utils.makedirs(args.out_dir)
utils.makedirs(args.checkpoints_dir)
writer = writer.Writer(args)
device = torch.device('cuda', args.device_idx)
torch.set_num_threads(args.n_threads)
# deterministic
torch.manual_seed(args.seed)
cudnn.benchmark = False
cudnn.deterministic = True
# load dataset
template_fp = osp.join('template', 'template.obj')
meshdata = MeshData(args.data_fp,
template_fp,
split=args.split,
test_exp=args.test_exp)
train_loader = DataLoader(meshdata.train_dataset,
batch_size=args.batch_size,
shuffle=True)
test_loader = DataLoader(meshdata.test_dataset, batch_size=args.batch_size)
# generate/load transform matrices
transform_fp = osp.join(args.data_fp, 'transform.pkl')
if not osp.exists(transform_fp):
print('Generating transform matrices...')
mesh = Mesh(filename=template_fp)
ds_factors = [4, 4, 4, 4]
_, A, D, U, F = mesh_sampling.generate_transform_matrices(mesh, ds_factors)
tmp = {'face': F, 'adj': A, 'down_transform': D, 'up_transform': U}
spiral_indices_list = [
utils.preprocess_spiral(tmp['face'][idx], seq_length[idx],
tmp['vertices'][idx], dilation[idx]).to(device)
for idx in range(len(tmp['face']) - 1)
]
down_transform_list = [
utils.to_sparse(down_transform).to(device)
for down_transform in tmp['down_transform']
]
up_transform_list = [
utils.to_sparse(up_transform).to(device)
for up_transform in tmp['up_transform']
]
meshdata = MeshData("data/CoMA",
"data/CoMA/template/template.obj",
split="interpolation",
test_exp="bareteeth")
mean = meshdata.mean
std = meshdata.std
model = AE(3, [32, 32, 32, 64], 16, spiral_indices_list,
down_transform_list, up_transform_list).to(device)
checkpoint = torch.load(
"out/interpolation_exp/checkpoints/checkpoint_300.pt")
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()
print(len(meshdata.train_dataset))
train_loader = DataLoader(meshdata.train_dataset,
batch_size=1,
# TODO read model_checkpoint path from command line
if not torch.cuda.is_available():
model.load_state_dict(
torch.load(osp.join(args.checkpoints_dir, 'vae_checkpoint.pt'),
map_location=torch.device('cpu'))['model_state_dict'])
else:
model.load_state_dict(
torch.load(osp.join(args.checkpoints_dir,
'vae_checkpoint.pt'))['model_state_dict'])
model.eval()
# generate random measurement matrix
template_mesh = Mesh(filename=template_fp)
meshdata = MeshData(args.data_fp,
template_fp,
split=args.split,
test_exp=args.test_exp)
test_loader = DataLoader(meshdata.test_dataset, batch_size=args.batch_size)
# get a random test example
for i, data in enumerate(test_loader):
v = torch.squeeze(data.x[0, :, :])
if i == 42: #61
break
# sample simple measurement matrix, which just subselects a set of vertices,
# this enables to use the a regular norm for the gradient descent
A_diag = np.random.binomial(1, 0.005, template_mesh.v.shape[0])
num_measurements = np.count_nonzero(A_diag)
A = torch.from_numpy(np.diag(A_diag)).float()
]
del tmp['face']
del tmp['vertices']
down_transform_list = [
utils.to_sparse(down_transform).to(device)
for down_transform in tmp['down_transform']
]
del tmp['down_transform']
up_transform_list = [
utils.to_sparse(up_transform).to(device)
for up_transform in tmp['up_transform']
]
del tmp
meshdata = MeshData(args.data_fp,
template_fp,
split='interpolation',
test_exp='bareteeth')
test_loader = DataLoader(meshdata.test_dataset, batch_size=100, shuffle=True)
mean = meshdata.mean
std = meshdata.std
del meshdata
for i, (data, idx) in enumerate(test_loader):
batch = data
if i == 0:
break
del test_loader
#### AE ####
model_ae = AE(args.in_channels, args.out_channels, args.latent_channels,