def __init__(self, quality, mode, maxSkip=0, joint_transform_list=None,
transform=None, target_transform=None, dump_images=False,
class_uniform_pct=0, class_uniform_tile=0, test=False,
cv_split=None, scf=None, hardnm=0):
self.quality = quality
self.mode = mode
self.maxSkip = maxSkip
self.joint_transform_list = joint_transform_list
self.transform = transform
self.target_transform = target_transform
self.dump_images = dump_images
self.class_uniform_pct = class_uniform_pct
self.class_uniform_tile = class_uniform_tile
self.scf = scf
self.hardnm = hardnm
self.cv_split = cv_split
self.centroids = []
self.imgs, self.aug_imgs = make_dataset(quality, mode, self.maxSkip, cv_split=self.cv_split, hardnm=self.hardnm)
assert len(self.imgs), 'Found 0 images, please check the data set'
# Centroids for GT data
if self.class_uniform_pct > 0:
json_fn = 'camvid_tile{}_cv{}_{}.json'.format(self.class_uniform_tile, self.cv_split, self.mode)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.centroids = {int(idx): centroids[idx] for idx in centroids}
else:
self.centroids = uniform.class_centroids_all(
self.imgs,
num_classes,
id2trainid=None,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.centroids, outfile, indent=4)
self.fine_centroids = self.centroids.copy()
if self.maxSkip > 0:
json_fn = 'camvid_tile{}_cv{}_{}_skip{}.json'.format(self.class_uniform_tile, self.cv_split, self.mode, self.maxSkip)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.aug_centroids = {int(idx): centroids[idx] for idx in centroids}
else:
self.aug_centroids = uniform.class_centroids_all(
self.aug_imgs,
num_classes,
id2trainid=None,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.aug_centroids, outfile, indent=4)
for class_id in range(num_classes):
self.centroids[class_id].extend(self.aug_centroids[class_id])
self.build_epoch()
def __init__(self,
quality,
mode,
joint_transform_list=None,
transform=None,
target_transform=None,
target_aux_transform=None,
image_in=False,
dump_images=False,
class_uniform_pct=0,
class_uniform_tile=768,
test=False):
"""
class_uniform_pct = Percent of class uniform samples. 1.0 means fully uniform.
0.0 means fully random.
class_uniform_tile_size = Class uniform tile size
"""
gen_id_to_ignore()
self.quality = quality
self.mode = mode
self.joint_transform_list = joint_transform_list
self.transform = transform
self.target_transform = target_transform
self.image_in = image_in
self.target_aux_transform = target_aux_transform
self.dump_images = dump_images
self.class_uniform_pct = class_uniform_pct
self.class_uniform_tile = class_uniform_tile
self.id2name = gen_colormap()
self.imgs_uniform = None
# find all images
self.imgs = make_dataset(quality, mode)
if len(self.imgs) == 0:
raise RuntimeError('Found 0 images, please check the data set')
if test:
np.random.shuffle(self.imgs)
self.imgs = self.imgs[:200]
if self.class_uniform_pct:
json_fn = 'mapillary_tile{}.json'.format(self.class_uniform_tile)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.centroids = {
int(idx): centroids[idx]
for idx in centroids
}
else:
# centroids is a dict (indexed by class) of lists of centroids
self.centroids = uniform.class_centroids_all(
self.imgs,
num_classes,
id2trainid=None,
tile_size=self.class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.centroids, outfile, indent=4)
else:
self.centroids = []
self.build_epoch()
def __init__(self, quality, mode, maxSkip=0, joint_transform_list=None,
transform=None, target_transform=None, dump_images=False,
