def style_transfer(subject, style, output, iterations):
pool_method = ['max', 'avg']
init_noise = 0.0
default_subject_weights = [(9, 1)]
default_style_weights = [(0, 1), (2, 1), (4, 1), (8, 1), (12, 1)]
subject_ratio = 2e-2
smoothness = 5e-8
learn_rate = 2.0
animation = 'animation'
transfer_learning = 'imagenet-vgg-verydeep-19.mat'
np.random.seed(None)
layers, pixel_mean = vgg_net(transfer_learning, pool_method=pool_method[1])
# Inputs
style_img = imread(style) - pixel_mean
subject_img = imread(subject) - pixel_mean
init_img = subject_img
noise = np.random.normal(size=init_img.shape, scale=np.std(init_img) * 1e-1)
init_img = init_img * (1 - init_noise) + noise * init_noise
# Setup network
subject_weights = weight_array(default_subject_weights) * subject_ratio
style_weights = weight_array(default_style_weights)
net = StyleNetwork(layers, to_bc01(init_img), to_bc01(subject_img),
to_bc01(style_img), subject_weights, style_weights,
smoothness)
# Repaint image
def net_img():
return to_rgb(net.image) + pixel_mean
if not os.path.exists(animation):
os.mkdir(animation)
params = net.params
learn_rule = dp.Adam(learn_rate=learn_rate)
learn_rule_states = [learn_rule.init_state(p) for p in params]
print 'working on {0} -> {1}'.format(style, subject)
for i in tqdm(range(iterations)):
imsave(os.path.join(animation, '%.4d.png' % i), net_img())
cost = np.mean(net.update())
for param, state in zip(params, learn_rule_states):
learn_rule.step(param, state)
# print('Iteration: {0}/{1}, cost: {2:.4f}'.format(i, iterations, cost))
imsave(output, net_img())
def run():
parser = argparse.ArgumentParser(
description='Neural artistic style. Generates an image by combining '
'the subject from one image and the style from another.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--subject',
required=True,
type=str,
help='Subject image.')
parser.add_argument('--style',
required=True,
type=str,
help='Style image.')
parser.add_argument('--output',
default='out.png',
type=str,
help='Output image.')
parser.add_argument('--init',
default=None,
type=str,
help='Initial image. Subject is chosen as default.')
parser.add_argument('--init-noise',
default=0.0,
type=float_range,
help='Weight between [0, 1] to adjust the noise level '
'in the initial image.')
parser.add_argument('--random-seed',
default=None,
type=int,
help='Random state.')
parser.add_argument('--animation',
default='animation',
type=str,
help='Output animation directory.')
parser.add_argument('--iterations',
default=500,
type=int,
help='Number of iterations to run.')
parser.add_argument('--learn-rate',
default=2.0,
type=float,
help='Learning rate.')
parser.add_argument('--smoothness',
type=float,
default=2e-7,
help='Weight of smoothing scheme.')
parser.add_argument('--subject-weights',
nargs='*',
type=weight_tuple,
default=[(9, 1)],
help='List of subject weights (conv_idx,weight).')
parser.add_argument('--style-weights',
nargs='*',
type=weight_tuple,
default=[(0, 1), (2, 1), (4, 1), (8, 1), (12, 1)],
help='List of style weights (conv_idx,weight).')
parser.add_argument('--subject-ratio',
type=float,
default=2e-2,
help='Weight of subject relative to style.')
parser.add_argument('--pool-method',
default='avg',
type=str,
choices=['max', 'avg'],
help='Subsampling scheme.')
parser.add_argument('--vgg19',
default='imagenet-vgg-verydeep-19.mat',
type=str,
help='VGG-19 .mat file.')
args = parser.parse_args()
if args.random_seed is not None:
np.random.seed(args.random_seed)
layers, img_mean = vgg19_net(args.vgg19, pool_method=args.pool_method)
# Inputs
pixel_mean = np.mean(img_mean, axis=(0, 1))
style_img = imread(args.style) - pixel_mean
subject_img = imread(args.subject) - pixel_mean
if args.init is None:
init_img = subject_img
else:
init_img = imread(args.init) - pixel_mean
noise = np.random.normal(size=init_img.shape, scale=np.std(init_img))
init_img = init_img * (1 - args.init_noise) + noise * args.init_noise
# Setup network
subject_weights = weight_array(args.subject_weights) * args.subject_ratio
style_weights = weight_array(args.style_weights)
net = StyleNetwork(layers, to_bc01(init_img), to_bc01(subject_img),
to_bc01(style_img), subject_weights, style_weights,
args.smoothness)
# Repaint image
def net_img():
return to_rgb(net.image) + pixel_mean
if not os.path.exists(args.animation):
os.mkdir(args.animation)
params = net._params
learn_rule = dp.Adam(learn_rate=args.learn_rate)
learn_rule_states = [learn_rule.init_state(p) for p in params]
for i in range(args.iterations):
imsave(os.path.join(args.animation, '%.4d.png' % i), net_img())
cost = np.mean(net._update())
for param, state in zip(params, learn_rule_states):
learn_rule.step(param, state)
print('Iteration: %i, cost: %.4f' % (i, cost))
imsave(args.output, net_img())
def process_image_pair(**kwargs):
start = time.time()
if kwargs['random_seed'] is not None:
np.random.seed(kwargs['random_seed'])
print('loading mat net...')
