def get_infile():
pydiffvg.set_use_gpu(False)
root = tk.Tk()
#root.withdraw()
file_path = filedialog.askopenfilename(initialdir = ".",title = "Select graphic to optimize",filetypes = (("SVG files","*.svg"),("all files","*.*")))
root.destroy()
return file_path
import pydiffvg
import torch
import skimage
import numpy as np
# Use GPU if available
pydiffvg.set_use_gpu(torch.cuda.is_available())
canvas_width, canvas_height = 256, 256
num_control_points = torch.tensor([2])
# points = torch.tensor([[120.0, 30.0], # base
# [150.0, 60.0], # control point
# [ 90.0, 198.0], # control point
# [ 60.0, 218.0], # base
# [ 90.0, 180.0], # control point
# [200.0, 65.0], # control point
# [210.0, 98.0], # base
# [220.0, 70.0], # control point
# [130.0, 55.0]]) # control point
points = torch.tensor([
[20.0, 128.0], # base
[50.0, 128.0], # control point
[170.0, 128.0], # control point
[200.0, 128.0]
]) # base
path = pydiffvg.Path(num_control_points=num_control_points,
points=points,
is_closed=False,
stroke_width=torch.tensor(10.0))
shapes = [path]
path_group = pydiffvg.ShapeGroup(shape_ids=torch.tensor([0]),
def main(args):
pydiffvg.set_use_gpu(torch.cuda.is_available())
canvas_width, canvas_height, shapes, shape_groups = pydiffvg.svg_to_scene(
args.content_file)
scene_args = pydiffvg.RenderFunction.serialize_scene(
canvas_width, canvas_height, shapes, shape_groups)
render = pydiffvg.RenderFunction.apply
img = render(
canvas_width, # width
canvas_height, # height
2, # num_samples_x
2, # num_samples_y
0, # seed
None,
*scene_args)
# Transform to gamma space
pydiffvg.imwrite(img.cpu(), 'results/style_transfer/init.png', gamma=1.0)
# HWC -> NCHW
img = img.unsqueeze(0)
img = img.permute(0, 3, 1, 2) # NHWC -> NCHW
loader = transforms.Compose([transforms.ToTensor()
]) # transform it into a torch tensor
def image_loader(image_name):
image = Image.open(image_name)
# fake batch dimension required to fit network's input dimensions
image = loader(image).unsqueeze(0)
return image.to(pydiffvg.get_device(), torch.float)
style_img = image_loader(args.style_img)
# alpha blend content with a gray background
content_img = img[:, :3, :, :] * img[:, 3, :, :] + \
0.5 * torch.ones([1, 3, img.shape[2], img.shape[3]]) * \
(1 - img[:, 3, :, :])
assert style_img.size() == content_img.size(), \
"we need to import style and content images of the same size"
# unloader = transforms.ToPILImage() # reconvert into PIL image
class ContentLoss(nn.Module):
def __init__(
self,
target,
):
super(ContentLoss, self).__init__()
# we 'detach' the target content from the tree used
# to dynamically compute the gradient: this is a stated value,
# not a variable. Otherwise the forward method of the criterion
# will throw an error.
self.target = target.detach()
def forward(self, input):
self.loss = F.mse_loss(input, self.target)
return input
def gram_matrix(input):
a, b, c, d = input.size() # a=batch size(=1)
# b=number of feature maps
# (c,d)=dimensions of a f. map (N=c*d)
features = input.view(a * b, c * d) # resise F_XL into \hat F_XL
G = torch.mm(features, features.t()) # compute the gram product
# we 'normalize' the values of the gram matrix
# by dividing by the number of element in each feature maps.
return G.div(a * b * c * d)
class StyleLoss(nn.Module):
def __init__(self, target_feature):
super(StyleLoss, self).__init__()
self.target = gram_matrix(target_feature).detach()
def forward(self, input):
G = gram_matrix(input)
self.loss = F.mse_loss(G, self.target)
return input
device = pydiffvg.get_device()
cnn = models.vgg19(pretrained=True).features.to(device).eval()
cnn_normalization_mean = torch.tensor([0.485, 0.456, 0.406]).to(device)
cnn_normalization_std = torch.tensor([0.229, 0.224, 0.225]).to(device)
# create a module to normalize input image so we can easily put it in a
# nn.Sequential
class Normalization(nn.Module):
def __init__(self, mean, std):
super(Normalization, self).__init__()
