encoder, False)
requires_grad(discriminator2, True)
for step_index in range(
args.n_step_e): # n_step_d2 is same as n_step_e
real_img = real_imgs[step_index]
if args.use_ema:
e_ema.eval()
g_ema.eval()
latent_real, _ = e_ema(real_img)
rec_img, _ = g_ema([latent_real], input_is_latent=True)
else:
latent_real, _ = encoder(real_img)
rec_img, _ = generator([latent_real], input_is_latent=True)
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p2)
rec_img, _ = augment(rec_img, ada_aug_p2)
else:
real_img_aug = real_img
rec_pred = discriminator2(rec_img)
real_pred = discriminator2(real_img_aug)
d2_loss_rec = F.softplus(rec_pred).mean()
d2_loss_real = F.softplus(-real_pred).mean()
d2_loss = d2_loss_real + d2_loss_rec
loss_dict["d2"] = d2_loss
loss_dict["recx_score"] = rec_pred.mean()
discriminator2.zero_grad()
d2_loss.backward()
d2_optim.step()
real_diff += torch.mean(real_pred - rec_pred).item()
d_regularize = args.d_reg_every > 0 and i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
real_pred = discriminator2(real_img)
r1_loss = d_r1_loss(real_pred, real_img)
discriminator2.zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every +
0 * real_pred[0]).backward()
d2_optim.step()
if args.augment and args.augment_p == 0:
ada_aug_p2 = ada_augment2.tune(rec_pred)
# r_t_stat2 = ada_augment2.r_t_stat
r_t_dict['recx'] = torch.sign(rec_pred).sum().item() / args.batch
# Train Encoder
requires_grad(encoder, True)
requires_grad(generator, args.train_ge)
requires_grad(discriminator, False)
requires_grad(discriminator2, False)
pix_loss = vgg_loss = adv_loss = torch.tensor(0., device=device)
for step_index in range(args.n_step_e):
real_img = real_imgs[step_index]
latent_real, _ = encoder(real_img)
if args.use_ema:
g_ema.eval()
rec_img, _ = g_ema([latent_real], input_is_latent=True)
else:
rec_img, _ = generator([latent_real], input_is_latent=True)
if args.lambda_pix > 0:
if args.pix_loss == 'l2':
pix_loss = torch.mean((rec_img - real_img)**2)
elif args.pix_loss == 'l1':
pix_loss = F.l1_loss(rec_img, real_img)
else:
raise NotImplementedError
if args.lambda_vgg > 0:
vgg_loss = torch.mean((vggnet(real_img) - vggnet(rec_img))**2)
if args.lambda_adv > 0:
if not args.decouple_d:
if args.augment:
rec_img_aug, _ = augment(rec_img, ada_aug_p)
else:
rec_img_aug = rec_img
rec_pred = discriminator(rec_img_aug)
else:
if args.augment:
rec_img_aug, _ = augment(rec_img, ada_aug_p2)
else:
rec_img_aug = rec_img
rec_pred = discriminator2(rec_img_aug)
adv_loss = g_nonsaturating_loss(rec_pred)
e_loss = pix_loss * args.lambda_pix + vgg_loss * args.lambda_vgg + adv_loss * args.lambda_adv
loss_dict["e"] = e_loss
loss_dict["pix"] = pix_loss
loss_dict["vgg"] = vgg_loss
loss_dict["adv"] = adv_loss
if args.train_ge:
encoder.zero_grad()
generator.zero_grad()
e_loss.backward()
e_optim.step()
if args.g_decay < 1:
manually_scale_grad(generator, g_scale)
g_scale *= args.g_decay
g_optim.step()
else:
encoder.zero_grad()
e_loss.backward()
e_optim.step()
# Train Generator
requires_grad(generator, True)
requires_grad(encoder, False)
requires_grad(discriminator, False)
requires_grad(discriminator2, False)
real_img = real_imgs[0]
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
fake_img, _ = generator(noise)
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = discriminator(fake_img)
g_loss_fake = g_nonsaturating_loss(fake_pred)
loss_dict["g"] = g_loss_fake
generator.zero_grad()
g_loss_fake.backward()
g_optim.step()
g_regularize = args.g_reg_every > 0 and i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing,
device)
fake_img, latents = generator(noise, return_latents=True)
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
mean_path_length_avg = (reduce_sum(mean_path_length).item() /
get_world_size())
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
# Update EMA
ema_nimg = args.ema_kimg * 1000
if args.ema_rampup is not None:
ema_nimg = min(ema_nimg, i * args.batch * args.ema_rampup)
accum = 0.5**(args.batch / max(ema_nimg, 1e-8))
accumulate(e_ema, e_module, accum)
accumulate(g_ema, g_module, accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
