def __call__(self, epoch):
""" Saves the training state.
Args:
epoch (int): Epoch number.
"""
if self._model:
ms_path_new = \
hlp.jn(self._path_dir, 'model_params_ep{}.h5'.format(epoch))
self.save_model_params(self._model, ms_path_new)
if self._verbose:
print('Saved model params in {}'.format(ms_path_new))
if self._ms_path and self._ms_path != ms_path_new:
os.remove(self._ms_path)
self._ms_path = ms_path_new
if self._opt:
os_path_new = \
hlp.jn(self._path_dir, 'optim_params_ep{}.pkl'.format(epoch))
self.save_optim_params(self._opt, os_path_new)
if self._verbose:
print('Saved optimizer params in {}'.format(os_path_new))
if self._os_path:
os.remove(self._os_path)
self._os_path = os_path_new
def get_file_list(path_root, seqs, exts=None):
""" Gets the list of paths to files contained in dirs listed in `seqs`.
The paths are relative w.r.t. the `path_root`.
Args:
path_root (str): Path to root dir containing sequences (directories)
of data samples.
seqs (lsit of str): Names of sequences (dirs) to load.
exts (list of str): Supported file name extensions.
Returns:
list of str: List of paths to files.
"""
return [
hlp.jn(s, f) for s in seqs
for f in hlp.ls(hlp.jn(path_root, s), exts=exts)
]
def __getitem__(self, idx):
d = np.load(hlp.jn(self._path_root, self._files[idx]))[self._key].\
astype(self._dtype)
if self._transform is not None:
d = self._transform(d)
return d
def process_sequence(obj):
# Get list of paths in obj. category in AtlasNet.
pth_files_an = jn(pth_root_an, obj, 'ply')
files = ls(pth_files_an, exts='txt')
# Peocess files.
for f in files:
# Extract file name.
fn_base = f.split('.')[0]
# Load .obj mesh from ShapeNet.
pth_f_sn = jn(
pth_root_sn, obj, fn_base, 'models', 'model_normalized.obj')
assert os.path.exists(pth_f_sn)
verts, faces = load_obj(pth_f_sn)
# Load tf and apply.
pth_f_an = jn(pth_files_an, f)
T, s = load_tf(pth_f_an)
verts = (verts - T) / s
# Compute area.
area = mesh_area(verts, faces)
# Write area to the file.
with open(pth_f_an, 'r') as fobj:
txt = fobj.read()
assert len(txt.splitlines()) == 2
has_nl = txt.endswith('\n')
with open(pth_f_an, 'a') as fobj:
fobj.write('{}{:.6f}'.format(('\n', '')[has_nl], area))
with num_samples_done.get_lock():
num_samples_done.value += 1
with finished_seqs.get_lock():
finished_seqs.value += 1
default='./data')
parser.add_argument('--resume',
help='Resume training from the given path',
default=False)
args = parser.parse_args()
args.resume = False
# Load the config file, prepare paths.
conf = helpers.load_conf(args.conf)
# Model type, color mode.
model_type = 'atlasnet_orig'
# Prepare TB writers.
writer_tr = SummaryWriter(helpers.jn(args.output, 'tr'))
writer_va = SummaryWriter(helpers.jn(args.output, 'va'))
