def run_bulk(config):
device = 'cpu' if config.cpu or not torch.cuda.is_available() else 'cuda:0'
device = torch.device(device)
transfer_at = set()
if config.transfer_at_encoder:
transfer_at.add('encoder')
if config.transfer_at_decoder:
transfer_at.add('decoder')
if config.transfer_at_skip:
transfer_at.add('skip')
# The filenames of the content and style pair should match
fnames = set(os.listdir(config.content)) & set(os.listdir(config.style))
if config.content_segment and config.style_segment:
fnames &= set(os.listdir(config.content_segment))
fnames &= set(os.listdir(config.style_segment))
for fname in tqdm.tqdm(fnames):
if not fname.endswith('.png'):
print('invalid file (should end with .png), ', fname)
continue
_content = os.path.join(config.content, fname)
_style = os.path.join(config.style, fname)
_content_segment = os.path.join(config.content_segment, fname) if config.content_segment else None
_style_segment = os.path.join(config.style_segment, fname) if config.style_segment else None
_output = os.path.join(config.output, fname)
content = open_image(_content, config.image_size).to(device)
style = open_image(_style, config.image_size).to(device)
content_segment = load_segment(_content_segment, config.image_size)
style_segment = load_segment(_style_segment, config.image_size)
if not config.transfer_all:
with Timer('Elapsed time in whole WCT: {}', config.verbose):
postfix = '_'.join(sorted(list(transfer_at)))
fname_output = _output.replace('.png', '_{}_{}.png'.format(config.option_unpool, postfix))
print('------ transfer:', _output)
wct2 = WCT2(transfer_at=transfer_at, option_unpool=config.option_unpool, device=device, verbose=config.verbose)
with torch.no_grad():
img = wct2.transfer(content, style, content_segment, style_segment, alpha=config.alpha)
save_image(img.clamp_(0, 1), fname_output, padding=0)
else:
for _transfer_at in get_all_transfer():
with Timer('Elapsed time in whole WCT: {}', config.verbose):
postfix = '_'.join(sorted(list(_transfer_at)))
fname_output = _output.replace('.png', '_{}_{}.png'.format(config.option_unpool, postfix))
print('------ transfer:', fname)
wct2 = WCT2(transfer_at=_transfer_at, option_unpool=config.option_unpool, device=device, verbose=config.verbose)
with torch.no_grad():
img = wct2.transfer(content, style, content_segment, style_segment, alpha=config.alpha)
save_image(img.clamp_(0, 1), fname_output, padding=0)
def run_bulk(config):
device = 'cpu' if config.cpu or not torch.cuda.is_available() else 'cuda:0'
device = torch.device(device)
transfer_at = set()
if config.transfer_at_encoder:
transfer_at.add('encoder')
if config.transfer_at_decoder:
transfer_at.add('decoder')
if config.transfer_at_skip:
transfer_at.add('skip')
# The filenames of the content and style pair should match
fnames = set(os.listdir(config.content)) & set(os.listdir(config.style))
if config.content_segment and config.style_segment:
fnames &= set(os.listdir(config.content_segment))
fnames &= set(os.listdir(config.style_segment))
for fname in tqdm.tqdm(fnames):
if not is_image_file(fname):
print('invalid file (is not image), ', fname)
continue
_content = os.path.join(config.content, fname)
_style = os.path.join(config.style, fname)
_content_segment = os.path.join(
config.content_segment, fname) if config.content_segment else None
_style_segment = os.path.join(config.style_segment,
fname) if config.style_segment else None
_output = os.path.join(config.output, fname)
# content_transfroms = tv.transforms.Compose([
# #tv.transforms.Resize(config.image_size),
# #tv.transforms.CenterCrop(config.image_size),
# tv.transforms.ToTensor(),
# # tv.transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD),
# tv.transforms.Lambda(lambda x: x * 255)
# ])
# content_dataset = tv.datasets.ImageFolder(_content, content_transfroms)
# content_dataloader = data.DataLoader(content_dataset, 1)
#
# style_transform = tv.transforms.Compose([
# # tv.transforms.Resize(256),
# # tv.transforms.RandomCrop(256),
# tv.transforms.ToTensor(),
# tv.transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD),
# # tv.transforms.Lambda(lambda x: x*255)
# ])
#
# style_image_dataset = tv.datasets.ImageFolder(_style, style_transform)
#
# style_dataloader = data.DataLoader(style_image_dataset, 1)
#content = open_image(_content, config.image_size).to(device)
content = Image.open(_content)
content_transforms = []
content = transforms.Resize(config.image_size)(content)
# _transforms.append(transforms.Resize(image_size))
w, h = content.size
content_transforms.append(
transforms.CenterCrop((h // 16 * 16, w // 16 * 16)))
content_transforms.append(transforms.ToTensor())
content_transforms.append(
transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD))
content_transforms.append(transforms.Lambda(lambda x: x * 255))