g_loss_val = loss_reduced["g"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
recx_score_val = loss_reduced["recx_score"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
pix_loss_val = loss_reduced["pix"].mean().item()
vgg_loss_val = loss_reduced["vgg"].mean().item()
adv_loss_val = loss_reduced["adv"].mean().item()
avg_pix_loss.update(pix_loss_val, real_img.shape[0])
avg_vgg_loss.update(vgg_loss_val, real_img.shape[0])
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}; "
f"pix: {pix_loss_val:.4f}; vgg: {vgg_loss_val:.4f}; adv: {adv_loss_val:.4f}"
))
if i % args.log_every == 0:
with torch.no_grad():
latent_x, _ = e_ema(sample_x)
fake_x, _ = generator([latent_x],
input_is_latent=True,
return_latents=False)
sample_pix_loss = torch.sum((sample_x - fake_x)**2)
with open(os.path.join(args.log_dir, 'log.txt'), 'a+') as f:
f.write((
f"{i:07d}; pix: {avg_pix_loss.avg:.4f}; vgg: {avg_vgg_loss.avg:.4f}; ref: {sample_pix_loss.item():.4f}; "
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"path: {path_loss_val:.4f}; mean_path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}; {'; '.join([f'{k}: {r_t_dict[k]:.4f}' for k in r_t_dict])}; "
f"real_score: {real_score_val:.4f}; fake_score: {fake_score_val:.4f}; recx_score: {recx_score_val:.4f}; "
f"real_diff: {real_diff/count:.4f}; fake_diff: {fake_diff/count:.4f};\n"
))
real_diff = fake_diff = count = 0
if args.eval_every > 0 and i % args.eval_every == 0:
with torch.no_grad():
fid_sa = fid_re = fid_hy = 0
# Sample FID
g_ema.eval()
if args.truncation < 1:
mean_latent = g_ema.mean_latent(4096)
features = extract_feature_from_samples(
g_ema, inception, args.truncation, mean_latent, 64,
args.n_sample_fid, args.device).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid_sa = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
# Recon FID
features = extract_feature_from_reconstruction(
e_ema,
g_ema,
inception,
args.truncation,
mean_latent,
loader2,
args.device,
mode='recon',
).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid_re = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
# Hybrid FID
if args.eval_hybrid:
features = extract_feature_from_reconstruction(
e_ema,
g_ema,
inception,
args.truncation,
mean_latent,
loader2,
args.device,
mode='hybrid', # shuffle_idx=fid_batch_idx
).numpy()
sample_mean = np.mean(features, 0)
sample_cov = np.cov(features, rowvar=False)
fid_hy = calc_fid(sample_mean, sample_cov, real_mean,
real_cov)
# print("Sample FID:", fid_sa, "Recon FID:", fid_re, "Hybrid FID:", fid_hy)
with open(os.path.join(args.log_dir, 'log_fid.txt'),
'a+') as f:
f.write(
f"{i:07d}; sample fid: {float(fid_sa):.4f}; recon fid: {float(fid_re):.4f}; hybrid fid: {float(fid_hy):.4f};\n"
)
if wandb and args.wandb:
wandb.log({
"Generator": g_loss_val,
"Discriminator": d_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1": r1_val,
"Path Length Regularization": path_loss_val,
"Mean Path Length": mean_path_length,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
"Path Length": path_length_val,
})
if i % args.log_every == 0:
with torch.no_grad():
g_ema.eval()
e_ema.eval()
nrow = int(args.n_sample**0.5)
nchw = list(sample_x.shape)[1:]
# Fixed fake samples
sample, _ = g_ema([sample_z])
utils.save_image(
sample,
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-sample.png"),
nrow=int(args.n_sample**0.5),
normalize=True,
value_range=(-1, 1),
)
# Reconstruction samples
latent_real, _ = e_ema(sample_x)
fake_img, _ = g_ema([latent_real],
input_is_latent=True,
return_latents=False)
sample = torch.cat(
(sample_x.reshape(args.n_sample // nrow, nrow, *nchw),
fake_img.reshape(args.n_sample // nrow, nrow, *nchw)),
1)
utils.save_image(
sample.reshape(2 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-recon.png"),
nrow=nrow,
normalize=True,
value_range=(-1, 1),
)
# Hybrid samples: [real_y1, real_y2; real_x1, fake_x2]
if args.eval_hybrid:
w1, _ = e_ema(sample_x1)
w2, _ = e_ema(sample_x2)
dw = w2 - w1
dw = torch.cat(
dw.chunk(2, 0)[::-1],
0) if sample_idx is None else dw[sample_idx, ...]
