def __init__(self, total_num_styles, network_parameter_path, device):
print("Setting up interpolator")
print(" Using network parameters from", network_parameter_path)
self.transfer_network = TransferNetwork(total_num_styles)
self.transfer_network.load_state_dict(
torch.load(network_parameter_path))
self.transfer_network.to(device).eval()
print(" Ready")
def __init__(self, model_path, save_dir, num_styles):
self.device = "cuda:0" if torch.cuda.is_available() else "cpu"
self.save_dir = save_dir
self.num_styles = num_styles
self.style_manager = StyleManager('../data/images/style/', self.device)
self.transfer_network = TransferNetwork(num_styles).to(self.device)
if self.device == "cpu":
self.transfer_network.load_state_dict(
torch.load(model_path, map_location='cpu'))
else:
self.transfer_network.load_state_dict(torch.load(model_path))
self.transfer_network.eval()
def main():
key = Cache.generate_key(str(Config()))
if Cache.check(key):
data = Cache.get(key)
points = data['points']
network = data['network']
else:
pass
# get points from trajectories
preprocessor = Preprocessor(Config.DATASET_ROOT_DIR,
Config.DATASET_SCALE)
points = preprocessor.get_points()
# use coherence expanded algorithm to form clusters
clusters = Cluster(points).coherence_expanding()
network = TransferNetwork(points, clusters)
# derive transfer probability
tp = TransferProbability(network)
tp.derive()
# save points and network to cache
Cache.save(key, {"points": points, "network": network})
# show the distribution of transfer probability
figure = Figure(width=8)
figure.transfer_probability(network, 8).show()
# search the most popular route
mpr = MostPopularRoute(network)
route = mpr.search(0, 4)
print(route)
figure = Figure()
figure.most_popular_route(points, network, route).show()
def process():
# get points from trajectories
preprocessor = Preprocessor(
Config.DATASET_ROOT_DIR,
Config.DATASET_SCALE)
points = preprocessor.get_points()
# use coherence expanded algorithm to form clusters
cluster = Cluster(points)
clusters = cluster.coherence_expanding()
network = TransferNetwork(points, clusters)
coherences = cluster.coherences
coherences.sort()
cluster_points = [point for cluster in clusters for point in cluster]
return points, clusters, network, coherences, cluster_points
class Interpolator(ABC):
def __init__(self, total_num_styles, network_parameter_path, device):
print("Setting up interpolator")
print(" Using network parameters from", network_parameter_path)
self.transfer_network = TransferNetwork(total_num_styles)
self.transfer_network.load_state_dict(
torch.load(network_parameter_path))
self.transfer_network.to(device).eval()
print(" Ready")
def render_interpolated_image(self,
interpolated_style_parameters,
test_image_tensor,
style_idx=0):
self.transfer_network.set_style_parameters(
interpolated_style_parameters, style_idx)
return self.transfer_network(test_image_tensor, style_idx)
class Renderer:
def __init__(self, model_path, save_dir, num_styles):
self.device = "cuda:0" if torch.cuda.is_available() else "cpu"
self.save_dir = save_dir
self.num_styles = num_styles
self.style_manager = StyleManager('../data/images/style/', self.device)
self.transfer_network = TransferNetwork(num_styles).to(self.device)
if self.device == "cpu":
self.transfer_network.load_state_dict(
torch.load(model_path, map_location='cpu'))
else:
self.transfer_network.load_state_dict(torch.load(model_path))
self.transfer_network.eval()
def render(self, content_image, style_idx):
content_image = content_image.to(self.device)
output = self.transfer_network(content_image, style_idx=style_idx)
output = output.detach().cpu()
return output[0]
def render_grid(self, content_dir, style_idx, style_id):
rendered_images = [self.style_manager[style_id]]
save_path = os.path.join(self.save_dir,
'style_grid_%d.png' % style_idx)
for content_path in os.listdir(content_dir):
content_image = load_image_as_tensor(
os.path.join(content_dir, content_path)).unsqueeze(0)
rendered_images.append(self.render(content_image, style_idx))
save_tensors_as_grid(rendered_images, save_path, nrow=4)
def render_all(self, content_image):
for style_idx in range(self.num_styles):
self.render_single(content_image, style_idx)
def render_single(self, content_image, style_idx):
save_path = os.path.join(self.save_dir, 'style_%d.png' % style_idx)
output = self.render(content_image, style_idx)
save_tensor_as_image(output, save_path)
sys.path.append(src_path)
from config import Config
from preprocessor import Preprocessor
from points import Points
from cluster import Cluster
from transfer_network import TransferNetwork
from transfer_probability import TransferProbability
from most_popular_route import MostPopularRoute
from figure import Figure
# get points from trajectories
preprocessor = Preprocessor(
Config.DATASET_ROOT_DIR,
Config.DATASET_SCALE)
points = preprocessor.get_points()
# use coherence expanded algorithm to form clusters
clusters = Cluster(points).coherence_expanding()
network = TransferNetwork(points, clusters)
# derive transfer probability
tp = TransferProbability(network)
tp.derive()
# search the most popular route
mpr = MostPopularRoute(network)
route = mpr.search(0, 6)
print(route)
figure = Figure()
figure.most_popular_route(points, network, route).show()