layers, img_mean = vgg19_net(kwargs['vgg19'],
pool_method='avg') #kwargs['pool_method'])
# Inputs
pixel_mean = np.mean(img_mean, axis=(0, 1))
style_img = imread(kwargs['style_path']) - pixel_mean
subject_img = imread(kwargs['content_path']) - pixel_mean
# if kwargs['init'] is None:
# init_img = subject_img
# else:
# init_img = imread(kwargs['init']) - pixel_mean
init_img = subject_img
noise = np.random.normal(size=init_img.shape,
scale=np.std(init_img) * 1e-1)
init_img = init_img * (1 -
kwargs['init_noise']) + noise * kwargs['init_noise']
# Setup network
print('building network...')
#FIXME: Set defaults
subject_weights = weight_array(
kwargs['subject_weights']) * kwargs['subject_ratio']
style_weights = weight_array(kwargs['style_weights'])
net = StyleNetwork(layers, to_bc01(init_img), to_bc01(subject_img),
to_bc01(style_img), subject_weights, style_weights,
kwargs['smoothness'])
# Repaint image
def net_img():
return to_rgb(net.image) + pixel_mean
if kwargs['animation']:
os.mkdir(kwargs['animation_dir'])
params = net.params
learn_rule = dp.Adam(learn_rate=kwargs['learn_rate'])
learn_rule_states = [learn_rule.init_state(p) for p in params]
for i in range(kwargs['iterations']):
print('iteration %d', i)
if kwargs['animation']:
imsave(os.path.join(kwargs['animation_dir'], '%.4d.png' % i),
net_img())
cost = np.mean(net.update())
for param, state in zip(params, learn_rule_states):
learn_rule.step(param, state)
print('Iteration: %i, cost: %.4f' % (i, cost))
imsave(kwargs['output_path'], net_img())
print("Time Elapsed: %0.2f", time.time() - start)
def run():
parser = argparse.ArgumentParser(
description='Neural artistic style. Generates an image by combining '
'the subject from one image and the style from another.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--subject',
required=True,
type=str,
help='Subject image.')
parser.add_argument('--style',
required=True,
type=str,
help='Style image.')
parser.add_argument('--output',
default='out.png',
type=str,
help='Output image.')
parser.add_argument('--init',
default=None,
type=str,
help='Initial image. Subject is chosen as default.')
parser.add_argument('--init-noise',
default=0.0,
type=float_range,
help='Weight between [0, 1] to adjust the noise level '
'in the initial image.')
parser.add_argument('--random-seed',
default=None,
type=int,
help='Random state.')
parser.add_argument('--animation',
default='animation',
type=str,
help='Output animation directory.')
parser.add_argument('--iterations',
default=500,
type=int,
help='Number of iterations to run.')
parser.add_argument('--scales',
default=4,
type=int,
help='Number of scales to use.')
parser.add_argument('--learn-rate',
default=2.0,
type=float,
help='Learning rate.')
parser.add_argument('--smoothness',
type=float,
default=5e-8,
help='Weight of smoothing scheme.')
parser.add_argument('--subject-weights',
nargs='*',
type=weight_tuple,
default=[(9, 1)],
help='List of subject weights (conv_idx,weight).')
parser.add_argument('--style-weights',
nargs='*',
type=weight_tuple,
default=[(0, 1), (2, 1), (4, 1), (8, 1), (12, 1)],
help='List of style weights (conv_idx,weight).')
parser.add_argument('--subject-ratio',
type=float,
default=2e-2,
help='Weight of subject relative to style.')
parser.add_argument('--pool-method',
default='avg',
type=str,
choices=['max', 'avg'],
help='Subsampling scheme.')
parser.add_argument('--network',
default='imagenet-vgg-verydeep-19.mat',
type=str,
help='Network in MatConvNet format).')
parser.add_argument('--outer_it',
default=8000,
type=int,
help='Outer iterations.')
parser.add_argument('--inner_it',
default=3,
type=int,
help='Inner iterations.')
parser.add_argument('--patch_size',
default=512,
type=int,
help='Patchsize.')