# .view the mean and std to make them [C x 1 x 1] so that they can
# directly work with image Tensor of shape [B x C x H x W].
# B is batch size. C is number of channels. H is height and W is width.
self.mean = mean.clone().view(-1, 1, 1)
self.std = std.clone().view(-1, 1, 1)
def forward(self, img):
# normalize img
return (img - self.mean) / self.std
# desired depth layers to compute style/content losses :
content_layers_default = ['conv_4']
style_layers_default = ['conv_1', 'conv_2', 'conv_3', 'conv_4', 'conv_5']
def get_style_model_and_losses(cnn,
normalization_mean,
normalization_std,
style_img,
content_img,
content_layers=content_layers_default,
style_layers=style_layers_default):
cnn = copy.deepcopy(cnn)
# normalization module
normalization = Normalization(normalization_mean,
normalization_std).to(device)
# just in order to have an iterable access to or list of content/syle
# losses
content_losses = []
style_losses = []
# assuming that cnn is a nn.Sequential, so we make a new nn.Sequential
# to put in modules that are supposed to be activated sequentially
model = nn.Sequential(normalization)
i = 0 # increment every time we see a conv
for layer in cnn.children():
if isinstance(layer, nn.Conv2d):
i += 1
name = 'conv_{}'.format(i)
elif isinstance(layer, nn.ReLU):
name = 'relu_{}'.format(i)
# The in-place version doesn't play very nicely with the ContentLoss
# and StyleLoss we insert below. So we replace with out-of-place
# ones here.
layer = nn.ReLU(inplace=False)
elif isinstance(layer, nn.MaxPool2d):
name = 'pool_{}'.format(i)
elif isinstance(layer, nn.BatchNorm2d):
name = 'bn_{}'.format(i)
else:
raise RuntimeError('Unrecognized layer: {}'.format(
layer.__class__.__name__))
model.add_module(name, layer)
if name in content_layers:
# add content loss:
target = model(content_img).detach()
content_loss = ContentLoss(target)
model.add_module("content_loss_{}".format(i), content_loss)
content_losses.append(content_loss)
if name in style_layers:
# add style loss:
target_feature = model(style_img).detach()
style_loss = StyleLoss(target_feature)
model.add_module("style_loss_{}".format(i), style_loss)
style_losses.append(style_loss)
# now we trim off the layers after the last content and style losses
for i in range(len(model) - 1, -1, -1):
if isinstance(model[i], ContentLoss) or isinstance(
model[i], StyleLoss):
break
model = model[:(i + 1)]
return model, style_losses, content_losses
def run_style_transfer(cnn,
normalization_mean,
normalization_std,
content_img,
style_img,
canvas_width,
canvas_height,
shapes,
shape_groups,
num_steps=500,
style_weight=5000,
content_weight=1):
"""Run the style transfer."""
print('Building the style transfer model..')
model, style_losses, content_losses = get_style_model_and_losses(
cnn, normalization_mean, normalization_std, style_img, content_img)
point_params = []
color_params = []
stroke_width_params = []
for shape in shapes:
if isinstance(shape, pydiffvg.Path):
point_params.append(shape.points.requires_grad_())
stroke_width_params.append(shape.stroke_width.requires_grad_())
for shape_group in shape_groups:
if isinstance(shape_group.fill_color, torch.Tensor):
color_params.append(shape_group.fill_color.requires_grad_())
elif isinstance(shape_group.fill_color, pydiffvg.LinearGradient):
point_params.append(
shape_group.fill_color.begin.requires_grad_())
point_params.append(
shape_group.fill_color.end.requires_grad_())
color_params.append(
shape_group.fill_color.stop_colors.requires_grad_())
if isinstance(shape_group.stroke_color, torch.Tensor):
color_params.append(shape_group.stroke_color.requires_grad_())
elif isinstance(shape_group.stroke_color, pydiffvg.LinearGradient):
point_params.append(
shape_group.stroke_color.begin.requires_grad_())
point_params.append(
shape_group.stroke_color.end.requires_grad_())
color_params.append(
shape_group.stroke_color.stop_colors.requires_grad_())
point_optimizer = optim.Adam(point_params, lr=1.0)
color_optimizer = optim.Adam(color_params, lr=0.01)
stroke_width_optimizers = optim.Adam(stroke_width_params, lr=0.1)
print('Optimizing..')
run = [0]
while run[0] <= num_steps:
point_optimizer.zero_grad()
color_optimizer.zero_grad()
stroke_width_optimizers.zero_grad()
scene_args = pydiffvg.RenderFunction.serialize_scene(
canvas_width, canvas_height, shapes, shape_groups)
render = pydiffvg.RenderFunction.apply
img = render(
canvas_width, # width
canvas_height, # height
2, # num_samples_x
2, # num_samples_y
0, # seed
None,
*scene_args)
# alpha blend img with a gray background
img = img[:, :, :3] * img[:, :, 3:4] + \
0.5 * torch.ones([img.shape[0], img.shape[1], 3]) * \
(1 - img[:, :, 3:4])