fake_img, _ = g_ema([w1 + dw],
input_is_latent=True,
return_latents=False)
drive = torch.cat(
(torch.cat(sample_x1.chunk(2, 0)[::-1], 0).reshape(
args.n_sample, 1, *nchw),
torch.cat(sample_x2.chunk(2, 0)[::-1], 0).reshape(
args.n_sample, 1, *nchw)), 1)
source = torch.cat(
(sample_x1.reshape(args.n_sample, 1, *nchw),
fake_img.reshape(args.n_sample, 1, *nchw)), 1)
sample = torch.cat(
(drive.reshape(args.n_sample // nrow, 2 * nrow, *
nchw),
source.reshape(args.n_sample // nrow, 2 * nrow, *
nchw)), 1)
utils.save_image(
sample.reshape(4 * args.n_sample, *nchw),
os.path.join(args.log_dir, 'sample',
f"{str(i).zfill(6)}-cross.png"),
nrow=2 * nrow,
normalize=True,
value_range=(-1, 1),
)
if i % args.save_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"d2":
d2_module.state_dict() if args.decouple_d else None,
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"d2_optim":
d2_optim.state_dict() if args.decouple_d else None,
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight',
f"{str(i).zfill(6)}.pt"),
)
if i % args.save_latest_every == 0:
torch.save(
{
"g": g_module.state_dict(),
"e": e_module.state_dict(),
"d": d_module.state_dict(),
"d2":
d2_module.state_dict() if args.decouple_d else None,
"g_ema": g_ema.state_dict(),
"e_ema": e_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"e_optim": e_optim.state_dict(),
"d_optim": d_optim.state_dict(),
"d2_optim":
d2_optim.state_dict() if args.decouple_d else None,
"args": args,
"ada_aug_p": ada_aug_p,
"iter": i,
},
os.path.join(args.log_dir, 'weight', f"latest.pt"),
)
def train(args, loader, generator, netsD, g_optim, rf_opt, info_opt, g_ema,
device):
loader = sample_data(loader)
pbar = range(args.iter)
if get_rank() == 0:
pbar = tqdm(pbar,
initial=args.start_iter,
dynamic_ncols=True,
smoothing=0.01)
mean_path_length = 0
d_loss_val = 0
r1_loss = torch.tensor(0.0, device=device)
g_loss_val = 0
path_loss = torch.tensor(0.0, device=device)
path_lengths = torch.tensor(0.0, device=device)
mean_path_length_avg = 0
loss_dict = {}
if args.distributed:
g_module = generator.module
d_module1 = netsD[1].module
d_module2 = netsD[2].module
else:
g_module = generator
d_module1 = netsD[1]
d_module2 = netsD[2]
accum = 0.5**(32 / (10 * 1000))
ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
r_t_stat = 0
if args.augment and args.augment_p == 0:
ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 256,
device)
sample_z = torch.randn(args.n_sample, args.latent, device=device)
criterion_class = nn.CrossEntropyLoss()
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Done!")