class_uniform_pct=0, class_uniform_tile=0, test=False,
cv_split=None, scf=None, hardnm=0):
self.quality = quality
self.mode = mode
self.maxSkip = maxSkip
self.joint_transform_list = joint_transform_list
self.transform = transform
self.target_transform = target_transform
self.dump_images = dump_images
self.class_uniform_pct = class_uniform_pct
self.class_uniform_tile = class_uniform_tile
self.scf = scf
self.hardnm = hardnm
if cv_split:
self.cv_split = cv_split
assert cv_split < cfg.DATASET.CV_SPLITS, \
'expected cv_split {} to be < CV_SPLITS {}'.format(
cv_split, cfg.DATASET.CV_SPLITS)
else:
self.cv_split = 0
if self.mode == 'test':
self.imgs, _ = make_test_dataset(quality, mode, self.maxSkip, cv_split=self.cv_split)
else:
self.imgs, _ = make_dataset(quality, mode, self.maxSkip, cv_split=self.cv_split, hardnm=self.hardnm)
assert len(self.imgs), 'Found 0 images, please check the data set'
# Centroids for GT data
#print(self.class_uniform_pct)
if self.class_uniform_pct > 0:
if self.scf:
json_fn = 'kitti_tile{}_cv{}_scf.json'.format(self.class_uniform_tile, self.cv_split)
else:
json_fn = 'kitti_tile{}_cv{}_{}_hardnm{}.json'.format(self.class_uniform_tile, self.cv_split, self.mode, self.hardnm)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.centroids = {int(idx): centroids[idx] for idx in centroids}
else:
if self.scf:
self.centroids = kitti_uniform.class_centroids_all(
self.imgs,
num_classes,
id2trainid=id_to_trainid,
tile_size=class_uniform_tile)
else:
print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
self.centroids = uniform.class_centroids_all(
self.imgs,
num_classes,
id2trainid=id_to_trainid,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.centroids, outfile, indent=4)
self.build_epoch()
def __init__(self,
mode,
maxSkip=0,
joint_transform_list=None,
sliding_crop=None,
transform=None,
target_transform=None,
target_aux_transform=None,
dump_images=False,
cv_split=None,
class_uniform_pct=0.5,
class_uniform_tile=1024,
test=False,
coarse_boost_classes=None):
self.mode = mode
self.maxSkip = maxSkip
self.joint_transform_list = joint_transform_list
self.sliding_crop = sliding_crop
self.transform = transform
self.target_transform = target_transform
self.target_aux_transform = target_aux_transform
self.dump_images = dump_images
self.class_uniform_pct = class_uniform_pct
self.class_uniform_tile = class_uniform_tile
self.coarse_boost_classes = coarse_boost_classes
if cv_split:
self.cv_split = cv_split
assert cv_split < cfg.DATASET.CV_SPLITS, \
'expected cv_split {} to be < CV_SPLITS {}'.format(
cv_split, cfg.DATASET.CV_SPLITS)
else:
self.cv_split = 0
self.imgs, self.aug_imgs = make_dataset(mode,
self.maxSkip,
cv_split=self.cv_split)
assert len(self.imgs), 'Found 0 images, please check the data set'
# Centroids for fine data
json_fn = 'bdd100k_{}_cv{}_tile{}.json'.format(self.mode,
self.cv_split,
self.class_uniform_tile)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.centroids = {int(idx): centroids[idx] for idx in centroids}
else:
self.centroids = uniform.class_centroids_all(
self.imgs,
num_classes,
id2trainid=trainid_to_trainid,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.centroids, outfile, indent=4)
self.fine_centroids = self.centroids.copy()
self.build_epoch()
def __init__(self,
quality,
mode,
maxSkip=0,
joint_transform_list=None,
sliding_crop=None,
transform=None,
target_transform=None,
dump_images=False,
cv_split=None,
class_uniform_pct=0.5,
class_uniform_tile=1024,
test=False,
coarse_boost_classes=None,
edge_map=False):
self.quality = quality
