def get_dataset_specs(args, model_specs):
dataset = model_specs['dataset']
meta = {}
meta_path = osp.join('issegm/data', dataset, 'meta.pkl')
if osp.isfile(meta_path):
with open(meta_path) as f:
meta = cPickle.load(f)
label_2_id = None
id_2_label = None
ident_size = False
cache_images = False
mx_workspace = 1650
if dataset == 'cityscapes':
sys.path.insert(0, 'data/cityscapesScripts/cityscapesscripts/helpers')
from labels import id2label, trainId2label
#
label_2_id = 255 * np.ones((256, ))
for l in id2label:
if l in (-1, 255):
continue
label_2_id[l] = id2label[l].trainId
id_2_label = np.array(
[trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)])
valid_labels = sorted(set(id_2_label.ravel()))
#
cmap = np.zeros((256, 3), dtype=np.uint8)
for i in id2label.keys():
cmap[i] = id2label[i].color
#
ident_size = True
#
max_shape = np.array((1024, 2048))
mx_workspace = 8000
meta['label_2_id'] = label_2_id
meta['id_2_label'] = id_2_label
meta['valid_labels'] = valid_labels
meta['cmap'] = cmap
meta['ident_size'] = ident_size
meta['max_shape'] = meta.get('max_shape', max_shape)
meta['cache_images'] = cache_images
meta['mx_workspace'] = mx_workspace
return meta
def get_dataset_specs(args, model_specs):
dataset = model_specs['dataset']
meta = {}
meta_path = osp.join('issegm/data', dataset, 'meta.pkl')
if osp.isfile(meta_path):
with open(meta_path) as f:
meta = cPickle.load(f)
label_2_id = None
id_2_label = None
ident_size = False
cmap = None
cmap_path = 'data/shared/cmap.pkl'
cache_images = args.phase == 'train'
mx_workspace = 1650
if dataset == 'cityscapes':
sys.path.insert(0, 'data/cityscapesscripts/helpers')
from labels import id2label, trainId2label
#
label_2_id = 255 * np.ones((256, ))
for l in id2label:
if l in (-1, 255):
continue
label_2_id[l] = id2label[l].trainId
id_2_label = np.array(
[trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)])
valid_labels = sorted(set(id_2_label.ravel()))
#
cmap = np.zeros((256, 3), dtype=np.uint8)
for i in id2label.keys():
cmap[i] = id2label[i].color
#
ident_size = True
#
max_shape = np.array((1024, 2048))
#
if args.split in ('train+', 'trainval+'):
cache_images = False
#
if args.phase in ('val', ):
mx_workspace = 8192
elif dataset == 'gta':
sys.path.insert(0, 'data/cityscapesscripts/helpers')
from labels import id2label, trainId2label
#
label_2_id = 255 * np.ones((256, ))
for l in id2label:
if l in (-1, 255):
continue
label_2_id[l] = id2label[l].trainId
id_2_label = np.array(
[trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)])
valid_labels = sorted(set(id_2_label.ravel()))
#
cmap = np.zeros((256, 3), dtype=np.uint8)
for i in id2label.keys():
cmap[i] = id2label[i].color
#
ident_size = True
#
max_shape = np.array((1052, 1914))
#
if args.phase in ('val', ):
mx_workspace = 8192
elif dataset == 'synthia':
sys.path.insert(0, 'data/cityscapesscripts/helpers')
from labels_synthia import id2label, trainId2label
#
label_2_id = 255 * np.ones((256, ))
for l in id2label:
if l in (-1, 255):
continue
label_2_id[l] = id2label[l].trainId
id_2_label = np.array(
[trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)])
valid_labels = sorted(set(id_2_label.ravel()))
#
cmap = np.zeros((256, 3), dtype=np.uint8)
for i in id2label.keys():
cmap[i] = id2label[i].color
#
ident_size = True
#
max_shape = np.array((760, 1280))
#
if args.phase in ('val', ):
