def animate(self,
mesh: MeshInference,
start: int,
end: int,
num_scene: int,
num_frames: Tuple[int, int],
zero_places: NoiseT = ()):
export_name = f'{self.opt.cp_folder}/inference/{mesh.mesh_name}/scene01'
start, end = self.trim(start, end)
if len(zero_places) == 1:
zero_places = zero_places * (end - start + 1)
z_a = self.growing(mesh.copy(), start, end, num_frames[0], zero_places)
z_b = self.get_z_sequence(mesh, end - start).to(self.device)
for i in range(min(len(z_a), len(zero_places))):
if zero_places[i]:
z_b[i] = 0
z_start = z_a
num_frames = num_frames[1]
for s in range(num_scene):
for i in range(num_frames):
alpha = (i + 1) / float(num_frames)
z = z_a * (1 - alpha) + z_b * alpha
m = mesh.copy()
out = self.generator(m, z, end, start, upsample=True)
out.export(f'{export_name}/{s * num_frames + i:02d}')
print(
f'frame {s * num_frames + i + 1} / {num_scene * num_frames}...'
)
z_a = z_b
if s == num_scene - 2:
z_b = z_start
else:
z_b = self.get_z_sequence(mesh, end - start).to(self.device)
def growing(self,
mesh: MeshInference,
start: int,
end: int,
num_frames: int,
zero_places: NoiseT = ()) -> factory.Noise:
export_name = f'{self.opt.cp_folder}/inference/{mesh.mesh_name}/scene00'
start, end = self.trim(start, end)
if len(zero_places) == 1:
zero_places = zero_places * (end - start + 1)
z = self.get_z_sequence(mesh, end - start)
for i in range(min(len(z), len(zero_places))):
if zero_places[i]:
z[i] = 0
deltas = self.generator.grow_forward(mesh.copy(), z, end, start)
mesh.export(f'{export_name}/{0:02d}')
for i in range(end - start + 1):
base_vs, cur_delta = mesh.vs.clone(), deltas[i]
for j in range(num_frames):
mesh.vs = base_vs + cur_delta * (j + 1) / num_frames
mesh.export(f'{export_name}/{(num_frames * i + j + 1):02d}')
print(
f'done: {num_frames * i + j + 1}/{num_frames * (end - start + 1)}'
)
if i < end - start:
mesh.upsample()
return z
class MeshGen(DGTS):
def __init__(self, opt: options.Options, device: D):
super(MeshGen, self).__init__(opt, device)
self.generator.eval()
template_name, template = load_template_mesh(opt, opt.start_level)
self.template = MeshInference(template_name, template, self.opt,
self.opt.start_level).to(self.device)
self.reconstruction_z = factory.NoiseMem(opt).load().to(device)
def compose_z(self, start_level) -> factory.Noise:
random_noise = self.get_z_sequence(self.template, len(self) - 1)
noise = self.reconstruction_z[:start_level] + random_noise[start_level:]
return noise
def generate_seq(self, num_seqs: int):
for seq in range(num_seqs):
z = self.compose_z(0)
self.generator.inference_forward(
self.template.copy(),
z,
len(self) - 1,
0,
f'{opt_.cp_folder}/inference/gen/{self.opt.mesh_name}_{seq}',
upsample=True)
def generate_all(self, num_samples: int):
for i in range(len(self.generator.levels)):
for j in range(num_samples):
out_mesh = self(i)
out_mesh.export(
f'{self.opt.cp_folder}/inference/gen/{self.opt.mesh_name}_{i}_{j:02d}'
)
print(
f'gen {self.opt.mesh_name} {i * num_samples + j +1:02d} / {len(self.generator.levels) * num_samples}'
)
def __call__(self, start_level: int):
with torch.no_grad():
if start_level < 0:
start_level = len(self)
start_level = min(len(self), start_level)
z = self.compose_z(start_level)
return self.generator(self.template.copy(), z, len(self) - 1)