pydiffvg.imwrite(img.cpu(),
'results/style_transfer/step_{}.png'.format(
run[0]),
gamma=1.0)
# HWC to NCHW
img = img.permute([2, 0, 1]).unsqueeze(0)
model(img)
style_score = 0
content_score = 0
for sl in style_losses:
style_score += sl.loss
for cl in content_losses:
content_score += cl.loss
style_score *= style_weight
content_score *= content_weight
loss = style_score + content_score
loss.backward()
run[0] += 1
if run[0] % 1 == 0:
print("run {}:".format(run))
print('Style Loss : {:4f} Content Loss: {:4f}'.format(
style_score.item(), content_score.item()))
print()
point_optimizer.step()
color_optimizer.step()
stroke_width_optimizers.step()
for color in color_params:
color.data.clamp_(0, 1)
for w in stroke_width_params:
w.data.clamp_(0.5, 4.0)
return shapes, shape_groups
shapes, shape_groups = run_style_transfer(cnn, cnn_normalization_mean,
cnn_normalization_std,
content_img, style_img,
canvas_width, canvas_height,
shapes, shape_groups)
scene_args = pydiffvg.RenderFunction.serialize_scene(shapes, shape_groups)
render = pydiffvg.RenderFunction.apply
img = render(
canvas_width, # width
canvas_height, # height
2, # num_samples_x
2, # num_samples_y
0, # seed
None,
*scene_args)
# Transform to gamma space
pydiffvg.imwrite(img.cpu(), 'results/style_transfer/output.png', gamma=1.0)
def train(args):
th.manual_seed(0)
np.random.seed(0)
pydiffvg.set_use_gpu(args.cuda)
# Initialize datasets
imsize = 28
dataset = Dataset(args.data_dir, imsize)
dataloader = DataLoader(dataset, batch_size=args.bs,
num_workers=4, shuffle=True)
if args.generator in ["vae", "ae"]:
LOG.info("Vector config:\n samples %d\n"
" paths: %d\n segments: %d\n"
" zdim: %d\n"
" conditional: %d\n"
" fc: %d\n",
args.samples, args.paths, args.segments,
args.zdim, args.conditional, args.fc)
model_params = dict(samples=args.samples, paths=args.paths,
segments=args.segments, conditional=args.conditional,
zdim=args.zdim, fc=args.fc)
if args.generator == "vae":
model = VectorMNISTVAE(variational=True, **model_params)
chkpt = VAE_OUTPUT
name = "mnist_vae"
elif args.generator == "ae":
model = VectorMNISTVAE(variational=False, **model_params)
chkpt = AE_OUTPUT
name = "mnist_ae"
else:
raise ValueError("unknown generator")
if args.conditional:
name += "_conditional"
chkpt += "_conditional"
if args.fc:
name += "_fc"
chkpt += "_fc"
# Resume from checkpoint, if any
checkpointer = ttools.Checkpointer(
chkpt, model, meta=model_params, prefix="g_")
extras, meta = checkpointer.load_latest()
if meta is not None and meta != model_params:
LOG.info("Checkpoint's metaparams differ from CLI, aborting: %s and %s",
meta, model_params)
# Hook interface
if args.generator in ["vae", "ae"]:
variational = args.generator == "vae"
if variational:
LOG.info("Using a VAE")
else:
LOG.info("Using an AE")
interface = VAEInterface(model, lr=args.lr, cuda=args.cuda,
variational=variational, w_kld=args.kld_weight)
trainer = ttools.Trainer(interface)
# Add callbacks
keys = ["loss_g", "loss_d"]
if args.generator == "vae":
keys = ["kld", "data_loss", "loss"]
elif args.generator == "ae":
keys = ["data_loss", "loss"]
port = 8097
trainer.add_callback(ttools.callbacks.ProgressBarCallback(
keys=keys, val_keys=keys))
trainer.add_callback(ttools.callbacks.VisdomLoggingCallback(
keys=keys, val_keys=keys, env=name, port=port))
trainer.add_callback(MNISTCallback(
env=name, win="samples", port=port, frequency=args.freq))
trainer.add_callback(ttools.callbacks.CheckpointingCallback(
checkpointer, max_files=2, interval=600, max_epochs=50))
# Start training
trainer.train(dataloader, num_epochs=args.num_epochs)
default=16,
type=int,
help="number of output to compute")
parser.add_argument("--imsize",
type=int,
help="if provided, override the raster output "
"resolution")
parser.add_argument("--nsteps",
default=9,
type=int,
help="number of "
"interpolation steps for the interpolation")
parser.add_argument("--nframes",
default=120,
type=int,
help="number of "
"frames for the interpolation video")
parser.add_argument("--invert",
default=False,
action="store_true",
help="if True, render black on white rather than the"
" opposite")
args = parser.parse_args()
pydiffvg.set_use_gpu(False)
ttools.set_logger(False)
run(args)
def main(args):
# Use GPU if available
pydiffvg.set_use_gpu(torch.cuda.is_available())
perception_loss = ttools.modules.LPIPS().to(pydiffvg.get_device())
#target = torch.from_numpy(skimage.io.imread('imgs/lena.png')).to(torch.float32) / 255.0
target = torch.from_numpy(skimage.io.imread(args.target)).to(
torch.float32) / 255.0
target = target.pow(gamma)
target = target.to(pydiffvg.get_device())
target = target.unsqueeze(0)
