def load_latent_vectors(experiment_directory, filename, lat_vecs):
full_filename = os.path.join(ws.get_latent_codes_dir(experiment_directory),
filename)
if not os.path.isfile(full_filename):
raise Exception(
'latent state file "{}" does not exist'.format(full_filename))
data = torch.load(full_filename)
if isinstance(data["latent_codes"], torch.Tensor):
# for backwards compatibility
if not lat_vecs.num_embeddings == data["latent_codes"].size()[0]:
raise Exception("num latent codes mismatched: {} vs {}".format(
lat_vecs.num_embeddings, data["latent_codes"].size()[0]))
if not lat_vecs.embedding_dim == data["latent_codes"].size()[2]:
raise Exception("latent code dimensionality mismatch")
for i, lat_vec in enumerate(data["latent_codes"]):
lat_vecs.weight.data[i, :] = lat_vec
else:
lat_vecs.load_state_dict(data["latent_codes"])
return data["epoch"]
def load_latent_vectors(experiment_directory, filename, lat_vecs):
full_filename = os.path.join(
ws.get_latent_codes_dir(experiment_directory), filename
)
if not os.path.isfile(full_filename):
raise Exception('latent state file "{}" does not exist'.format(full_filename))
data = torch.load(full_filename)
if not len(lat_vecs) == data["latent_codes"].size()[0]:
raise Exception(
"num latent codes mismatched: {} vs {}".format(
len(lat_vecs), data["latent_codes"].size()[2]
)
)
if not lat_vecs[0].size()[1] == data["latent_codes"].size()[2]:
raise Exception("latent code dimensionality mismatch")
for i in range(len(lat_vecs)):
lat_vecs[i] = data["latent_codes"][i].cuda()
return data["epoch"]
def save_latent_vectors(experiment_directory, filename, latent_vec, epoch):
latent_codes_dir = ws.get_latent_codes_dir(experiment_directory, True)
all_latents = latent_vec.state_dict()
torch.save(
{"epoch": epoch, "latent_codes": all_latents},
os.path.join(latent_codes_dir, filename),
)
def save_latent_vectors(experiment_directory, filename, latent_vec, epoch):
latent_codes_dir = ws.get_latent_codes_dir(experiment_directory, True)
all_latents = torch.zeros(0)
for l in latent_vec:
all_latents = torch.cat([all_latents, l.cpu().unsqueeze(0)], 0)
torch.save(
{"epoch": epoch, "latent_codes": all_latents},
os.path.join(latent_codes_dir, filename),
)
if not os.path.isdir(reconstruction_codes_dir):
os.makedirs(reconstruction_codes_dir)
latent_codes = []
data_root = "/".join(npz_filenames[0].split("/")[:-1])
for npz in args.sample_interp:
npz = data_root + "/" + npz + ".npz"
ii = npz_filenames.index(npz)
full_filename = os.path.join(args.data_source, ws.sdf_samples_subdir, npz)
logging.info("reconstructing {}".format(npz))
err, latent = None, None
latent = torch.load(os.path.join(ws.get_latent_codes_dir(args.experiment_directory), "latest.pth"))['latent_codes']['weight'].cuda()[ii, :]
logging.debug("latent: {}".format(latent.detach().cpu().numpy()))
latent_codes.append(latent)
num_interpol_steps = 12
for step in range(num_interpol_steps + 1):
mesh_filename = os.path.join(
reconstruction_meshes_dir, npz[:-4] + "-" + str(step)
)
latent_filename = os.path.join(
reconstruction_codes_dir, npz[:-4] + "-" + str(step) + ".pth"
)
decoder.eval()
npz[:-4] + ".pth")
if (args.skip and os.path.isfile(mesh_filename + ".ply")
and os.path.isfile(latent_filename)):
continue
logging.info("reconstructing {}".format(npz))
start = time.time()
use_saved_latent_vector = True
err, latent = None, None
if use_saved_latent_vector:
latent = torch.load(
os.path.join(
ws.get_latent_codes_dir(args.experiment_directory),
"latest.pth"))['latent_codes']['weight'].cuda()[ii, :]
else:
data_sdf = deep_sdf.data.read_sdf_samples_into_ram(
full_filename)
data_sdf[0] = data_sdf[0][torch.randperm(data_sdf[0].shape[0])]
data_sdf[1] = data_sdf[1][torch.randperm(data_sdf[1].shape[0])]
err, latent = reconstruct(
decoder,
int(args.iterations),
latent_size,
data_sdf,
0.01, # [emp_mean,emp_var],
0.1,
num_samples=8000,
