def __init__(self,
npop,
target,
cpkt="stylegan2/stylegan2-ffhq-config-f.pt",
lamb=5.0,
sigma=0.4,
device="cuda",
trunc=.6,
mask=None,
ploss=False):
with torch.no_grad():
self.mask = mask
self.device = device
self.npop = npop
self.target = target
self.lamb = lamb
self.sigma = sigma
self.generator = Generator(IMG_DIM,
LATENT_DIM,
8,
channel_multiplier=2).to(self.device)
self.generator.eval()
checkpoint = torch.load(cpkt)
self.generator.load_state_dict(checkpoint['g_ema'])
sample = torch.randn(npop, LATENT_DIM, device=device)
self.trunc_target = self.generator.mean_latent(4096)
# matrix of genomes
self.trunc = trunc
self.genomes = self.trunc_target + self.trunc * (
self.generator.get_latent(sample) - self.trunc_target)
self.fitness = []
self.ranks = torch.zeros([npop], device="cpu")
self.faces = torch.zeros([npop, 3, IMG_DIM, IMG_DIM], device="cpu")
if ploss:
self.percept = lpips.PerceptualLoss(
model='net-lin',
net='vgg',
use_gpu=device.startswith('cuda'))
else:
self.percept = None
self.generate()
def main(args):
ensure_checkpoint_exists(args.ckpt)
text_inputs = torch.cat([clip.tokenize(args.description)]).cuda()
os.makedirs(args.results_dir, exist_ok=True)
g_ema = Generator(args.size, 512, 8)
g_ema.load_state_dict(torch.load(args.ckpt)["g_ema"], strict=False)
g_ema.eval()
g_ema = g_ema.cuda()
mean_latent = g_ema.mean_latent(4096)
if args.latent_path:
latent_code_init = torch.load(args.latent_path).cuda()
elif args.mode == "edit":
latent_code_init_not_trunc = torch.randn(1, 512).cuda()
with torch.no_grad():
_, latent_code_init = g_ema([latent_code_init_not_trunc],
return_latents=True,
truncation=args.truncation,
truncation_latent=mean_latent)
else:
latent_code_init = mean_latent.detach().clone().repeat(1, 18, 1)
latent = latent_code_init.detach().clone()
latent.requires_grad = True
clip_loss = CLIPLoss()
optimizer = optim.Adam([latent], lr=args.lr)
pbar = tqdm(range(args.step))
for i in pbar:
t = i / args.step
lr = get_lr(t, args.lr)
optimizer.param_groups[0]["lr"] = lr
img_gen, _ = g_ema([latent],
input_is_latent=True,
randomize_noise=False)
c_loss = clip_loss(img_gen, text_inputs)
if args.mode == "edit":
l2_loss = ((latent_code_init - latent)**2).sum()
loss = c_loss + args.l2_lambda * l2_loss
else:
loss = c_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
pbar.set_description((f"loss: {loss.item():.4f};"))
if args.save_intermediate_image_every > 0 and i % args.save_intermediate_image_every == 0:
with torch.no_grad():
img_gen, _ = g_ema([latent],
input_is_latent=True,
randomize_noise=False)
torchvision.utils.save_image(img_gen,
f"results/{str(i).zfill(5)}.png",
normalize=True,
range=(-1, 1))
if args.mode == "edit":
with torch.no_grad():
img_orig, _ = g_ema([latent_code_init],
input_is_latent=True,
randomize_noise=False)
final_result = torch.cat([img_orig, img_gen])
else:
final_result = img_gen
return final_result
if __name__ == '__main__':
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
torch.cuda.set_device(device)
encoder_HG = HG_softmax2020(num_classes=5, heatmap_size=64)
# encoder_HG.load_state_dict(torch.load("/home/ibespalov/pomoika/hg2_e29.pt", map_location="cpu"))
print("HG")
latent = 512
n_mlp = 5
size = 256
generator = CondGen3(Generator(
size, latent, n_mlp, channel_multiplier=1,
), heatmap_channels=5)
discriminator = CondDisc3(
size, heatmap_channels=5, channel_multiplier=1
)
style_encoder = StyleEncoder(style_dim=latent)
starting_model_number = 190000 # 170000
weights = torch.load(
f'{Paths.default.models()}/stylegan2_MAFL_{str(starting_model_number).zfill(6)}.pt',
# f'{Paths.default.nn()}/zhores/stylegan2_w300_{str(starting_model_number).zfill(6)}.pt',
map_location="cpu"
)
def generate(args, g_ema, device, mean_latent):
with torch.no_grad():
g_ema.eval()
for i in tqdm(range(args.pics)):
sample_z = torch.randn(args.sample, args.latent, device=device)
sample, _ = g_ema(
[sample_z], truncation=args.truncation, truncation_latent=mean_latent
)
utils.save_image(
sample,
f"sample/{str(i).zfill(6)}.png",
nrow=1,
normalize=True,
range=(-1, 1),
)
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
stylegan_gen = Generator(size=1024, style_dim=512, n_mlp=8, channel_multiplier=2).to(device)