target = target.permute(0, 3, 1, 2) # NHWC -> NCHW
#target = torch.nn.functional.interpolate(target, size = [256, 256], mode = 'area')
canvas_width, canvas_height = target.shape[3], target.shape[2]
num_paths = args.num_paths
max_width = args.max_width
random.seed(1234)
torch.manual_seed(1234)
shapes = []
shape_groups = []
if args.use_blob:
for i in range(num_paths):
num_segments = random.randint(3, 5)
num_control_points = torch.zeros(num_segments,
dtype=torch.int32) + 2
points = []
p0 = (random.random(), random.random())
points.append(p0)
for j in range(num_segments):
radius = 0.05
p1 = (p0[0] + radius * (random.random() - 0.5),
p0[1] + radius * (random.random() - 0.5))
p2 = (p1[0] + radius * (random.random() - 0.5),
p1[1] + radius * (random.random() - 0.5))
p3 = (p2[0] + radius * (random.random() - 0.5),
p2[1] + radius * (random.random() - 0.5))
points.append(p1)
points.append(p2)
if j < num_segments - 1:
points.append(p3)
p0 = p3
points = torch.tensor(points)
points[:, 0] *= canvas_width
points[:, 1] *= canvas_height
path = pydiffvg.Path(num_control_points=num_control_points,
points=points,
stroke_width=torch.tensor(1.0),
is_closed=True)
shapes.append(path)
path_group = pydiffvg.ShapeGroup(shape_ids=torch.tensor(
[len(shapes) - 1]),
fill_color=torch.tensor([
random.random(),
random.random(),
random.random(),
random.random()
]))
shape_groups.append(path_group)
else:
for i in range(num_paths):
num_segments = random.randint(1, 3)
num_control_points = torch.zeros(num_segments,
dtype=torch.int32) + 2
points = []
p0 = (random.random(), random.random())
points.append(p0)
for j in range(num_segments):
radius = 0.05
p1 = (p0[0] + radius * (random.random() - 0.5),
p0[1] + radius * (random.random() - 0.5))
p2 = (p1[0] + radius * (random.random() - 0.5),
p1[1] + radius * (random.random() - 0.5))
p3 = (p2[0] + radius * (random.random() - 0.5),
p2[1] + radius * (random.random() - 0.5))
points.append(p1)
points.append(p2)
points.append(p3)
p0 = p3
points = torch.tensor(points)
points[:, 0] *= canvas_width
points[:, 1] *= canvas_height
#points = torch.rand(3 * num_segments + 1, 2) * min(canvas_width, canvas_height)
path = pydiffvg.Path(num_control_points=num_control_points,
points=points,
stroke_width=torch.tensor(1.0),
is_closed=False)
shapes.append(path)
path_group = pydiffvg.ShapeGroup(shape_ids=torch.tensor(
[len(shapes) - 1]),
fill_color=None,
stroke_color=torch.tensor([
random.random(),
random.random(),
random.random(),
random.random()
]))
shape_groups.append(path_group)
scene_args = pydiffvg.RenderFunction.serialize_scene(\
canvas_width, canvas_height, shapes, shape_groups)
render = pydiffvg.RenderFunction.apply
img = render(
canvas_width, # width
canvas_height, # height
2, # num_samples_x
2, # num_samples_y
0, # seed
None,
*scene_args)
pydiffvg.imwrite(img.cpu(),
'results/painterly_rendering/init.png',
gamma=gamma)
points_vars = []
stroke_width_vars = []
color_vars = []
for path in shapes:
path.points.requires_grad = True
points_vars.append(path.points)
if not args.use_blob:
for path in shapes:
path.stroke_width.requires_grad = True
stroke_width_vars.append(path.stroke_width)
if args.use_blob:
for group in shape_groups:
group.fill_color.requires_grad = True
color_vars.append(group.fill_color)
else:
for group in shape_groups:
group.stroke_color.requires_grad = True
color_vars.append(group.stroke_color)
# Optimize
points_optim = torch.optim.Adam(points_vars, lr=1.0)
if len(stroke_width_vars) > 0:
width_optim = torch.optim.Adam(stroke_width_vars, lr=0.1)
color_optim = torch.optim.Adam(color_vars, lr=0.01)
# Adam iterations.
for t in range(args.num_iter):
print('iteration:', t)
points_optim.zero_grad()
if len(stroke_width_vars) > 0:
width_optim.zero_grad()
color_optim.zero_grad()
# Forward pass: render the image.
scene_args = pydiffvg.RenderFunction.serialize_scene(\
canvas_width, canvas_height, shapes, shape_groups)
img = render(
canvas_width, # width
canvas_height, # height
2, # num_samples_x
2, # num_samples_y
t, # seed
None,
*scene_args)
# Compose img with white background
img = img[:, :, 3:4] * img[:, :, :3] + torch.ones(
img.shape[0], img.shape[1], 3,
device=pydiffvg.get_device()) * (1 - img[:, :, 3:4])
# Save the intermediate render.
pydiffvg.imwrite(img.cpu(),
'results/painterly_rendering/iter_{}.png'.format(t),
gamma=gamma)
img = img[:, :, :3]
# Convert img from HWC to NCHW
img = img.unsqueeze(0)
img = img.permute(0, 3, 1, 2) # NHWC -> NCHW
if args.use_lpips_loss:
loss = perception_loss(
img, target) + (img.mean() - target.mean()).pow(2)
else:
loss = (img - target).pow(2).mean()
print('render loss:', loss.item())
# Backpropagate the gradients.
loss.backward()