self.mode = mode
self.maxSkip = maxSkip
self.joint_transform_list = joint_transform_list
self.sliding_crop = sliding_crop
self.transform = transform
self.target_transform = target_transform
self.dump_images = dump_images
self.class_uniform_pct = class_uniform_pct
self.class_uniform_tile = class_uniform_tile
self.coarse_boost_classes = coarse_boost_classes
self.edge_map = edge_map
if cv_split:
self.cv_split = cv_split
assert cv_split < cfg.DATASET.CV_SPLITS, \
'expected cv_split {} to be < CV_SPLITS {}'.format(
cv_split, cfg.DATASET.CV_SPLITS)
else:
self.cv_split = 0
self.imgs, self.aug_imgs = make_dataset(quality,
mode,
self.maxSkip,
cv_split=self.cv_split)
assert len(self.imgs), 'Found 0 images, please check the data set'
# Centroids for fine data
json_fn = 'cityscapes_{}_cv{}_tile{}.json'.format(
self.mode, self.cv_split, self.class_uniform_tile)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.centroids = {int(idx): centroids[idx] for idx in centroids}
else:
self.centroids = uniform.class_centroids_all(
self.imgs,
num_classes,
id2trainid=id_to_trainid,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.centroids, outfile, indent=4)
self.fine_centroids = self.centroids.copy()
# Centroids for augmented data
if self.maxSkip > 0:
json_fn = 'cityscapes_{}_cv{}_tile{}_skip{}.json'.format(
self.mode, self.cv_split, self.class_uniform_tile,
self.maxSkip)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.aug_centroids = {
int(idx): centroids[idx]
for idx in centroids
}
else:
self.aug_centroids = uniform.class_centroids_all(
self.aug_imgs,
num_classes,
id2trainid=id_to_trainid,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.aug_centroids, outfile, indent=4)
# add centroids for augmented data
# TODO: later, we can also pick classes for augmented data
for class_id in range(num_classes):
self.centroids[class_id].extend(self.aug_centroids[class_id])
# Add in coarse centroids for certain classes
if self.coarse_boost_classes is not None:
json_fn = 'cityscapes_coarse_{}_tile{}.json'.format(
self.mode, self.class_uniform_tile)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.coarse_centroids = {
int(idx): centroids[idx]
for idx in centroids
}
else:
self.coarse_imgs, _ = make_dataset('coarse', mode, cv_split=0)
self.coarse_centroids = uniform.class_centroids_all(
self.coarse_imgs,
num_classes,
id2trainid=id_to_trainid,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.coarse_centroids, outfile, indent=4)
# add centroids for boost classes
for class_id in self.coarse_boost_classes:
self.centroids[class_id].extend(
self.coarse_centroids[class_id])
self.build_epoch()
def __init__(self,
mode,
maxSkip=0,
joint_transform_list=None,
sliding_crop=None,
transform=None,
target_transform=None,
target_aux_transform=None,
dump_images=False,
cv_split=None,
class_uniform_pct=0.5,
class_uniform_tile=1024,
test=False,
coarse_boost_classes=None,
pos_rfactor=8):
self.mode = mode
self.maxSkip = maxSkip
self.joint_transform_list = joint_transform_list
self.sliding_crop = sliding_crop
self.transform = transform
self.target_transform = target_transform
self.target_aux_transform = target_aux_transform
self.dump_images = dump_images
self.class_uniform_pct = class_uniform_pct
self.class_uniform_tile = class_uniform_tile
self.coarse_boost_classes = coarse_boost_classes
self.pos_rfactor = pos_rfactor
# position information
self.pos_h = torch.arange(0, 1024).unsqueeze(0).unsqueeze(2).expand(
-1, -1, 2048) // 8
self.pos_w = torch.arange(0, 2048).unsqueeze(0).unsqueeze(1).expand(