args = parser.parse_args()
if args.random_seed is not None:
np.random.seed(args.random_seed)
layers, pixel_mean = vgg_net(args.network, pool_method=args.pool_method)
# Inputs
style_img = imread(args.style) - pixel_mean
subject_img = imread(args.subject) - pixel_mean
print "style_img", style_img.shape
print "subject_img", subject_img.shape
if args.init is None: init_img = subject_img
else: init_img = imread(args.init) - pixel_mean
noise = np.random.normal(size=init_img.shape,
scale=np.std(init_img) * 1e-1)
init_img = init_img * (1 - args.init_noise) + noise * args.init_noise
# Setup network
subject_weights = weight_array(args.subject_weights) * args.subject_ratio
style_weights = weight_array(args.style_weights)
max_patch_size = np.array([args.patch_size, args.patch_size])
inner_iterations = 10
for outer_it in range(0, args.outer_it):
scale = 1.0
for sc in range(0, args.scales - 1):
if random.random() < 0.25: scale *= 2.0
print "OUTER_IT: ", outer_it, "SCALE:", scale
effective_patch_size = max_patch_size * scale
patch_size = (np.min([
style_img.shape[0], subject_img.shape[0], init_img.shape[0],
effective_patch_size[0]
]),
np.min([
style_img.shape[1], subject_img.shape[1],
init_img.shape[1], effective_patch_size[1]
]))
rel_x = random.random()
rel_y = random.random()
x = int(math.floor(rel_x * (style_img.shape[0] - patch_size[0])))
y = int(math.floor(rel_y * (style_img.shape[1] - patch_size[1])))
style_patch = style_img[x:x + patch_size[0], y:y + patch_size[1], :]
x = int(math.floor(rel_x * (subject_img.shape[0] - patch_size[0])))
y = int(math.floor(rel_y * (subject_img.shape[1] - patch_size[1])))
subject_patch = subject_img[x:x + patch_size[0],
y:y + patch_size[1], :]
x = int(math.floor(rel_x * (init_img.shape[0] - patch_size[0])))
y = int(math.floor(rel_y * (init_img.shape[1] - patch_size[1])))
init_patch = init_img[x:x + patch_size[0], y:y + patch_size[1], :]
style_patch_scaled = nd.zoom(style_patch,
(1.0 / scale, 1.0 / scale, 1),
order=1)
subject_patch_scaled = nd.zoom(subject_patch,
(1.0 / scale, 1.0 / scale, 1),
order=1)
init_patch_scaled = nd.zoom(init_patch, (1.0 / scale, 1.0 / scale, 1),
order=1)
net = StyleNetwork(layers, to_bc01(init_patch_scaled),
to_bc01(subject_patch_scaled),
to_bc01(style_patch_scaled), subject_weights,
style_weights, args.smoothness)
params = net._params
learn_rule = dp.Adam(learn_rate=args.learn_rate)
learn_rule_states = [learn_rule.init_state(p) for p in params]
for i in range(args.inner_it):
cost = np.mean(net._update())
for param, state in zip(params, learn_rule_states):
learn_rule.step(param, state)
print('Iteration: %i, cost: %.4f' % (i, cost))
result_patch_scaled = to_rgb(net.image) - init_patch_scaled
result_patch = nd.zoom(result_patch_scaled, (scale, scale, 1), order=1)
border_cut = 10
init_img[x + border_cut:x + result_patch.shape[0] - border_cut,
y + border_cut:y + result_patch.shape[1] -
border_cut, :] += result_patch[border_cut:-border_cut,
border_cut:-border_cut]
imsave("%s.jpg" % args.output, init_img + pixel_mean)
def run():
parser = argparse.ArgumentParser(
description='Neural artistic style.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--subject',
required=True,
type=str,
help='subject image')
parser.add_argument('--style', required=True, type=str, help='style image')
parser.add_argument('--output',
default='out.png',
type=str,
help='output image')
parser.add_argument('--animation',
default='animation',
type=str,
help='output animation directory')
parser.add_argument('--iterations',
default=500,
type=int,
help='Number of iterations')
parser.add_argument('--learn-rate',
default=5.0,
type=float,
help='Learning rate')
parser.add_argument('--subject-weights',
nargs='*',
type=weight_tuple,
default=[(9, 1)],
help='list of subject weights (conv_idx,weight)')
parser.add_argument('--style-weights',
nargs='*',
type=weight_tuple,
default=[(0, 1), (2, 1), (4, 1), (8, 1), (12, 1)],
help='list of style weights (conv_idx,weight)')
parser.add_argument('--subject-ratio',
type=float,
default=2e-2,
help='weight of subject relative to style')
parser.add_argument('--vgg19',
default='imagenet-vgg-verydeep-19.mat',
type=str,
help='VGG-19 .mat file')
args = parser.parse_args()
layers, img_mean = vgg19_net(args.vgg19, pool_method='avg')
# Inputs
pixel_mean = np.mean(img_mean, axis=(0, 1))
style_img = imread(args.style)
subject_img = imread(args.subject)
style_img -= pixel_mean
subject_img -= pixel_mean
# Setup network
subject_weights = weight_array(args.subject_weights) * args.subject_ratio
style_weights = weight_array(args.style_weights)
net = StyleNetwork(layers, to_bc01(subject_img), to_bc01(style_img),
subject_weights, style_weights)
# Repaint image
def net_img():
return to_rgb(net.image) + pixel_mean
if not os.path.exists(args.animation):
os.mkdir(args.animation)
params = net._params
learn_rule = dp.Adam(learn_rate=args.learn_rate)
learn_rule_states = [learn_rule.init_state(p) for p in params]
for i in range(args.iterations):
imsave(os.path.join(args.animation, '%.4d.png' % i), net_img())
cost = np.mean(net._update())
for param, state in zip(params, learn_rule_states):
learn_rule.step(param, state)
print('Iteration: %i, cost: %.4f' % (i, cost))
imsave(args.output, net_img())