# Build a model.
model = AtlasNetReimplEncImg(
M=conf['M'],
code=conf['code'],
num_patches=conf['num_patches'],
normalize_cw=conf['normalize_cw'],
freeze_encoder=conf['enc_freeze'],
enc_load_weights=conf['enc_weights'],
dec_activ_fns=conf['dec_activ_fns'],
dec_use_tanh=conf['dec_use_tanh'],
dec_batch_norm=conf['dec_batch_norm'],
loss_scaled_isometry=conf['loss_scaled_isometry'],
loss_patch_areas=conf['loss_patch_area'], # zhantao
def __init__(self, rootimg, rootpc, path_category_file,
class_choice='chair', train=True, npoints=2500, balanced=False,
gen_view=False, SVR=False, idx=0, load_area=False):
super(ShapeNet, self).__init__()
self.balanced = balanced
self.train = train
self.rootimg = rootimg
self.rootpc = rootpc
self.npoints = npoints
self.datapath = []
self.catfile = path_category_file
self.cat = {}
self.meta = {}
self.SVR = SVR
self.gen_view = gen_view
self.idx = idx
self._load_area = load_area
self._zero = torch.tensor(0., dtype=torch.float32)
with open(self.catfile, 'r') as f:
for line in f:
ls = line.strip().split()
self.cat[ls[0]] = ls[1]
if class_choice is not None:
self.cat = {k: v for k, v in self.cat.items()
if k in class_choice}
print(self.cat)
empty = []
for item in self.cat:
dir_img = os.path.join(self.rootimg, self.cat[item])
fns_img = sorted(os.listdir(dir_img))
try:
dir_point = os.path.join(self.rootpc, self.cat[item], 'ply')
fns_pc = sorted(os.listdir(dir_point))
except:
fns_pc = []
fns = [val for val in fns_img if val + '.points.ply' in fns_pc]
print('category ', self.cat[item], 'files ' + str(len(fns)),
len(fns) / float(len(fns_img)), "%"),
if train:
fns = fns[:int(len(fns) * 0.8)]
else:
fns = fns[int(len(fns) * 0.8):]
if len(fns) != 0:
self.meta[item] = []
for fn in fns:
objpath = jn(
'/cvlabdata2/cvlab/datasets_jan/shapenet/'
'ShapeNetCore.v2', self.cat[item], fn,
'models/model_normalized.obj')
self.meta[item].append(
(os.path.join(dir_img, fn, 'rendering'),
os.path.join(dir_point, fn + '.points.ply'),
os.path.join(dir_point, fn + '.points.ply2.txt'),
item, objpath, fn))
else:
empty.append(item)
for item in empty:
del self.cat[item]
self.idx2cat = {}
self.size = {}
i = 0
for item in self.cat:
self.idx2cat[i] = item
self.size[i] = len(self.meta[item])
i = i + 1
for fn in self.meta[item]:
self.datapath.append(fn)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
self.transforms = transforms.Compose([
transforms.Resize(size=224, interpolation=2),
transforms.ToTensor(),
# normalize,
])
# RandomResizedCrop or RandomCrop
self.dataAugmentation = transforms.Compose([
transforms.RandomCrop(127),
transforms.RandomHorizontalFlip(),
])
self.validating = transforms.Compose([
transforms.CenterCrop(127),
])
self.perCatValueMeter = {}
for item in self.cat:
self.perCatValueMeter[item] = AverageValueMeter()
self.perCatValueMeter_metro = {}
for item in self.cat:
self.perCatValueMeter_metro[item] = AverageValueMeter()
self.transformsb = transforms.Compose([
transforms.Resize(size=224, interpolation=2),
])
def __init__(self,
rootimg,
rootpc,
path_category_file,
class_choice='chair',
train=True,
test=False,
npoints=2500,
gen_view=False,
SVR=False,
idx=0,
load_area=False):
super(ShapeNet, self).__init__()
self.train = train
self.test = test
self.rootimg = rootimg
self.rootpc = rootpc
self.npoints = npoints
self.datapath = []
self.catfile = path_category_file
self.cat = {}
self.meta = {}
self.SVR = SVR
self.gen_view = gen_view
self.idx = idx
self._load_area = load_area
self._zero = torch.tensor(0., dtype=torch.float32)
# Find all category names in the dataset.
with open(self.catfile, 'r') as f:
for line in f:
ls = line.strip().split()
self.cat[ls[0]] = ls[1]
# Keep only selected object categories.
if class_choice is not None:
self.cat = {k: v for k, v in self.cat.items() if k in class_choice}
print(self.cat)
empty = []
for item in self.cat:
# Get image file names.
dir_img = os.path.join(self.rootimg, self.cat[item])
fns_img = sorted(os.listdir(dir_img))
# Get pts file names.
try:
dir_point = os.path.join(self.rootpc, self.cat[item], 'ply')
fns_pc = sorted(os.listdir(dir_point))
except:
fns_pc = []
# Select only those file names present both in imgs and pts dirs.
fns = [val for val in fns_img if val + '.points.ply' in fns_pc]