-1, 1024, -1) // 16
self.pos_h = self.pos_h[0].byte().numpy()
self.pos_w = self.pos_w[0].byte().numpy()
# pos index to image
self.pos_h = Image.fromarray(self.pos_h, mode="L")
self.pos_w = Image.fromarray(self.pos_w, mode="L")
# position information
if cv_split:
self.cv_split = cv_split
assert cv_split < cfg.DATASET.CV_SPLITS, \
'expected cv_split {} to be < CV_SPLITS {}'.format(
cv_split, cfg.DATASET.CV_SPLITS)
else:
self.cv_split = 0
self.imgs, self.aug_imgs = make_dataset(mode,
self.maxSkip,
cv_split=self.cv_split)
assert len(self.imgs), 'Found 0 images, please check the data set'
# Centroids for fine data
json_fn = 'bdd100k_{}_cv{}_tile{}.json'.format(self.mode,
self.cv_split,
self.class_uniform_tile)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.centroids = {int(idx): centroids[idx] for idx in centroids}
else:
self.centroids = uniform.class_centroids_all(
self.imgs,
num_classes,
id2trainid=trainid_to_trainid,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.centroids, outfile, indent=4)
self.fine_centroids = self.centroids.copy()
self.build_epoch()
def __init__(self, quality, mode, maxSkip=0, joint_transform_list=None,
transform=None, target_transform=None, target_aux_transform=None, dump_images=False,
class_uniform_pct=0, class_uniform_tile=0, test=False,
cv_split=None, scf=None, hardnm=0, pos_rfactor=8):
self.quality = quality
self.mode = mode
self.maxSkip = maxSkip
self.joint_transform_list = joint_transform_list
self.transform = transform
self.target_transform = target_transform
self.target_aux_transform = target_aux_transform
self.dump_images = dump_images
self.class_uniform_pct = class_uniform_pct
self.class_uniform_tile = class_uniform_tile
self.scf = scf
self.hardnm = hardnm
self.cv_split = cv_split
self.centroids = []
self.pos_rfactor = pos_rfactor
# position information
self.pos_h = torch.arange(0, 1024).unsqueeze(0).unsqueeze(2).expand(-1,-1,2048)//8
self.pos_w = torch.arange(0, 2048).unsqueeze(0).unsqueeze(1).expand(-1,1024,-1)//16
self.pos_h = self.pos_h[0].byte().numpy()
self.pos_w = self.pos_w[0].byte().numpy()
# pos index to image
self.pos_h = Image.fromarray(self.pos_h, mode="L")
self.pos_w = Image.fromarray(self.pos_w, mode="L")
# position information
self.imgs, self.aug_imgs = make_dataset(quality, mode, self.maxSkip, cv_split=self.cv_split, hardnm=self.hardnm)
assert len(self.imgs), 'Found 0 images, please check the data set'
# Centroids for GT data
if self.class_uniform_pct > 0:
json_fn = 'camvid_tile{}_cv{}_{}.json'.format(self.class_uniform_tile, self.cv_split, self.mode)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.centroids = {int(idx): centroids[idx] for idx in centroids}
else:
self.centroids = uniform.class_centroids_all(
self.imgs,
num_classes,
id2trainid=None,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.centroids, outfile, indent=4)
self.fine_centroids = copy.deepcopy(self.centroids)
if self.maxSkip > 0:
json_fn = 'camvid_tile{}_cv{}_{}_skip{}.json'.format(self.class_uniform_tile, self.cv_split, self.mode, self.maxSkip)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.aug_centroids = {int(idx): centroids[idx] for idx in centroids}
else:
self.aug_centroids = uniform.class_centroids_all(
self.aug_imgs,
num_classes,
id2trainid=None,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.aug_centroids, outfile, indent=4)
for class_id in range(num_classes):
self.centroids[class_id].extend(self.aug_centroids[class_id])
self.build_epoch()
def __init__(self,
quality,
mode,
maxSkip=0,
