def get(self,
idx=None,
model_args=None,
random_aug=True,
id=None,
svg: SVG = None):
if id is None:
idx = idx % len(self.df)
id = self.idx_to_id(idx)
# new
elif idx is None:
idx = np.where((self.df['id'] == id).to_numpy())[0][0]
# new
if svg is None:
svg = self._load_svg(id)
svg = SVGDataset.preprocess(svg, augment=random_aug)
t_sep, fillings = svg.to_tensor(
concat_groups=False, PAD_VAL=self.PAD_VAL), svg.to_fillings()
label = self.get_label(idx)
return self.get_data(t_sep,
fillings,
model_args=model_args,
label=label)
def visualize(self, model, output, train_vars, step, epoch, summary_writer,
visualization_dir):
device = next(model.parameters()).device
# Reconstruction
for i, data in enumerate(train_vars.x_inputs_train):
model_args = batchify((data[key] for key in self.model_args),
device)
commands_y, args_y = model.module.greedy_sample(*model_args)
tensor_pred = SVGTensor.from_cmd_args(commands_y[0].cpu(),
args_y[0].cpu())
try:
svg_path_sample = SVG.from_tensor(
tensor_pred.data, viewbox=Bbox(256),
allow_empty=True).normalize().split_paths().set_color(
"random")
except:
continue
tensor_target = data["tensor_grouped"][0].copy().drop_sos().unpad()
svg_path_gt = SVG.from_tensor(
tensor_target.data, viewbox=Bbox(
256)).normalize().split_paths().set_color("random")
img = make_grid([svg_path_sample,
svg_path_gt]).draw(do_display=False,
return_png=True,
fill=False,
with_points=False)
summary_writer.add_image(f"reconstructions_train/{i}",
TF.to_tensor(img), step)
def get(self,
idx=0,
model_args=None,
random_aug=True,
id=None,
svg: SVG = None):
item = self.data[idx]
if self.svg and self.svg_cmds:
tens_scene = []
tens_path = []
if self.add_scene_svg and len(item['path']) != 0:
tens_scene = self.simplify(SVG.from_tensor(
item['path'])).split_paths().to_tensor(concat_groups=False)
if self.add_history_agent_svg and len(item['history_agent']) != 0:
tens_path = self.normalize_history(
SVG.from_tensor(
item['history_agent'])).split_paths().to_tensor(
concat_groups=False)
if self.add_scene_svg and len(item['path_type']) != 0:
tens_scene = apply_colors(tens_scene, item['path_type'])
if self.add_history_agent_svg and len(
item['history_agent_type']) != 0:
tens_path = apply_colors(tens_path, item['history_agent_type'])
MAX_NUM_AGENTS = self.MAX_NUM_AGENTS
item['agents_availabilty'] = np.ones((item['agents_num']),
dtype=bool)
if item['agents_availabilty'].shape[0] < MAX_NUM_AGENTS:
item['agents_availabilty'] = np.concatenate(
(item['agents_availabilty'],
np.zeros((MAX_NUM_AGENTS - item['agents_num']),
dtype=bool)))
del item['path']
del item['path_type']
del item['history_agent']
del item['history_agent_type']
if self.add_history_agent_svg and self.add_scene_svg:
tens = tens_scene + tens_path
elif not self.add_history_agent_svg:
tens = tens_scene
elif not self.add_scene_svg:
tens = tens_path
item['image'] = self.get_data(idx,
tens,
None,
model_args=model_args,
label=None)
if item['image'] is None:
return None
return item
def __init__(self, index, keyframe=False, svg=None):
self.index = index
self.keyframe = keyframe
if svg is None:
svg = SVG([], viewbox=Bbox(256))
self.svg = svg
self.kivy_bezierpaths = None
def _load_svg(self, icon_id):
svg = SVG.load_svg(os.path.join(self.data_dir, f"{icon_id}.svg"))
if not self.already_preprocessed:
svg.fill_(False)
svg.normalize().zoom(0.9)
svg.canonicalize()
svg = svg.simplify_heuristic()
return svg
def get(self, idx=0, model_args=None, random_aug=True, id=None, svg: SVG=None):
if id is None:
idx = idx % len(self.df)
id = self.idx_to_id(idx)
if svg is None:
tensors, fillings = self._load_tensor(id)
t_sep = random.choice(tensors) if random_aug else tensors[0]
else:
t_sep, fillings = svg.to_tensor(concat_groups=False, PAD_VAL=self.PAD_VAL), svg.to_fillings()
label = self.get_label(idx)
return self.get_data(t_sep, fillings, model_args=model_args, label=label)
def preprocess_svg(svg_file, output_folder, meta_data):
filename = os.path.splitext(os.path.basename(svg_file))[0]
svg = SVG.load_svg(svg_file)
svg.fill_(False)
svg.normalize()
svg.zoom(4)
svg.canonicalize()
svg = svg.simplify_heuristic()
svg.save_svg(os.path.join(output_folder, f"{filename}.svg"))
len_groups = [path_group.total_len() for path_group in svg.svg_path_groups]
meta_data[filename] = {
"id": filename,
"total_len": sum(len_groups),
"nb_groups": len(len_groups),
"len_groups": len_groups,
"max_len_group": max(len_groups)
}
def decode(z):
commands_y, args_y = model.greedy_sample(z=z)
tensor_pred = SVGTensor.from_cmd_args(commands_y[0].cpu(), args_y[0].cpu())
svg_path_sample = SVG.from_tensor(tensor_pred.data, viewbox=Bbox(256))
return svg_path_sample
def normalized_path(svg_path):
svg = SVG([svg_path.copy().to_group()], viewbox=Bbox(256)).normalize()
return svg[0].path
def preprocess_svg_path(svg_path: SVGPath, force_smooth=False):
svg = SVG([svg_path.to_group()], viewbox=Bbox(256)).normalize()
svg.canonicalize()
svg.filter_duplicates()
svg = svg.simplify_heuristic(force_smooth=force_smooth)
svg.normalize()
svg.numericalize(256)
return svg[0].path
def load_frames(self, frames):
self.frames = [Frame.load_dict(frame) for frame in frames]
for frame in self.frames:
frame.svg = SVG.load_svg(
os.path.join(self.base_dir, f"{frame.index}.svg"))