print('category ', self.cat[item], 'files ' + str(len(fns)),
len(fns) / float(len(fns_img)), "%"),
# Select train/valid/test subset.
if train:
# the first 80% data
print('construct training set')
fns = fns[:int(len(fns) * 0.8)]
elif test:
# test last 10% data
print('construct testing set')
fns = fns[int(len(fns) * 0.9):]
else:
# the 80% - 90% data
print('construct validation set')
fns = fns[int(len(fns) * 0.8):int(len(fns) * 0.9)]
if len(fns) != 0:
self.meta[item] = []
for fn in fns:
objpath = jn(
'/cvlabdata2/cvlab/datasets_jan/shapenet/'
'ShapeNetCore.v2', self.cat[item], fn,
'models/model_normalized.obj')
self.meta[item].append(
(os.path.join(dir_img, fn, 'rendering'),
os.path.join(dir_point, fn + '.points.ply'),
os.path.join(dir_point, fn + '.points.ply2.txt'),
item, objpath, fn))
else:
empty.append(item)
for item in empty:
del self.cat[item]
self.idx2cat = {}
self.size = {}
i = 0
for item in self.cat:
self.idx2cat[i] = item
self.size[i] = len(self.meta[item])
i = i + 1
for fn in self.meta[item]:
self.datapath.append(fn)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
self.transforms = transforms.Compose([
transforms.Resize(size=224, interpolation=2),
transforms.ToTensor(),
# normalize,
])
# RandomResizedCrop or RandomCrop
self.dataAugmentation = transforms.Compose([
transforms.RandomCrop(127),
transforms.RandomHorizontalFlip(),
])
self.validating = transforms.Compose([
transforms.CenterCrop(127),
])
def __getitem__(self, idx):
return hlp.load_img(hlp.jn(self._path_root_imgs, self._files[idx]))
gpu = torch.cuda.is_available()
# Parse arguments.
parser = argparse.ArgumentParser()
parser.add_argument('--conf', help='Path to the main config file of the model.',
default='config.yaml')
parser.add_argument('--output', help='Path to the output directory for storing '
'weights and tensorboard data.',
default='config.yaml')
args = parser.parse_args()
# Load the config file, prepare paths.
conf = helpers.load_conf(args.conf)
# Prepare TB writers.
writer_tr = SummaryWriter(helpers.jn(args.output, 'tr'))
writer_va = SummaryWriter(helpers.jn(args.output, 'va'))
# Build a model.
model = AtlasNetReimpl(
M=conf['M'], code=conf['code'], num_patches=conf['num_patches'],
normalize_cw=conf['normalize_cw'],
freeze_encoder=conf['enc_freeze'],
enc_load_weights=conf['enc_weights'],
dec_activ_fns=conf['dec_activ_fns'],
dec_use_tanh=conf['dec_use_tanh'],
dec_batch_norm=conf['dec_batch_norm'],
loss_scaled_isometry=conf['loss_scaled_isometry'],
alpha_scaled_isometry=conf['alpha_scaled_isometry'],
alphas_sciso=conf['alphas_sciso'], gpu=gpu)
help='Path to the main config file of the model.',
default='conf_texless_defsurf.yaml')
parser.add_argument('--output',
help='Path to the output directory for storing '
'weights and tensorboard data.',
default='./data')
parser.add_argument('--resume',
help='Resume training from the given path',
default=False)
args = parser.parse_args()
# Load the config file, prepare paths.
conf = helpers.load_conf(args.conf)
# Prepare TB writers.
writer_tr = SummaryWriter(helpers.jn(args.output, 'tr'))
writer_va = SummaryWriter(helpers.jn(args.output, 'va'))
# Build a model.
model = AtlasNetReimplEncImg(
M=conf['M'],
code=conf['code'],
num_patches=conf['num_patches'],
normalize_cw=conf['normalize_cw'],
freeze_encoder=conf['enc_freeze'],
enc_load_weights=conf['enc_weights'],
dec_activ_fns=conf['dec_activ_fns'],
dec_use_tanh=conf['dec_use_tanh'],
dec_batch_norm=conf['dec_batch_norm'],
loss_scaled_isometry=conf['loss_scaled_isometry'],
loss_patch_areas=conf['loss_patch_area'], # zhantao
def get_total_num_samples():
num_total = 0
seqs = lsd(pth_root_an)
for s in seqs:
num_total += len(ls(jn(pth_root_an, s), exts='txt'))
return num_total