def __init__(self,
root_path,
h5_folder,
split,
transform3d=None,
range01=False,
pos_int16=False,
random_crop=False,
factor_rand=False,
factor=1):
self.root_path = root_path
self.h5_path = os.path.join(self.root_path, h5_folder)
self.split = utils.read_string_list(os.path.join(
self.root_path, split))
self.h5_folder = h5_folder
self.transform3d = transform3d
self.range01 = range01
self.pos_int16 = pos_int16
self.random_crop = random_crop
self.factor_rand = factor_rand
self.factor = factor
def extract_features(model, loader, split_path, h5_path, nLayer, cuda=True):
model.eval()
model.obtain_intermediate(nLayer)
features = SaveFeatures(list(model.children())[nLayer])
print("This layer will be used to extract the features: ",
list(model.children())[nLayer])
loader = tqdm(loader, ncols=100)
files = utils.read_string_list(split_path)
for i, batch in enumerate(loader, start=0):
if cuda:
batch = batch.to('cuda:0')
graph = model(batch)
feat = features.features
if type(feat) == torch_geometric.data.batch.Batch:
feat = feat.x
if type(graph) == torch_geometric.data.batch.Batch:
if feat.size(0) == graph.x.size(0) and graph.x.size(
0) == graph.pos.size(0) and graph.pos.size(0) == len(
graph.batch):
graph.x = feat
x, _ = torch_geometric.utils.to_dense_batch(graph.x,
batch=graph.batch)
pos, _ = torch_geometric.utils.to_dense_batch(
graph.pos, batch=graph.batch)
labels = graph.y
file_indexes = graph.c
for i in range(0, x.size(0)):
x_i = x[i, :, :]
pos_i = pos[i, :, :]
y_i = labels[i]
h5_file_name = os.path.join(h5_path,
files[file_indexes[i]] + '.h5')
save_h5_features(h5_file_name,
x_i.detach().cpu().numpy(),
pos_i.detach().cpu().numpy(),
y_i.detach().cpu().numpy())
else:
print('Dimensions doesn\'t match')
exit()
else:
raise RuntimeError('wrong input data')
features.close()
def __init__(self,
root_path,
h5_folder,
split,
transform=None,
range01=False):
self.root_path = root_path
self.h5_path = os.path.join(self.root_path, h5_folder)
self.split = utils.read_string_list(os.path.join(
self.root_path, split))
self.h5_folder = h5_folder
self.transform = transform
self.range01 = range01
def __init__(self,
root_path,
h5_folder_b1,
h5_folder_b2,
split,
transform3d=None,
pos_int16=False):
self.root_path = root_path
self.h5_path_b1 = os.path.join(self.root_path, h5_folder_b1)
self.h5_path_b2 = os.path.join(self.root_path, h5_folder_b2)
self.split = utils.read_string_list(os.path.join(
self.root_path, split))
self.h5_folder_b1 = h5_folder_b1
self.h5_folder_b2 = h5_folder_b2
self.transform3d = transform3d
self.pos_int16 = pos_int16
return img[start[0]:end[0], start[1]:end[1]]
if __name__ == "__main__":
# dataset folder path
dataset_path = '../../nyuv2'
# path to the data
dataset_mat_path = dataset_path + '/nyu_depth_v2_labeled.mat'
# folder to store the h5 files
path_h5 = dataset_path + '/h5/h5_feat2d/'
utils.create_folder(path_h5)
train_split = utils.read_string_list(dataset_path + "/list/train_list.txt")
val_split = utils.read_string_list(dataset_path + "/list/test_list.txt")
scenes_types = utils.read_string_list(dataset_path +
"/list/scenes_labels27.txt")
dataset = train_split + val_split
print("Loading .mat")
f = h5py.File(dataset_mat_path, 'r')
depths = np.round(np.transpose(np.asarray(f['depths'])), 4)
scenes = np.transpose(np.asarray(f['sceneTypes'])).squeeze()
mapping_10 = np.asarray([
10, 4, 1, 9, 10, 8, 10, 10, 10, 5, 10, 10, 10, 7, 10, 10, 2, 10, 3, 6,