# Take a gradient descent step.
points_optim.step()
if len(stroke_width_vars) > 0:
width_optim.step()
color_optim.step()
if len(stroke_width_vars) > 0:
for path in shapes:
path.stroke_width.data.clamp_(1.0, max_width)
if args.use_blob:
for group in shape_groups:
group.fill_color.data.clamp_(0.0, 1.0)
else:
for group in shape_groups:
group.stroke_color.data.clamp_(0.0, 1.0)
if t % 10 == 0 or t == args.num_iter - 1:
pydiffvg.save_svg(
'results/painterly_rendering/iter_{}.svg'.format(t),
canvas_width, canvas_height, shapes, shape_groups)
# Render the final result.
img = render(
target.shape[1], # width
target.shape[0], # height
2, # num_samples_x
2, # num_samples_y
0, # seed
None,
*scene_args)
# Save the intermediate render.
pydiffvg.imwrite(img.cpu(),
'results/painterly_rendering/final.png'.format(t),
gamma=gamma)
# Convert the intermediate renderings to a video.
from subprocess import call
call([
"ffmpeg", "-framerate", "24", "-i",
"results/painterly_rendering/iter_%d.png", "-vb", "20M",
"results/painterly_rendering/out.mp4"
])
def gen_and_optimize(self, writer=None, color_optimisation_activated=False):
# Thanks to Katherine Crowson for this.
# In the CLIPDraw code used to generate examples, we don't normalize images
# before passing into CLIP, but really you should. Turn this to True to do that.
use_normalized_clip = True
pydiffvg.set_print_timing(False)
gamma = 1.0
# Use GPU if available
pydiffvg.set_use_gpu(torch.cuda.is_available())
pydiffvg.set_device(device)
max_width = 50
shapes, shape_groups = self.generator_func() # self.setup_parameters(colors)
# Just some diffvg setup
scene_args = pydiffvg.RenderFunction.serialize_scene(
self.canvas_width, self.canvas_height, shapes, shape_groups)
render = pydiffvg.RenderFunction.apply
img = render(self.canvas_width, self.canvas_height, 2, 2, 0, None, *scene_args)
background_image = torch.ones(img.shape)
points_vars = []
for path in shapes:
path.points.requires_grad = True
points_vars.append(path.points)
color_vars = list()
for group in shape_groups:
group.stroke_color.requires_grad = True
color_vars.append(group.stroke_color)
stroke_vars = list()
for path in shapes:
path.stroke_width.requires_grad = True
stroke_vars.append(path.stroke_width)
# Optimizers
points_optim = torch.optim.Adam(points_vars, lr=1.0)
color_optim = torch.optim.Adam(color_vars, lr=0.1)
stroke_optim = torch.optim.Adam(stroke_vars, lr=0.01)
# Run the main optimization loop
#all_groups = sum([g.param_groups for g in [points_optim, color_optim, stroke_optim]], [])
for t in range(self.num_iter):
# Anneal learning rate (makes videos look cleaner)
if t == int(self.num_iter * 0.5):
print(f"Iter {t}")
for g in points_optim.param_groups:
g['lr'] *= 0.5
if t == int(self.num_iter * 0.75):
print(f"Iter {t}")
for g in points_optim.param_groups:
g['lr'] *= 0.5
points_optim.zero_grad()
if color_optimisation_activated:
color_optim.zero_grad()
stroke_optim.zero_grad()
img = self.gen_image_from_curves(t, shapes, shape_groups, gamma, background_image)
im_batch = self.data_augment(img, self.n_augms, use_normalized_clip)
loss = self.forward_model_func(im_batch)
# Back-propagate the gradients.
loss.backward()
# Take a gradient descent step.
points_optim.step()
if color_optimisation_activated:
color_optim.step()
stroke_optim.step()
for path in shapes:
path.stroke_width.data.clamp_(1.0, max_width)
for group in shape_groups:
group.stroke_color.data.clamp_(0.0, 1.0)
if t % int(self.num_iter / 10) == 0 and writer is not None:
writer.add_scalars("neuron_excitation", {"loss": loss}, t)
writer.add_image('Rendering', img[0], t)
return shapes, shape_groups
def main(args):
if args.seed:
np.random.seed(args.seed)
random.seed(args.seed)
torch.manual_seed(args.seed)
pydiffvg.set_print_timing(False)
outdir = os.path.join(args.results_dir, args.prompt, args.subdir)
# Use GPU if available
pydiffvg.set_use_gpu(torch.cuda.is_available())
canvas_width, canvas_height = 224, 224
margin = args.initial_margin
total_paths = args.open_paths + args.closed_paths
step = min(args.step, total_paths)
if step == 0:
step = total_paths
fill_color = None
stroke_color = None
shapes = []
shape_groups = []
losses = []
tt = 0
for num_paths in range(step, total_paths + 1, step):
for i in range(num_paths - step, num_paths):
num_segments = random.randint(1, args.extra_segments + 1)
p0 = (margin + random.random() * (1 - 2 * margin),
margin + random.random() * (1 - 2 * margin))
points = [p0]
