|
import torch |
|
from torch import nn |
|
import torch.optim as optim |
|
import torch.nn.functional as F |
|
from torch.utils.data.dataloader import DataLoader |
|
from torchvision import transforms |
|
from torchvision import utils as vutils |
|
|
|
from models import Generator |
|
from utils import copy_G_params, load_params |
|
|
|
|
|
|
|
def get_early_features(net, noise): |
|
with torch.no_grad(): |
|
feat_4 = net._init(noise) |
|
feat_8 = net._upsample_8(feat_4) |
|
feat_16 = net._upsample_16(feat_8) |
|
feat_32 = net._upsample_32(feat_16) |
|
feat_64 = net._upsample_64(feat_32) |
|
return feat_8, feat_16, feat_32, feat_64 |
|
|
|
def get_late_features(net, feat_64, feat_8, feat_16, feat_32): |
|
with torch.no_grad(): |
|
feat_128 = net._upsample_128(feat_64) |
|
feat_128 = net._sle_128(feat_8, feat_128) |
|
|
|
feat_256 = net._upsample_256(feat_128) |
|
feat_256 = net._sle_256(feat_16, feat_256) |
|
|
|
feat_512 = net._upsample_512(feat_256) |
|
feat_512 = net._sle_512(feat_32, feat_512) |
|
|
|
feat_1024 = net._upsample_1024(feat_512) |
|
|
|
return net._out_1024(feat_1024) |
|
|
|
def style_mix(model_name_or_path, bs, device): |
|
_in_channels = 256 |
|
im_size = 1024 |
|
|
|
netG = Generator(in_channels=_in_channels, out_channels=3) |
|
netG = netG.from_pretrained(model_name_or_path, in_channels=256, out_channels=3) |
|
_ = netG.to(device) |
|
_ = netG.eval() |
|
|
|
avg_param_G = copy_G_params(netG) |
|
load_params(netG, avg_param_G) |
|
|
|
noise_a = torch.randn(bs, 256, 1, 1, device=device).to(device) |
|
noise_b = torch.randn(bs, 256, 1, 1, device=device).to(device) |
|
|
|
feat_8_a, feat_16_a, feat_32_a, feat_64_a = get_early_features(netG, noise_a) |
|
feat_8_b, feat_16_b, feat_32_b, feat_64_b = get_early_features(netG, noise_b) |
|
|
|
images_b = get_late_features(netG, feat_64_b, feat_8_b, feat_16_b, feat_32_b) |
|
images_a = get_late_features(netG, feat_64_a, feat_8_a, feat_16_a, feat_32_a) |
|
|
|
imgs = [ torch.ones(1, 3, im_size, im_size) ] |
|
|
|
imgs.append(images_b.cpu()) |
|
for i in range(bs): |
|
imgs.append(images_a[i].unsqueeze(0).cpu()) |
|
gimgs = get_late_features(netG, feat_64_a[i].unsqueeze(0).repeat(bs, 1, 1, 1), feat_8_b, feat_16_b, feat_32_b) |
|
imgs.append(gimgs.cpu()) |
|
|
|
imgs = torch.cat(imgs) |
|
|
|
|
|
return imgs |
|
|