break
real_img = next(loader)
real_img = real_img.to(device)
############# train child discriminator #############
requires_grad(generator, False)
requires_grad(netsD[2], True)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
c_code = sample_c_code(args.batch, args.c_categories, device)
# fake_img, _ = generator(noise, c_code)
image_li, _, _ = generator(noise, c_code)
fake_img = image_li[0]
if args.augment:
real_img_aug, _ = augment(real_img, ada_aug_p)
fake_img, _ = augment(fake_img, ada_aug_p)
else:
real_img_aug = real_img
fake_pred = netsD[2](fake_img)[0]
real_pred = netsD[2](real_img)[0]
d_loss = d_logistic_loss(real_pred, fake_pred)
loss_dict["d"] = d_loss
loss_dict["real_score"] = real_pred.mean()
loss_dict["fake_score"] = fake_pred.mean()
netsD[2].zero_grad()
d_loss.backward()
rf_opt[2].step()
if args.augment and args.augment_p == 0:
ada_aug_p = ada_augment.tune(real_pred)
r_t_stat = ada_augment.r_t_stat
d_regularize = i % args.d_reg_every == 0
if d_regularize:
real_img.requires_grad = True
real_pred = netsD[2](real_img)[0]
r1_loss = d_r1_loss(real_pred, real_img)
netsD[2].zero_grad()
(args.r1 / 2 * r1_loss * args.d_reg_every +
0 * real_pred[0]).backward()
rf_opt[2].step()
loss_dict["r1"] = r1_loss
############# train generator and info discriminator #############
requires_grad(generator, True)
requires_grad(netsD[1], True)
requires_grad(netsD[2], True)
noise = mixing_noise(args.batch, args.latent, args.mixing, device)
c_code = sample_c_code(args.batch, args.c_categories, device)
image_li, code_li, _ = generator(noise, c_code)
fake_img = image_li[0]
mkd_images = image_li[2]
masks = image_li[3]
p_code = code_li[1]
c_code = code_li[2]
if args.augment:
fake_img, _ = augment(fake_img, ada_aug_p)
fake_pred = netsD[2](fake_img)[0]
g_loss = g_nonsaturating_loss(fake_pred)
loss_dict["g"] = g_loss
pred_p = netsD[1](mkd_images[1])[1]
p_info_loss = criterion_class(
pred_p,
torch.nonzero(p_code.long(), as_tuple=False)[:, 1])
pred_c = netsD[2](mkd_images[2])[1]
c_info_loss = criterion_class(
pred_c,
torch.nonzero(c_code.long(), as_tuple=False)[:, 1])
loss_dict["p_info"] = p_info_loss
loss_dict["c_info"] = c_info_loss
binary_loss = binarization_loss(masks[1]) * 2e1
# oob_loss = torch.sum(bg_mk * ch_mk, dim=(-1,-2)).mean() * 1e-2
ms = masks[1].size()
min_fg_cvg = 0.2 * ms[2] * ms[3]
fg_cvg_loss = F.relu(min_fg_cvg -
torch.sum(masks[1], dim=(-1, -2))).mean() * 1e-2
ms = masks[1].size()
min_bg_cvg = 0.2 * ms[2] * ms[3]
bg_cvg_loss = F.relu(min_bg_cvg - torch.sum(
torch.ones_like(masks[1]) - masks[1], dim=(-1, -2))).mean() * 1e-2
loss_dict["bin"] = binary_loss
loss_dict["cvg"] = fg_cvg_loss + bg_cvg_loss
generator_loss = g_loss + p_info_loss + c_info_loss + binary_loss + fg_cvg_loss + bg_cvg_loss
generator.zero_grad()
netsD[1].zero_grad()
netsD[2].zero_grad()
generator_loss.backward()
g_optim.step()
info_opt[1].step()
info_opt[2].step()
g_regularize = i % args.g_reg_every == 0
if g_regularize:
path_batch_size = max(1, args.batch // args.path_batch_shrink)
noise = mixing_noise(path_batch_size, args.latent, args.mixing,
device)