is_closed = i >= args.open_paths
if is_closed:
num_segments += 2
for j in range(num_segments):
p1 = (p0[0] + radius * (random.random() - 0.5),
p0[1] + radius * (random.random() - 0.5))
p2 = (p1[0] + radius * (random.random() - 0.5),
p1[1] + radius * (random.random() - 0.5))
p3 = (p2[0] + radius * (random.random() - 0.5),
p2[1] + radius * (random.random() - 0.5))
points.append(p1)
points.append(p2)
if is_closed and j < num_segments - 1 or not is_closed:
points.append(p3)
p0 = p3
points = torch.tensor(points)
points[:, 0] *= canvas_width
points[:, 1] *= canvas_height
stroke_width = torch.tensor(1.0)
color = torch.tensor([
random.random(),
random.random(),
random.random(),
random.random()
])
num_control_points = torch.zeros(num_segments,
dtype=torch.int32) + 2
path = pydiffvg.Path(num_control_points=num_control_points,
points=points,
stroke_width=stroke_width,
is_closed=is_closed)
shapes.append(path)
path_group = pydiffvg.ShapeGroup(
shape_ids=torch.tensor([len(shapes) - 1]),
fill_color=color if is_closed else None,
stroke_color=None if is_closed else color)
shape_groups.append(path_group)
scene_args = pydiffvg.RenderFunction.serialize_scene(\
canvas_width, canvas_height, shapes, shape_groups)
render = pydiffvg.RenderFunction.apply
img = render(
canvas_width, # width
canvas_height, # height
2, # num_samples_x
2, # num_samples_y
0, # seed
None,
*scene_args)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
pydiffvg.imwrite(img.cpu(),
os.path.join(outdir, 'init.png'),
gamma=gamma)
points_vars = []
stroke_width_vars = []
color_vars = []
for path in shapes:
path.points.requires_grad = True
points_vars.append(path.points)
if not path.is_closed and args.max_width > 1:
path.stroke_width.requires_grad = True
stroke_width_vars.append(path.stroke_width)
for group in shape_groups:
if group.fill_color is not None:
group.fill_color.requires_grad = True
color_vars.append(group.fill_color)
else:
group.stroke_color.requires_grad = True
color_vars.append(group.stroke_color)
# Embed prompt
text_features = clip_utils.embed_text(args.prompt)
# Optimize
points_optim = torch.optim.Adam(points_vars, lr=args.points_lr)
if len(stroke_width_vars) > 0:
width_optim = torch.optim.Adam(stroke_width_vars, lr=args.width_lr)
color_optim = torch.optim.Adam(color_vars, lr=args.color_lr)
# Adam iterations.
final = False
this_step_iters = max(1, round(args.num_iter * step / total_paths))
if num_paths + step > total_paths:
final = True
this_step_iters += args.extra_iter
for t in range(this_step_iters):
points_optim.zero_grad()
if len(stroke_width_vars) > 0:
width_optim.zero_grad()
color_optim.zero_grad()
# Forward pass: render the image.
scene_args = pydiffvg.RenderFunction.serialize_scene(\
canvas_width, canvas_height, shapes, shape_groups)
img = render(
canvas_width, # width
canvas_height, # height
2, # num_samples_x
2, # num_samples_y
tt, # seed
None,
*scene_args)
# Save the intermediate render.
with warnings.catch_warnings():
warnings.simplefilter("ignore")
pydiffvg.imwrite(img.cpu(),
os.path.join(outdir,
'iter_{}.png'.format(tt)),
gamma=gamma)
image_features = clip_utils.embed_image(img)
loss = -torch.cosine_similarity(
text_features, image_features, dim=-1).mean()
# Backpropagate the gradients.
loss.backward()
losses.append(loss.item())
# Take a gradient descent step.
points_optim.step()
if len(stroke_width_vars) > 0:
width_optim.step()
color_optim.step()
for path in shapes:
path.points.data[:, 0].clamp_(0.0, canvas_width)
path.points.data[:, 1].clamp_(0.0, canvas_height)
if not path.is_closed:
path.stroke_width.data.clamp_(1.0, args.max_width)
for group in shape_groups:
if group.fill_color is not None:
group.fill_color.data[:3].clamp_(0.0, 1.0)
group.fill_color.data[3].clamp_(args.min_alpha, 1.0)
else:
group.stroke_color.data[:3].clamp_(0.0, 1.0)
group.stroke_color.data[3].clamp_(args.min_alpha, 1.0)
if tt % 10 == 0 or final and t == this_step_iters - 1:
print('%d loss=%.3f' % (tt, 1 + losses[-1]))
pydiffvg.save_svg(
os.path.join(outdir, 'iter_{}.svg'.format(tt)),
canvas_width, canvas_height, shapes, shape_groups)
clip_utils.plot_losses(losses, outdir)
tt += 1
# Render the final result.
img = render(
args.final_px, # width
args.final_px, # height
2, # num_samples_x
2, # num_samples_y
0, # seed
None,
*scene_args)
# Save the intermediate render
with warnings.catch_warnings():
warnings.simplefilter("ignore")
pydiffvg.imwrite(img.cpu(),
os.path.join(outdir, 'final.png'),
gamma=gamma)
