def _get_cityscapes_path(exists=False):
return settings.get_data_path(config_name='cityscapes',
dnnlib_template=os.path.join(
'<DATASETS>', 'research', 'gfrench',
'cityscapes_full',
'cityscapes_segmentation.zip'),
exists=exists)
def _get_isic2017_path(exists=False):
return settings.get_data_path(
config_name='isic2017',
dnnlib_template=os.path.join('<DATASETS>', 'research', 'gfrench', 'isic2017',
'isic2017_segmentation.zip'),
exists=exists
)
def __init__(self, n_val, val_rng, trainval_perm):
super(CamVidDataSource, self).__init__(
settings.get_data_path(config_name='camvid',
dnnlib_template=os.path.join(
'<DATASETS>', 'research', 'gfrench',
'camvid', 'CamVidData.zip')))
sample_names = set()
sample_name_to_dir_name = {}
for filename in self.zip_file.namelist():
dir_name, sample_name = os.path.split(filename)
if not dir_name.endswith('annot') and os.path.splitext(
sample_name)[1].lower() == '.png':
sample_names.add(sample_name)
sample_name_to_dir_name[sample_name] = dir_name
sample_names = list(sample_names)
sample_names.sort()
self.x_names = [
sample_name_to_dir_name[name] + '/' + name for name in sample_names
]
self.y_names = [
sample_name_to_dir_name[name] + 'annot/' + name
for name in sample_names
]
self.sample_names = sample_names
self.train_ndx = np.array([
i for i in range(len(self.x_names))
if os.path.split(self.x_names[i])[0].endswith('train')
])
self.val_ndx = np.array([
i for i in range(len(self.x_names))
if os.path.split(self.x_names[i])[0].endswith('val')
])
self.test_ndx = np.array([
i for i in range(len(self.x_names))
if os.path.split(self.x_names[i])[0].endswith('test')
])
if n_val > 0 and n_val < len(self.val_ndx):
self.val_ndx = self.val_ndx[val_rng.permutation(len(
self.val_ndx))[:n_val]]
# weights when using median frequency balancing used in SegNet paper
# https://arxiv.org/pdf/1511.00561.pdf
# The numbers were generated by https://github.com/yandex/segnet-torch/blob/master/datasets/camvid-gen.lua
self.class_weights = np.array([
0.58872014284134, 0.51052379608154, 2.6966278553009,
0.45021694898605, 1.1785038709641, 0.77028578519821,
2.4782588481903, 2.5273461341858, 1.0122526884079, 3.2375309467316,
4.1312313079834, 0
])
self.class_names = [
'Sky', 'Building', 'Pole', 'Road', 'Pavement', 'Tree',
'SignSymbol', 'Fence', 'Car', 'Pedestrian', 'Bicyclist', 'void'
]
self.num_classes_all = len(self.class_names)
self.num_classes = len(self.class_names) - 1
def _get_pascal_path(exists=False):
return settings.get_data_path(config_name='pascal_voc',
dnnlib_template=os.path.join(
'<DATASETS>', 'research', 'pascal_voc'),
exists=exists)
def _get_pascal_path(exists=False):
return settings.get_data_path(config_name='pascal_voc', exists=exists)