c_code = sample_c_code(path_batch_size, args.c_categories, device)
image_li, _, latents = generator(noise,
c_code,
return_latents=True)
fake_img = image_li[0]
path_loss, mean_path_length, path_lengths = g_path_regularize(
fake_img, latents, mean_path_length)
generator.zero_grad()
weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss
if args.path_batch_shrink:
weighted_path_loss += 0 * fake_img[0, 0, 0, 0]
weighted_path_loss.backward()
g_optim.step()
mean_path_length_avg = (reduce_sum(mean_path_length).item() /
get_world_size())
loss_dict["path"] = path_loss
loss_dict["path_length"] = path_lengths.mean()
accumulate(g_ema, g_module, accum)
loss_reduced = reduce_loss_dict(loss_dict)
d_loss_val = loss_reduced["d"].mean().item()
g_loss_val = loss_reduced["g"].mean().item()
p_info_loss_val = loss_reduced["p_info"].mean().item()
c_info_loss_val = loss_reduced["c_info"].mean().item()
binary_loss_val = loss_reduced["bin"].mean().item()
cvg_loss_val = loss_reduced["cvg"].mean().item()
r1_val = loss_reduced["r1"].mean().item()
path_loss_val = loss_reduced["path"].mean().item()
real_score_val = loss_reduced["real_score"].mean().item()
fake_score_val = loss_reduced["fake_score"].mean().item()
path_length_val = loss_reduced["path_length"].mean().item()
if get_rank() == 0:
pbar.set_description((
f"d: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; "
f"p_info: {p_info_loss_val:.4f}; c_info: {c_info_loss_val:.4f}; "
f"bin: {binary_loss_val:.4f}; cvg: {cvg_loss_val:.4f}; "
f"path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}; "
f"augment: {ada_aug_p:.4f}"))
if wandb and args.wandb:
wandb.log({
"Generator": g_loss_val,
"Discriminator": d_loss_val,
"p_info": p_info_loss_val,
"c_info": c_info_loss_val,
"Augment": ada_aug_p,
"Rt": r_t_stat,
"R1": r1_val,
"Path Length Regularization": path_loss_val,
"Mean Path Length": mean_path_length,
"Real Score": real_score_val,
"Fake Score": fake_score_val,
"Path Length": path_length_val,
})
if i % 1000 == 0:
with torch.no_grad():
g_ema.eval()
c_code = sample_c_code(args.n_sample, args.c_categories,
device)
image_li, _, _ = g_ema([sample_z], c_code)
utils.save_image(
image_li[0],
f"sample/{str(i).zfill(6)}_0.png",
nrow=int(args.n_sample**0.5),
normalize=True,
range=(-1, 1),
)
for j in range(3):
utils.save_image(
image_li[1][j],
f"sample/{str(i).zfill(6)}_{str(1+j)}.png",
nrow=int(args.n_sample**0.5),
normalize=True,
range=(-1, 1),
)
for j in range(3):
utils.save_image(
image_li[2][j],
f"sample/{str(i).zfill(6)}_{str(4+j)}.png",
nrow=int(args.n_sample**0.5),
normalize=True,
range=(-1, 1),
)
for j in range(2):
utils.save_image(
image_li[3][j],
f"sample/{str(i).zfill(6)}_{str(7+j)}.png",
nrow=int(args.n_sample**0.5),
normalize=True,
range=(0, 1),
)
if i % 10000 == 0:
torch.save(
{
"g": g_module.state_dict(),
# "d0": d_module0.state_dict(),
"d1": d_module1.state_dict(),
"d2": d_module2.state_dict(),
"g_ema": g_ema.state_dict(),
"g_optim": g_optim.state_dict(),
"rf_optim2": rf_opt[2].state_dict(),
"info_optim1": info_opt[1].state_dict(),
"info_optim2": info_opt[2].state_dict(),
"args": args,
"ada_aug_p": ada_aug_p,
},
f"checkpoint/{str(i).zfill(6)}.pt",
)