# Convert the intermediate renderings to a video with a white background.
from subprocess import call
call([
"ffmpeg", "-framerate", "24", "-i",
os.path.join(outdir, "iter_%d.png"), "-vb", "20M", "-filter_complex",
"color=white,format=rgb24[c];[c][0]scale2ref[c][i];[c][i]overlay=format=auto:shortest=1,setsar=1",
"-c:v", "libx264", "-pix_fmt", "yuv420p", "-profile:v", "baseline",
"-movflags", "+faststart",
os.path.join(outdir, "out.mp4")
])
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--svg",
default=os.path.join("imgs", "seamcarving",
"hokusai.svg"))
parser.add_argument("--optim_steps", default=10, type=int)
parser.add_argument("--lr", default=1e-1, type=int)
args = parser.parse_args()
name = os.path.splitext(os.path.basename(args.svg))[0]
root = os.path.join("results", "seam_carving", name)
svg_root = os.path.join(root, "svg")
os.makedirs(root, exist_ok=True)
os.makedirs(os.path.join(root, "svg"), exist_ok=True)
pydiffvg.set_use_gpu(False)
# pydiffvg.set_device(th.device('cuda'))
# Load SVG
print("loading svg %s" % args.svg)
canvas_width, canvas_height, shapes, shape_groups = \
pydiffvg.svg_to_scene(args.svg)
print("done loading")
max_size = 512
scale_factor = max_size / max(canvas_width, canvas_height)
print("rescaling from %dx%d with scale %f" %
(canvas_width, canvas_height, scale_factor))
canvas_width = int(canvas_width * scale_factor)
canvas_height = int(canvas_height * scale_factor)
print("new shape %dx%d" % (canvas_width, canvas_height))
vector_rescale(shapes, scale_x=scale_factor, scale_y=scale_factor)
# Shrink image by 33 %
# num_seams_to_remove = 2
num_seams_to_remove = canvas_width // 3
new_canvas_width = canvas_width - num_seams_to_remove
scaling = new_canvas_width * 1.0 / canvas_width
# Naive scaling baseline
print("rendering naive rescaling...")
vector_rescale(shapes, scale_x=scaling)
resized = render(new_canvas_width, canvas_height, shapes, shape_groups)
pydiffvg.imwrite(resized.cpu(),
os.path.join(root, 'uniform_scaling.png'),
gamma=2.2)
pydiffvg.save_svg(os.path.join(svg_root, 'uniform_scaling.svg'),
canvas_width,
canvas_height,
shapes,
shape_groups,
use_gamma=False)
vector_rescale(shapes,
scale_x=1.0 / scaling) # bring back original coordinates
print("saved naiving scaling")
# Save initial state
print("rendering initial state...")
im = render(canvas_width, canvas_height, shapes, shape_groups)
pydiffvg.imwrite(im.cpu(), os.path.join(root, 'init.png'), gamma=2.2)
pydiffvg.save_svg(os.path.join(svg_root, 'init.svg'),
canvas_width,
canvas_height,
shapes,
shape_groups,
use_gamma=False)
print("saved initial state")
# Optimize
# color_optim = th.optim.Adam(color_vars, lr=0.01)
retargeted = im[..., :3].cpu().numpy()
previous_width = canvas_width
print("carving seams")
for seam_idx in range(num_seams_to_remove):
print('\nseam', seam_idx + 1, 'of', num_seams_to_remove)
# Remove a seam
retargeted = carve_seam(retargeted)
current_width = canvas_width - seam_idx - 1
scale_factor = current_width * 1.0 / previous_width
previous_width = current_width
padded = np.zeros((canvas_height, canvas_width, 4))
padded[:, :-seam_idx - 1, :3] = retargeted
padded[:, :-seam_idx - 1, -1] = 1.0 # alpha
padded = th.from_numpy(padded).to(im.device)
# Remap points to the smaller canvas and
# collect variables to optimize
points_vars = []
# width_vars = []
mini, maxi = canvas_width, 0
for path in shapes:
path.points.requires_grad = False
x = path.points[..., 0]
y = path.points[..., 1]
# rescale
x = x * scale_factor
# clip to canvas
path.points[..., 0] = th.clamp(x, 0, current_width)
path.points[..., 1] = th.clamp(y, 0, canvas_height)
path.points.requires_grad = True
points_vars.append(path.points)
path.stroke_width.requires_grad = True
# width_vars.append(path.stroke_width)
mini = min(mini, path.points.min().item())
maxi = max(maxi, path.points.max().item())
print("points", mini, maxi, "scale", scale_factor)
# recreate an optimizer so we don't carry over the previous update
# (momentum)?
geom_optim = th.optim.Adam(points_vars, lr=args.lr)
for step in range(args.optim_steps):
geom_optim.zero_grad()
img = render(canvas_width,
canvas_height,
shapes,
shape_groups,
samples=2)
pydiffvg.imwrite(img.cpu(),
os.path.join(
root,
"seam_%03d_iter_%02d.png" % (seam_idx, step)),
gamma=2.2)
# NO alpha
loss = (img - padded)[..., :3].pow(2).mean()
# loss = (img - padded).pow(2).mean()
print('render loss:', loss.item())
# Backpropagate the gradients.
loss.backward()