mx_workspace = 8192
else:
raise NotImplementedError('Unknow dataset: {}'.format(dataset))
if cmap is None and cmap_path is not None:
if osp.isfile(cmap_path):
with open(cmap_path) as f:
cmap = cPickle.load(f)
meta['label_2_id'] = label_2_id
meta['id_2_label'] = id_2_label
meta['valid_labels'] = valid_labels
meta['cmap'] = cmap
meta['ident_size'] = ident_size
meta['max_shape'] = meta.get('max_shape', max_shape)
meta[
'cache_images'] = args.cache_images if args.cache_images is not None else cache_images
meta['mx_workspace'] = mx_workspace
return meta
def get_dataset_specs(args, model_specs):
dataset = model_specs['dataset']
meta = {}
meta_path = osp.join('issegm/data', dataset, 'meta.pkl')
if osp.isfile(meta_path):
with open(meta_path) as f:
meta = cPickle.load(f)
label_2_id = None
id_2_label = None
cmap = None
cmap_path = 'data/shared/cmap.pkl'
ident_size = False
cache_images = args.phase == 'train'
mx_workspace = 1650
if dataset == 'ade20k':
num_classes = model_specs.get('classes', 150)
label_2_id = np.arange(-1, 150)
label_2_id[0] = 255
id_2_label = np.arange(1, 256 + 1)
id_2_label[255] = 0
valid_labels = range(1, 150 + 1)
#
if args.split == 'test':
cmap_path = None
#
max_shape = np.array((2100, 2100))
if model_specs.get('balanced', False) and args.split == 'trainval':
meta['image_classes']['trainval'] = meta['image_classes'][
'train'] + meta['image_classes']['val']
elif dataset == 'cityscapes':
sys.path.insert(0, 'data/cityscapesScripts/cityscapesscripts/helpers')
from labels import id2label, trainId2label
#
num_classes = model_specs.get('classes', 19)
label_2_id = 255 * np.ones((256, ))
for l in id2label:
if l in (-1, 255):
continue
label_2_id[l] = id2label[l].trainId
id_2_label = np.array(
[trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)])
valid_labels = sorted(set(id_2_label.ravel()))
#
cmap = np.zeros((256, 3), dtype=np.uint8)
for i in id2label.keys():
cmap[i] = id2label[i].color
#
ident_size = True
#
max_shape = np.array((1024, 2048))
#
if args.split in ('train+', 'trainval+'):
cache_images = False
#
if args.phase in ('val', ):
mx_workspace = 8000
elif dataset == 'coco':
sys.path.insert(0, osp.join(args.data_root, 'PythonAPI'))
from pycocotools.coco import COCO
coco = COCO(
osp.join(args.data_root, 'annotations',
'instances_minival2014.json'))
#
id_2_label = np.array([0] + sorted(coco.getCatIds()))
assert len(id_2_label) == 81
valid_labels = id_2_label.tolist()
num_classes = model_specs.get('classes', 81)
label_2_id = 255 * np.ones((256, ))
for i, l in enumerate(id_2_label):
label_2_id[l] = i
#
max_shape = np.array((640, 640))
elif dataset == 'pascal-context':
num_classes = model_specs.get('classes', 60)
valid_labels = range(num_classes)
#
max_shape = np.array((500, 500))
elif dataset == 'voc':
num_classes = model_specs.get('classes', 21)
valid_labels = range(num_classes)
#
if args.split in ('train++', ):
max_shape = np.array((640, 640))
else:
max_shape = np.array((500, 500))
elif dataset == 'segnet':
num_classes = 13
label_2_id = 255 * np.ones((256, ))
valid_labels = range(num_classes)
#
ident_size = True
#
max_shape = np.array((720, 960))
#
if args.split in ('train+', 'trainval+'):
cache_images = False
#
if args.phase in ('val', ):
mx_workspace = 8000
else:
raise NotImplementedError('Unknow dataset: {}'.format(dataset))
if cmap is None and cmap_path is not None:
if osp.isfile(cmap_path):
with open(cmap_path) as f:
cmap = cPickle.load(f)
meta['label_2_id'] = label_2_id
meta['id_2_label'] = id_2_label
meta['valid_labels'] = valid_labels
meta['cmap'] = cmap
meta['ident_size'] = ident_size
meta['max_shape'] = meta.get('max_shape', max_shape)
meta[
'cache_images'] = args.cache_images if args.cache_images is not None else cache_images
meta['mx_workspace'] = mx_workspace
return meta
def get_dataset_specs_tgt(args, model_specs):
dataset = args.dataset
dataset_tgt = args.dataset_tgt
meta = {}
mine_id = None
mine_id_priority = None
mine_port = args.mine_port
mine_th = args.mine_thresh
cmap_path = 'data/shared/cmap.pkl'
cache_images = args.phase == 'train'
mx_workspace = 1650
sys.path.insert(0, 'data/cityscapesscripts/helpers')
if args.phase == 'train':
mine_id = np.load(args.mine_id_address + '/mine_id.npy')
mine_id_priority = np.load(args.mine_id_address +
'/mine_id_priority.npy')
mine_th = np.zeros(len(mine_id)) # trainId starts from 0
if dataset == 'gta' and dataset_tgt == 'cityscapes':
from labels import id2label, trainId2label
#
label_2_id_tgt = 255 * np.ones((256, ))
for l in id2label:
if l in (-1, 255):
continue
label_2_id_tgt[l] = id2label[l].trainId
id_2_label_tgt = np.array(
[trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)])
valid_labels_tgt = sorted(set(id_2_label_tgt.ravel()))
id_2_label_src = id_2_label_tgt
label_2_id_src = label_2_id_tgt
valid_labels_src = valid_labels_tgt
#
cmap = np.zeros((256, 3), dtype=np.uint8)
for i in id2label.keys():
cmap[i] = id2label[i].color
#
ident_size = True
#
# max_shape_src = np.array((1052, 1914))
max_shape_src = np.array((1024, 2048))
max_shape_tgt = np.array((1024, 2048))
#
if args.split in ('train+', 'trainval+'):
cache_images = False
#
if args.phase in ('val', ):
mx_workspace = 8192
elif dataset == 'synthia' and dataset_tgt == 'cityscapes':
from labels_cityscapes_synthia import id2label as id2label_tgt
from labels_cityscapes_synthia import trainId2label as trainId2label_tgt
from labels_synthia import id2label as id2label_src
label_2_id_src = 255 * np.ones((256, ))
for l in id2label_src:
if l in (-1, 255):
continue
label_2_id_src[l] = id2label_src[l].trainId
label_2_id_tgt = 255 * np.ones((256, ))
for l in id2label_tgt:
if l in (-1, 255):
continue
label_2_id_tgt[l] = id2label_tgt[l].trainId
id_2_label_tgt = np.array([
trainId2label_tgt[_].id for _ in trainId2label_tgt
if _ not in (-1, 255)
])
valid_labels_tgt = sorted(set(id_2_label_tgt.ravel()))
id_2_label_src = None
valid_labels_src = None
#
cmap = np.zeros((256, 3), dtype=np.uint8)
for i in id2label_tgt.keys():
cmap[i] = id2label_tgt[i].color
#
ident_size = True
#
max_shape_src = np.array((760, 1280))
max_shape_tgt = np.array((1024, 2048))
#
if args.split in ('train+', 'trainval+'):
cache_images = False
#
if args.phase in ('val', ):
mx_workspace = 8192
else:
raise NotImplementedError('Unknow dataset: {}'.format(args.dataset))
if cmap is None and cmap_path is not None:
if osp.isfile(cmap_path):
with open(cmap_path) as f:
cmap = cPickle.load(f)
meta['gpus'] = args.gpus
meta['mine_port'] = mine_port
meta['mine_id'] = mine_id
meta['mine_id_priority'] = mine_id_priority
meta['mine_th'] = mine_th
meta['label_2_id_tgt'] = label_2_id_tgt
meta['id_2_label_tgt'] = id_2_label_tgt
meta['valid_labels_tgt'] = valid_labels_tgt
