def main():
args = train_common.parse_args()
args.logs_dir = osp.join(here, 'logs')
# Dataset.
args.dataset = 'arc2017_real'
train_data = instance_occlsegm_lib.datasets.apc.\
ARC2017InstanceSegmentationDataset(split='train', aug='standard')
test_data = instance_occlsegm_lib.datasets.apc.\
ARC2017InstanceSegmentationDataset(split='test')
args.class_names = tuple(test_data.class_names)
# Model.
args.min_size = 600
args.max_size = 1000
args.anchor_scales = (4, 8, 16, 32)
# Run training!.
train_common.train(
args=args,
train_data=train_data,
test_data=test_data,
evaluator_type='coco',
)
def main():
args = train_common.parse_args()
args.logs_dir = osp.join(here, 'logs')
# Dataset. For demonstration with few images, we use same dataset
# for both train and test.
root_dir = osp.join(
here,
'src/labelme/examples/instance_segmentation/data_dataset_voc',
)
args.dataset = 'custom'
# 1 epoch = 3 images -> 60 images
train_data = [VOCLikeDataset(root_dir=root_dir)] * 20
train_data = chainer.datasets.ConcatenatedDataset(*train_data)
test_data = VOCLikeDataset(root_dir=root_dir)
args.class_names = tuple(VOCLikeDataset.class_names.tolist())
# Model.
args.min_size = 600
args.max_size = 1000
args.anchor_scales = (4, 8, 16, 32)
train_common.train(
args=args,
train_data=train_data,
test_data=test_data,
evaluator_type='voc',
)
def main():
args = train_common.parse_args()
args.logs_dir = osp.join(here, 'logs')
# Dataset.
args.dataset = 'coco'
train_data = chainer.datasets.ConcatenatedDataset(
cmr.datasets.COCOInstanceSegmentationDataset('train'),
cmr.datasets.COCOInstanceSegmentationDataset('valminusminival'),
)
test_data = cmr.datasets.COCOInstanceSegmentationDataset(
'minival',
use_crowd=True,
return_crowd=True,
return_area=True,
)
args.class_names = tuple(test_data.class_names.tolist())
# Model.
args.min_size = 800
args.max_size = 1333
args.anchor_scales = (2, 4, 8, 16, 32)
# Run training!.
train_common.train(
args=args,
train_data=train_data,
test_data=test_data,
evaluator_type='coco',
)
def main():
parser = train_common.get_parser()
parser.add_argument(
'--exclude-arc2017',
action='store_true',
help='Exclude ARC2017 objects from synthetic',
)
parser.add_argument(
'--background',
choices=['tote', 'tote+shelf'],
default='tote',
help='background image in 2D synthesis',
)
args = parser.parse_args()
args.logs_dir = osp.join(here, 'logs')
# Dataset.
args.dataset = 'synthetic'
train_data = \
grasp_fusion.datasets.SyntheticInstanceSegmentationDataset(
augmentation=True,
augmentation_level='all',
exclude_arc2017=args.exclude_arc2017,
background=args.background,
)
test_data = \
grasp_fusion.datasets.RealInstanceSegmentationDataset()
args.class_names = tuple(test_data.class_names.tolist())
# Model.
args.min_size = 600
args.max_size = 1000
args.anchor_scales = (4, 8, 16, 32)
# Run training!.
train_common.train(
args=args,
train_data=train_data,
test_data=test_data,
evaluator_type='coco',
)
def main():
args = train_common.parse_args()
args.logs_dir = osp.join(here, 'logs')
# Dataset.
# args.dataset = 'coco'
# train_data = chainer.datasets.ConcatenatedDataset(
# cmr.datasets.COCOInstanceSegmentationDataset('train'),
# cmr.datasets.COCOInstanceSegmentationDataset('valminusminival'),
# )
# test_data = cmr.datasets.COCOInstanceSegmentationDataset(
# 'minival',
# use_crowd=True,
# return_crowd=True,
# return_area=True,
# )
# args.class_names = tuple(test_data.class_names.tolist())
args.dataset = 'original'
train_data = dataset.OomugiDataset()
test_data = dataset.OomugiDataset(test=True)
args.class_names = ['leaf']
# Model.
args.min_size = 800
args.max_size = 1333
args.anchor_scales = (2, 4, 8, 16, 32)
args.ratios = (0.11, 0.14, 0.2, 0.25, 0.33, 0.5, 1, 2, 3, 4, 5, 7, 9)
# Run training!.
train_common.train(
args=args,
train_data=train_data,
test_data=test_data,
evaluator_type='coco',
)
def main():
args = train_common.parse_args()
args.logs_dir = osp.join(here, 'logs')
# Dataset.
args.dataset = 'voc'
train_data = cmr.datasets.SBDInstanceSegmentationDataset('train')
test_data = cmr.datasets.SBDInstanceSegmentationDataset('val')
args.class_names = tuple(train_data.class_names.tolist())
# Model.
args.min_size = 600
args.max_size = 1000
args.anchor_scales = (4, 8, 16, 32)
train_common.train(
args=args,
train_data=train_data,
test_data=test_data,
evaluator_type='voc',
)