content_transform = transforms.Compose(content_transforms)
content = content_transform(content).unsqueeze(0).to(device)
#style = open_image(_style, config.image_size).to(device)
style = Image.open(_style)
style_transforms = []
style = transforms.Resize(config.image_size)(style)
# _transforms.append(transforms.Resize(image_size))
w, h = style.size
style_transforms.append(
transforms.CenterCrop((h // 16 * 16, w // 16 * 16)))
style_transforms.append(transforms.ToTensor())
style_transforms.append(
transforms.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD))
#style_transforms.append(transforms.Lambda(lambda x: x * 255))
style_transform = transforms.Compose(style_transforms)
style = style_transform(style).unsqueeze(0).to(device)
content_segment = load_segment(_content_segment, config.image_size)
style_segment = load_segment(_style_segment, config.image_size)
_, ext = os.path.splitext(fname)
if not config.transfer_all:
with Timer('Elapsed time in whole WCT: {}', config.verbose):
postfix = '_'.join(sorted(list(transfer_at)))
fname_output = _output.replace(
ext, '_{}_{}.{}'.format(config.option_unpool, postfix,
ext))
print('------ transfer:', _output)
wct2 = WCT2(transfer_at=transfer_at,
option_unpool=config.option_unpool,
device=device,
verbose=config.verbose)
with torch.no_grad():
img = wct2.transfer(content,
style,
content_segment,
style_segment,
alpha=config.alpha)
save_image(img.clamp_(0, 1), fname_output, padding=0)
else:
for _transfer_at in get_all_transfer():
with Timer('Elapsed time in whole WCT: {}', config.verbose):
postfix = '_'.join(sorted(list(_transfer_at)))
fname_output = _output.replace(
ext, '_{}_{}.{}'.format(config.option_unpool, postfix,
ext))
print('------ transfer:', fname)
wct2 = WCT2(transfer_at=_transfer_at,
option_unpool=config.option_unpool,
device=device,
verbose=config.verbose)
with torch.no_grad():
img = wct2.transfer(content,
style,
content_segment=content_segment,
style_segment=style_segment,
alpha=config.alpha)
img = utils_.normalize_batch(img)
#img = img.data.cpu()[0] * 0.225 + 0.45
save_image(img.clamp_(0, 1), fname_output, padding=0)
def run_bulk(config):
device = 'cpu' if config.cpu or not torch.cuda.is_available() else 'cuda:0'
device = torch.device(device)
transfer_at = set()
if config.transfer_at_encoder:
transfer_at.add('encoder')
if config.transfer_at_decoder:
transfer_at.add('decoder')
if config.transfer_at_skip:
transfer_at.add('skip')
# The filenames of the content and style pair should match
fnames = set(os.listdir(config.content)) & set(os.listdir(config.style))
if config.content_segment and config.style_segment:
fnames &= set(os.listdir(config.content_segment))
fnames &= set(os.listdir(config.style_segment))
for fname in tqdm.tqdm(fnames):
if not is_image_file(fname):
print('invalid file (is not image), ', fname)
continue
_content = os.path.join(config.content, fname)
_style = os.path.join(config.style, fname)
_content_segment = os.path.join(
config.content_segment, fname) if config.content_segment else None
_style_segment = os.path.join(config.style_segment,
fname) if config.style_segment else None
_output = os.path.join(config.output, fname)
content = open_image(_content, config.image_size).to(device)
style = open_image(_style, config.image_size).to(device)
content_segment = load_segment(_content_segment, config.image_size)
style_segment = load_segment(_style_segment, config.image_size)
_, ext = os.path.splitext(fname)
if not config.transfer_all:
with Timer('Elapsed time in whole WCT: {}', config.verbose):
postfix = '_'.join(sorted(list(transfer_at)))
fname_output = _output.replace(
ext, '_{}_{}.{}'.format(config.option_unpool, postfix,
ext))
print('------ transfer:', _output)
LapSobGaus = Lap_Sob_Gaus(transfer_at=transfer_at,
option_unpool=config.option_unpool,
device=device,
verbose=config.verbose)
with torch.no_grad():
img = LapSobGaus.transfer(content,
style,
content_segment,
style_segment,
alpha=config.alpha)
save_image(img.clamp_(0, 1), fname_output, padding=0)
else:
for _transfer_at in get_all_transfer():
with Timer('Elapsed time in whole WCT: {}', config.verbose):
postfix = '_'.join(sorted(list(_transfer_at)))
fname_output = _output.replace(
ext, '_{}_{}.{}'.format(config.option_unpool, postfix,
ext))
print('------ transfer:', fname)
LapSobGaus = Lap_Sob_Gaus(
transfer_at=_transfer_at,
option_unpool=config.option_unpool,
device=device,
verbose=config.verbose)
with torch.no_grad():
# content = content * 255
# style = style * 255
img = LapSobGaus.transfer(
content,
style,
content_segment=content_segment,
style_segment=style_segment,
alpha=config.alpha,
fname_output=None)
# img = utils_.normalize_batch(img)
#img = img.data.cpu()[0] * 0.225 + 0.45
save_image(img.clamp_(0, 1), fname_output, padding=0)