indices[BATCH_SIZE // 2:] = torch.tensor(np.random.choice(
negative_indices, BATCH_SIZE // 2),
device=device)
sdf_net.zero_grad()
predicted_sdf = sdf_net(points[indices, :], latent_code)
batch_sdf = sdf[indices]
loss = torch.mean(torch.abs(predicted_sdf - batch_sdf))
loss.backward()
optimizer.step()
if loss.item() < error_targets[image_index]:
try:
viewer.set_mesh(
sdf_net.get_mesh(latent_code[0, :],
voxel_resolution=64,
raise_on_empty=True))
if save_images:
image = viewer.get_image(flip_red_blue=True)
cv2.imwrite("images/frame-{:05d}.png".format(image_index),
image)
image_index += 1
except ValueError:
pass
step += 1
print('Step {:04d}, Image {:04d}, loss: {:.6f}, target: {:.6f}'.format(
step, image_index, loss.item(), error_targets[image_index]))
except KeyboardInterrupt:
viewer.stop()
break
checkpoints_latent_codes[i])).detach()
latent_code = latent_codes[MODEL_INDEX, :]
plot.set_image(render_image(sdf_net, latent_code, crop=True), i)
plot.save('plots/deepsdf-checkpoints.pdf')
if "deepsdf-interpolation-stl" in sys.argv:
from rendering.raymarching import render_image_for_index, render_image
sdf_net, latent_codes = load_sdf_net(return_latent_codes=True)
STEPS = 5
indices = random.sample(list(range(latent_codes.shape[0])), 2)
print(indices)
code_start = latent_codes[indices[0], :]
code_end = latent_codes[indices[1], :]
print("Generating codes...")
with torch.no_grad():
codes = torch.zeros([STEPS, LATENT_CODE_SIZE], device=device)
for i in range(STEPS):
codes[
i, :] = code_start * (1.0 - i /
(STEPS - 1)) + code_end * i / (STEPS - 1)
for i in range(STEPS):
print(i)
mesh = sdf_net.get_mesh(codes[i, :],
voxel_resolution=256,
sphere_only=False)
mesh.export('plots/mesh-{:d}.stl'.format(i))