meta['label_2_id_src'] = label_2_id_src
meta['id_2_label_src'] = id_2_label_src
meta['valid_labels_src'] = valid_labels_src
meta['cmap'] = cmap
meta['ident_size'] = ident_size
meta['max_shape_src'] = meta.get('max_shape_src', max_shape_src)
meta['max_shape_tgt'] = meta.get('max_shape_tgt', max_shape_tgt)
meta[
'cache_images'] = args.cache_images if args.cache_images is not None else cache_images
meta['mx_workspace'] = mx_workspace
return meta
def get_dataset_specs(args, model_specs):
dataset = model_specs['dataset']
meta = {}
meta_path = osp.join('issegm/data', dataset, 'meta.pkl')
if osp.isfile(meta_path):
with open(meta_path) as f:
meta = cPickle.load(f)
label_2_id = None
id_2_label = None
cmap = None
cmap_path = 'data/shared/cmap.pkl'
ident_size = False
cache_images = args.phase == 'train'
mx_workspace = 1650
if dataset == 'ade20k':
num_classes = model_specs.get('classes', 150)
label_2_id = np.arange(-1, 150)
label_2_id[0] = 255
id_2_label = np.arange(1, 256+1)
id_2_label[255] = 0
valid_labels = range(1, 150+1)
#
if args.split == 'test':
cmap_path = None
#
max_shape = np.array((2100, 2100))
if model_specs.get('balanced', False) and args.split == 'trainval':
meta['image_classes']['trainval'] = meta['image_classes']['train'] + meta['image_classes']['val']
elif dataset == 'cityscapes':
sys.path.insert(0, 'data/cityscapesScripts/cityscapesscripts/helpers')
from labels import id2label, trainId2label
#
num_classes = model_specs.get('classes', 19)
label_2_id = 255 * np.ones((256,))
for l in id2label:
if l in (-1, 255):
continue
label_2_id[l] = id2label[l].trainId
id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)])
valid_labels = sorted(set(id_2_label.ravel()))
#
cmap = np.zeros((256,3), dtype=np.uint8)
for i in id2label.keys():
cmap[i] = id2label[i].color
#
ident_size = True
#
max_shape = np.array((1024, 2048))
#
if args.split in ('train+', 'trainval+'):
cache_images = False
#
if args.phase in ('val',):
mx_workspace = 8000
elif dataset == 'coco':
sys.path.insert(0, osp.join(args.data_root, 'PythonAPI'))
from pycocotools.coco import COCO
coco = COCO(osp.join(args.data_root, 'annotations', 'instances_minival2014.json'))
#
id_2_label = np.array([0] + sorted(coco.getCatIds()))
assert len(id_2_label) == 81
valid_labels = id_2_label.tolist()
num_classes = model_specs.get('classes', 81)
label_2_id = 255 * np.ones((256,))
for i, l in enumerate(id_2_label):
label_2_id[l] = i
#
max_shape = np.array((640, 640))
elif dataset == 'pascal-context':
num_classes = model_specs.get('classes', 60)
valid_labels = range(num_classes)
#
max_shape = np.array((500, 500))
elif dataset == 'voc':
num_classes = model_specs.get('classes', 21)
valid_labels = range(num_classes)
#
if args.split in ('train++',):
max_shape = np.array((640, 640))
else:
max_shape = np.array((500, 500))
else:
raise NotImplementedError('Unknow dataset: {}'.format(dataset))
if cmap is None and cmap_path is not None:
if osp.isfile(cmap_path):
with open(cmap_path) as f:
cmap = cPickle.load(f)
meta['label_2_id'] = label_2_id
meta['id_2_label'] = id_2_label
meta['valid_labels'] = valid_labels
meta['cmap'] = cmap
meta['ident_size'] = ident_size
meta['max_shape'] = meta.get('max_shape', max_shape)
meta['cache_images'] = args.cache_images if args.cache_images is not None else cache_images
meta['mx_workspace'] = mx_workspace
return meta