joint_transform_list=None,
sliding_crop=None,
transform=None,
target_transform=None,
target_aux_transform=None,
dump_images=False,
cv_split=None,
class_uniform_pct=0.5,
class_uniform_tile=1024,
test=False,
coarse_boost_classes=None,
pos_rfactor=8):
print("######## CityScapesUniformWithPos #########")
self.quality = quality
self.mode = mode
self.maxSkip = maxSkip
self.joint_transform_list = joint_transform_list
self.sliding_crop = sliding_crop
self.transform = transform
self.target_transform = target_transform
self.target_aux_transform = target_aux_transform
self.dump_images = dump_images
self.class_uniform_pct = class_uniform_pct
self.class_uniform_tile = class_uniform_tile
self.coarse_boost_classes = coarse_boost_classes
self.pos_rfactor = pos_rfactor
# position information
self.pos_h = torch.arange(0, 1024).unsqueeze(0).unsqueeze(2).expand(
-1, -1, 2048) // 8
self.pos_w = torch.arange(0, 2048).unsqueeze(0).unsqueeze(1).expand(
-1, 1024, -1) // 16
self.pos_h = self.pos_h[0].byte().numpy()
self.pos_w = self.pos_w[0].byte().numpy()
# pos index to image
self.pos_h = Image.fromarray(self.pos_h, mode="L")
self.pos_w = Image.fromarray(self.pos_w, mode="L")
# position information
if cv_split:
self.cv_split = cv_split
assert cv_split < cfg.DATASET.CV_SPLITS, \
'expected cv_split {} to be < CV_SPLITS {}'.format(
cv_split, cfg.DATASET.CV_SPLITS)
else:
self.cv_split = 0
self.imgs, self.aug_imgs = make_dataset(quality,
mode,
self.maxSkip,
cv_split=self.cv_split)
assert len(self.imgs), 'Found 0 images, please check the data set'
# Centroids for fine data
json_fn = 'cityscapes_{}_cv{}_tile{}.json'.format(
self.mode, self.cv_split, self.class_uniform_tile)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.centroids = {int(idx): centroids[idx] for idx in centroids}
else:
self.centroids = uniform.class_centroids_all(
self.imgs,
num_classes,
id2trainid=id_to_trainid,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.centroids, outfile, indent=4)
self.fine_centroids = copy.deepcopy(self.centroids)
# Centroids for augmented data
if self.maxSkip > 0:
json_fn = 'cityscapes_{}_cv{}_tile{}_skip{}.json'.format(
self.mode, self.cv_split, self.class_uniform_tile,
self.maxSkip)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.aug_centroids = {
int(idx): centroids[idx]
for idx in centroids
}
else:
self.aug_centroids = uniform.class_centroids_all(
self.aug_imgs,
num_classes,
id2trainid=id_to_trainid,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.aug_centroids, outfile, indent=4)
# add centroids for augmented data
# TODO: later, we can also pick classes for augmented data
for class_id in range(num_classes):
self.centroids[class_id].extend(self.aug_centroids[class_id])
# Add in coarse centroids for certain classes
if self.coarse_boost_classes is not None:
json_fn = 'cityscapes_coarse_{}_tile{}.json'.format(
self.mode, self.class_uniform_tile)
if os.path.isfile(json_fn):
with open(json_fn, 'r') as json_data:
centroids = json.load(json_data)
self.coarse_centroids = {
int(idx): centroids[idx]
for idx in centroids
}
else:
self.coarse_imgs, _ = make_dataset('coarse',
'train',
cv_split=0)
self.coarse_centroids = uniform.class_centroids_all(
self.coarse_imgs,
num_classes,
id2trainid=id_to_trainid,
tile_size=class_uniform_tile)
with open(json_fn, 'w') as outfile:
json.dump(self.coarse_centroids, outfile, indent=4)
# add centroids for boost classes
for class_id in self.coarse_boost_classes:
self.centroids[class_id].extend(
self.coarse_centroids[class_id])
self.build_epoch()
