opt = parser.parse_args()
checkpoint_dir = os.path.join('../results/2D',opt.name)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
utils.save_opt(checkpoint_dir,opt)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print('loading dataset')
if opt.init is not None:
init_pose_np = np.load(opt.init)
init_pose = torch.from_numpy(init_pose_np)
else:
init_pose = None
dataset = SimulatedPointCloud(opt.data_dir,init_pose)
loader = DataLoader(dataset,batch_size=opt.batch_size,shuffle=False)
loss_fn = eval('loss.'+opt.loss)
print('creating model')
model = DeepMapping2D(loss_fn=loss_fn,n_obs=dataset.n_obs, n_samples=opt.n_samples).to(device)
optimizer = optim.Adam(model.parameters(),lr=opt.lr)
if opt.model is not None:
utils.load_checkpoint(opt.model,model,optimizer)
print('start training')
for epoch in range(opt.n_epochs):
training_loss= 0.0
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
utils.save_opt(checkpoint_dir, opt)
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print('loading dataset')
if opt.init is not None:
init_pose_np = np.load(opt.init)
init_pose = torch.from_numpy(init_pose_np)
else:
init_pose = None
instance_num = 256
latent_size = opt.lat_size
dataset = SimulatedPointCloud(opt.data_dir, instance_num, init_pose)
loader = DataLoader(dataset, batch_size=opt.batch_size, shuffle=False)
loss_fn = eval('loss.' + opt.loss)
print('creating model')
model = DeepMapping2D(n_lat=latent_size,
loss_fn=loss_fn,
n_obs=dataset.n_obs,
n_samples=opt.n_samples).to(device)
latent_vecs = []
for i in range(instance_num):
vec = nn.Parameter(torch.ones(latent_size).normal_(0, 0.9).to(device),
requires_grad=True)
latent_vecs.append(vec)
default='point',
choices=['point', 'plane'],
help='minimization metric')
parser.add_argument('-d',
'--data_dir',
type=str,
default='../data/2D/',
help='dataset path')
opt = parser.parse_args()
checkpoint_dir = os.path.join('../results/2D', opt.name)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
utils.save_opt(checkpoint_dir, opt)
dataset = SimulatedPointCloud(opt.data_dir)
n_pc = len(dataset)
pose_est = np.zeros((n_pc, 3), dtype=np.float32)
print('running icp')
for idx in range(n_pc - 1):
dst, valid_dst = dataset[idx]
src, valid_src = dataset[idx + 1]
dst = dst[valid_dst, :].numpy()
src = src[valid_src, :].numpy()
_, R0, t0 = utils.icp(src, dst, metrics=opt.metric)
if idx == 0:
R_cum = R0
t_cum = t0
init_pose = None
latent_vecs = []
latent_size = opt.latent_size
print("latent_size", latent_size)
instances_per_scene = 256
for i in range(instances_per_scene):
vec = (torch.ones(latent_size).normal_(0, 0.8).to(device))
vec.requires_grad = True
latent_vecs.append(vec)
w = torch.ones(latent_size).normal_(0, 0.8).to(device)
w_r = torch.ones(latent_size).normal_(0, 0.8).to(device)
dataset = SimulatedPointCloud(opt.data_dir, instances_per_scene, init_pose)
loader = DataLoader(dataset, batch_size=opt.batch_size, shuffle=False)
#print(len(dataset))
loss_fn = eval('loss.' + opt.loss)
print('creating model')
if opt.op == 'cat':
latent_size *= 2
model = DeepMapping2D(loss_fn=loss_fn,
n_obs=dataset.n_obs,
latent_size=latent_size,
n_samples=opt.n_samples).to(device)
optimizer = optim.Adam([{
"params": model.parameters(),
"lr": opt.lr,
}, {