# Take a gradient descent step.
geom_optim.step()
pydiffvg.save_svg(os.path.join(svg_root, "seam%03d.svg" % seam_idx),
canvas_width - seam_idx,
canvas_height,
shapes,
shape_groups,
use_gamma=False)
for path in shapes:
mini = min(mini, path.points.min().item())
maxi = max(maxi, path.points.max().item())
print("points", mini, maxi)
img = render(canvas_width, canvas_height, shapes, shape_groups)
img = img[:, :-num_seams_to_remove]
pydiffvg.imwrite(img.cpu(), os.path.join(root, 'final.png'), gamma=2.2)
pydiffvg.imwrite(retargeted, os.path.join(root, 'ref.png'), gamma=2.2)
pydiffvg.save_svg(os.path.join(svg_root, 'final.svg'),
canvas_width - num_seams_to_remove + 1,
canvas_height,
shapes,
shape_groups,
use_gamma=False)
# Convert the intermediate renderings to a video.
from subprocess import call
call([
"ffmpeg", "-framerate", "24", "-i",
os.path.join(root, "seam_%03d_iter_00.png"), "-vb", "20M",
os.path.join(root, "out.mp4")
])
def main(args):
# set device -> use cpu now since I haven't solved the nvcc issue
pydiffvg.set_use_gpu(False)
# pydiffvg.set_device(torch.device('cuda:1'))
# use L2 for now
# perception_loss = ttools.modules.LPIPS().to(pydiffvg.get_device())
# generate a texture synthesized
target_img = texture_syn(args.target)
tar_h, tar_w = target_img.shape[1], target_img.shape[0]
canvas_width, canvas_height, shapes, shape_groups = \
pydiffvg.svg_to_scene(args.svg_path)
# svgpathtools for checking the bounding box
# paths, _, _ = svg2paths2(args.svg_path)
# print(len(paths))
# xmin, xmax, ymin, ymax = big_bounding_box(paths)
# print(xmin, xmax, ymin, ymax)
# input("check")
print('tar h : %d tar w : %d' % (tar_h, tar_w))
print('canvas h : %d canvas w : %d' % (canvas_height, canvas_width))
scale_ratio = tar_h / canvas_height
print("scale ratio : ", scale_ratio)
# input("check")
for path in shapes:
path.points[..., 0] = path.points[..., 0] * scale_ratio
path.points[..., 1] = path.points[..., 1] * scale_ratio
init_img = render(tar_w, tar_h, shapes, shape_groups)
pydiffvg.imwrite(init_img.cpu(),
'results/texture_synthesis/%d/init.png' % (args.case),
gamma=2.2)
# input("check")
random.seed(1234)
torch.manual_seed(1234)
points_vars = []
for path in shapes:
path.points.requires_grad = True
points_vars.append(path.points)
color_vars = []
for group in shape_groups:
group.fill_color.requires_grad = True
color_vars.append(group.fill_color)
# Optimize
points_optim = torch.optim.Adam(points_vars, lr=1.0)
color_optim = torch.optim.Adam(color_vars, lr=0.01)
target = torch.from_numpy(target_img).to(torch.float32) / 255.0
target = target.pow(2.2)
target = target.to(pydiffvg.get_device())
target = target.unsqueeze(0)
target = target.permute(0, 3, 1, 2) # NHWC -> NCHW
canvas_width, canvas_height = target.shape[3], target.shape[2]
# print('canvas h : %d canvas w : %d' % (canvas_height, canvas_width))
# input("check")
for t in range(args.max_iter):
print('iteration:', t)
points_optim.zero_grad()
color_optim.zero_grad()
cur_img = render(canvas_width, canvas_height, shapes, shape_groups)
pydiffvg.imwrite(cur_img.cpu(),
'results/texture_synthesis/%d/iter_%d.png' %
(args.case, t),
gamma=2.2)
cur_img = cur_img[:, :, :3]
cur_img = cur_img.unsqueeze(0)
cur_img = cur_img.permute(0, 3, 1, 2) # NHWC -> NCHW
# perceptual loss
# loss = perception_loss(cur_img, target)
# l2 loss
loss = (cur_img - target).pow(2).mean()
print('render loss:', loss.item())
loss.backward()
points_optim.step()
color_optim.step()
for group in shape_groups:
group.fill_color.data.clamp_(0.0, 1.0)
# write svg
if t % 10 == 0 or t == args.max_iter - 1:
pydiffvg.save_svg(
'results/texture_synthesis/%d/iter_%d.svg' % (args.case, t),
canvas_width, canvas_height, shapes, shape_groups)
# render final result
final_img = render(tar_h, tar_w, shapes, shape_groups)
pydiffvg.imwrite(final_img.cpu(),
'results/texture_synthesis/%d/final.png' % (args.case),
gamma=2.2)
from subprocess import call
call([
"ffmpeg", "-framerate", "24", "-i",
"results/texture_synthesis/%d/iter_%d.png" % (args.case), "-vb", "20M",
"results/texture_synthesis/%d/out.mp4" % (args.case)
])
# make gif
make_gif("results/texture_synthesis/%d" % (args.case),
"results/texture_synthesis/%d/out.gif" % (args.case),
frame_every_X_steps=1,
repeat_ending=3,
total_iter=args.max_iter)