checkpoint = torch.load('./twdne3.pt')
stylegan_gen.load_state_dict(checkpoint["g_ema"])
mean_latent = None
#generate(args, stylegan_gen, device, mean_latent)
torch.cuda.set_device(device)
cont_style_encoder: MunitEncoder = cont_style_munit_enc(
munit_args,
None, # "/home/ibespalov/pomoika/munit_content_encoder15.pt",
None # "/home/ibespalov/pomoika/munit_style_encoder_1.pt"
) #.to(device)
args.latent = 512
args.n_mlp = 5
args.start_iter = 0
generator = CondGen2(
Generator(args.size,
args.latent,
args.n_mlp,
channel_multiplier=args.channel_multiplier)) #.to(device)
discriminator = CondStyleDisc2Wrapper(
Discriminator(
args.size,
channel_multiplier=args.channel_multiplier)) #.to(device)
g_reg_ratio = args.g_reg_every / (args.g_reg_every + 1)
d_reg_ratio = args.d_reg_every / (args.d_reg_every + 1)
image_size = args.size
transform = albumentations.Compose([
albumentations.HorizontalFlip(),
albumentations.Resize(image_size, image_size),
albumentations.ElasticTransform(p=0.5,
print(device)
torch.cuda.set_device(device)
encoder_HG = HG_softmax2020(num_classes=68, heatmap_size=64)
encoder_HG.load_state_dict(
torch.load("/trinity/home/n.buzun/PycharmProjects/saved/hg2_e29.pt",
map_location="cpu"))
encoder_HG = encoder_HG.cuda()
print("HG")
latent = 512
n_mlp = 5
size = 256
generator = CondGen3(Generator(size, latent, n_mlp, channel_multiplier=1))
discriminator = CondDisc3(size, channel_multiplier=1)
starting_model_number = 110000
generator = generator.cuda()
discriminator = discriminator.to(device)
generator = nn.DataParallel(generator, [0, 1, 2, 3])
discriminator = nn.DataParallel(discriminator, [0, 1, 2, 3])
encoder_HG = nn.DataParallel(encoder_HG, [0, 1, 2, 3])
style_encoder = StyleEncoder(style_dim=latent).cuda()
weights = torch.load(
vae_ckpt = torch.load(args.vae)
vae_args = vae_ckpt['args']
vae = VAE(3, n_latent=vae_args.n_latent, size=vae_args.input_size)
vae.load_state_dict(vae_ckpt['vae'])
vae = vae.enc.to(device)
args.latent = 512
args.n_mlp = 8
args.start_iter = 0
generator = Generator(
args.size,
args.latent,
args.n_mlp,
channel_multiplier=args.channel_multiplier,
style_in_dim=args.latent + vae_args.n_latent,
).to(device)
discriminator = Discriminator(
args.size, channel_multiplier=args.channel_multiplier).to(device)
g_ema = Generator(
args.size,
args.latent,
args.n_mlp,
channel_multiplier=args.channel_multiplier,
style_in_dim=args.latent + vae_args.n_latent,
).to(device)
g_ema.eval()
accumulate(g_ema, generator, 0)
parser.add_argument('--space', choices=['z', 'w'])
parser.add_argument('--batch', type=int, default=64)
parser.add_argument('--n_sample', type=int, default=5000)
parser.add_argument('--size', type=int, default=256)
parser.add_argument('--eps', type=float, default=1e-4)
parser.add_argument('--crop', action='store_true')
parser.add_argument('ckpt', metavar='CHECKPOINT')
args = parser.parse_args()
latent_dim = 512
ckpt = torch.load(args.ckpt)
g = Generator(args.size, latent_dim, 8).to(device)
g.load_state_dict(ckpt['g_ema'])
g.eval()
percept = lpips.PerceptualLoss(model='net-lin',
net='vgg',
use_gpu=device.startswith('cuda'))
distances = []
n_batch = args.n_sample // args.batch
resid = args.n_sample - (n_batch * args.batch)
batch_sizes = [args.batch] * n_batch + [resid]
with torch.no_grad():
for batch in tqdm(batch_sizes):
import torch
from stylegan2.model import Generator
from stylegan2 import ppl
import matplotlib.pyplot as plt
import numpy as np
# Test out style mixing, interpolation
# Load the model
device = "cuda"
g_ema = Generator(1024, 512, 8, channel_multiplier=2).to(device)
checkpoint = torch.load("stylegan2/stylegan2-ffhq-config-f.pt")
g_ema.load_state_dict(checkpoint['g_ema'])
with torch.no_grad():
# Test new mutation heuristic
latent1 = torch.randn(1, 512, device=device)
latent2 = torch.randn(1, 512, device=device)
trunc_target = g_ema.mean_latent(4096)
w1 = g_ema.get_latent(latent1)
w1 = trunc_target + .7 * (w1 - trunc_target)
w2 = g_ema.get_latent(latent2)
trunc_target = g_ema.mean_latent(4096)
fig, axs = plt.subplots(3, 3)
for i in range(8):
torch.cuda.empty_cache()
im = w1.clone()
p1 = 64 * i
p2 = 64 * (i + 1)
diff = (w2[0, :] - trunc_target)