10, 10, 10, 10, 10, 10, 10
]) - 1
args.pretrain_path)
model.load_state_dict(model_state)
else:
print('Wrong pretrain path')
exit()
if args.cuda is True and args.multigpu is False:
model = model.to('cuda:0')
print(model)
label_path = os.path.join(args.dataset_path, args.classname)
if not os.path.isfile(label_path):
raise RuntimeError("label file does not exist")
label_names = utils.read_string_list(label_path)
assert args.batch_size % args.batch_parts == 0
dataset = PCH5Dataset(args.dataset_path,
args.dataset_folder,
args.dataset,
range01=args.range01,
pos_int16=args.pos_int16)
loader = torch_geometric.data.DataLoader(dataset,
batch_size=int(args.batch_size /
args.batch_parts),
num_workers=args.nworkers,
shuffle=False,
pin_memory=False)
parameters = model.parameters()
if args.optim == "adam":
optimizer = torch.optim.Adam(parameters, lr=args.lr, betas=args.betas, weight_decay=args.wd)
elif args.optim == "sgd":
optimizer = torch.optim.SGD(parameters, lr=args.lr, momentum=args.momentum, weight_decay=args.wd)
elif args.optim == 'radam':
print('radam')
optimizer = RAdam(parameters, lr=args.lr, betas=args.betas, weight_decay=args.wd)
weights = None
if args.weights != '' and args.weights != '-':
weights_path = os.path.join(args.dataset_path, args.weights)
if not os.path.isfile(weights_path):
raise RuntimeError("weights file does not exist")
weights = torch.FloatTensor([float(i) for i in utils.read_string_list(weights_path)]).cuda(0)
loss_criterion = torch.nn.CrossEntropyLoss(weight=weights)
label_path = os.path.join(args.dataset_path, args.classname)
if not os.path.isfile(label_path):
raise RuntimeError("label file does not exist")
label_names = utils.read_string_list(label_path)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.wd)
elif args.optim == 'radam':
optimizer = RAdam(parameters,
lr=args.lr,
betas=args.betas,
weight_decay=args.wd)
weights = None
if args.weights != '' and args.weights != '-':
weights_path = os.path.join(args.dataset_path, args.weights)
if not os.path.isfile(weights_path):
raise RuntimeError("weights file does not exist")
weights = torch.FloatTensor([
float(i) for i in utils.read_string_list(weights_path)
]).cuda(args.lastgpu)
loss_criterion = torch.nn.CrossEntropyLoss(weight=weights)
label_path = os.path.join(args.dataset_path, args.classname)
if not os.path.isfile(label_path):
raise RuntimeError("label file does not exist")
label_names = utils.read_string_list(label_path)
assert args.batch_size % args.batch_parts == 0
transform3d = {
"dropout": args.pc_augm_input_dropout,
"rot": args.pc_augm_rot,
"mirror": args.pc_augm_mirror_prob
}
if __name__ == "__main__":
# dataset folder path
dataset_path = '../../sunrgbd'
# path to the data
images_list = dataset_path + '/list/sun_list.txt'
label_list = dataset_path + '/list/scenes_labels.txt'
# folder to store the h5 files
path_h5 = dataset_path + '/h5/h5_feat2d/'
utils.create_folder(path_h5)
images = utils.read_string_list(images_list)
dataset_labels = utils.read_string_list(label_list)
newSize = (420, 560)
for i in tqdm(range(0, len(images)), ncols=100):
img_folder = dataset_path + "/" + images[i]
depth_path = glob.glob(img_folder + "/depth_bfx/*.png")[0]
depth_img = read_depth_sunrgbd(depth_path)
intrinsic = np.loadtxt(img_folder + '/intrinsics.txt')
readlabel = open(img_folder + "/scene.txt", "r")
label = readlabel.read()