def factory_from_args(args):
# Data
if args.glob:
args.images += glob.glob(args.glob)
if not args.images:
raise Exception("no image files given")
if args.path_gt is None:
args.show_all = True
# Models
dic_models = download_checkpoints(args)
args.checkpoint = dic_models['keypoints']
logger.configure(args, LOG) # logger first
# Devices
args.device = torch.device('cpu')
args.pin_memory = False
if torch.cuda.is_available():
args.device = torch.device('cuda')
args.pin_memory = True
LOG.debug('neural network device: %s', args.device)
# Add visualization defaults
args.figure_width = 10
args.dpi_factor = 1.0
if args.mode == 'stereo':
args.batch_size = 2
args.images = sorted(args.images)
else:
args.batch_size = 1
# Patch for stereo images with batch_size = 2
if args.batch_size == 2 and not args.long_edge:
args.long_edge = 1238
LOG.info("Long-edge set to %i", args.long_edge)
# Make default pifpaf argument
args.force_complete_pose = True
LOG.info("Force complete pose is active")
# Configure
decoder.configure(args)
network.Factory.configure(args)
show.configure(args)
visualizer.configure(args)
return args, dic_models
def factory_from_args(args):
# Data
if args.glob:
args.images += glob.glob(args.glob)
if not args.images:
raise Exception("no image files given")
# Model
if not args.checkpoint:
if os.path.exists(OPENPIFPAF_PATH):
args.checkpoint = OPENPIFPAF_PATH
else:
LOG.info("Checkpoint for OpenPifPaf not specified and default model not found in 'data/models'. "
"Using a ShuffleNet backbone")
args.checkpoint = 'shufflenetv2k30'
logger.configure(args, LOG) # logger first
# Devices
args.device = torch.device('cpu')
args.pin_memory = False
if torch.cuda.is_available():
args.device = torch.device('cuda')
args.pin_memory = True
LOG.debug('neural network device: %s', args.device)
# Add visualization defaults
args.figure_width = 10
args.dpi_factor = 1.0
if args.net == 'monstereo':
args.batch_size = 2
else:
args.batch_size = 1
# Make default pifpaf argument
args.force_complete_pose = True
LOG.info("Force complete pose is active")
# Configure
decoder.configure(args)
network.Factory.configure(args)
show.configure(args)
visualizer.configure(args)
return args
def __init__(self,
checkpoint=None,
basenet=None,
headnets=None,
pretrained=True,
two_scale=False,
multi_scale=False,
multi_scale_hflip=True,
cross_talk=0.0,
download_progress=True,
head_dropout=0.0,
head_quad=1,
seed_threshold=0.5,
instance_threshold=0.1,
keypoint_threshold=None,
decoder_workers=None,
dense_connections=False,
dense_coupling=0.01,
caf_seeds=False,
force_complete_pose=True,
profile_decoder=None,
cif_th=0.1,
caf_th=0.1,
connection_method="blend",
greedy=False,
video_output="",
xls_output="",
quiet=False,
debug=False,
skip_frames=1,
max_frames=False,
start_frame=0,
confidence_threshold=0.1):
self.checkpoint = checkpoint
self.basenet = basenet
self.headnets = headnets
self.pretrained = pretrained
self.two_scale = two_scale
self.multi_scale = multi_scale
self.multi_scale_hflip = multi_scale_hflip
self.cross_talk = cross_talk
self.download_progress = download_progress
self.head_dropout = head_dropout
self.head_quad = head_quad
self.seed_threshold = seed_threshold
self.instance_threshold = instance_threshold
self.keypoint_threshold = keypoint_threshold
self.decoder_workers = decoder_workers
self.dense_connections = dense_connections
self.dense_coupling = dense_coupling
self.caf_seeds = caf_seeds
self.force_complete_pose = force_complete_pose
self.profile_decoder = profile_decoder
self.cif_th = cif_th
self.caf_th = caf_th
self.connection_method = connection_method
self.greedy = greedy
self.quiet = quiet
self.debug = debug
self.skip_frames = skip_frames
self.max_frames = max_frames
self.start_frame = start_frame
self.confidence_threshold = confidence_threshold
class InternalConfig:
def __init__(self):
self.debug_cifhr = None
self.debug_cif_c = None
self.debug_cif_v = None
self.debug_cifdet_c = None
self.debug_cifdet_v = None
self.debug_caf_c = None
self.debug_caf_v = None
self.debug_indices = []
self.debug_images = False
self.show_box = False,
self.show_joint_scales = False,
self.show_decoding_order = False,
self.show_frontier_order = False,
self.show_only_decoded_connections = False,
self.show_joint_confidences = False,
# configure logging
self.log_level = logging.INFO
if quiet:
self.log_level = logging.WARNING
if debug:
self.log_level = logging.DEBUG
logging.basicConfig()
logging.getLogger('openpifpaf').setLevel(self.log_level)
LOG.setLevel(self.log_level)
internal_config = InternalConfig()
network.configure(self)
pifpaf.show.configure(internal_config)
visualizer.configure(internal_config)
# add args.device
self.device = torch.device('cpu')
if torch.cuda.is_available():
self.device = torch.device('cuda')
LOG.debug('neural network device: %s', self.device)
self.model, _ = network.factory_from_args(self)
self.model = self.model.to(self.device)
self.processor = decoder.factory_from_args(self, self.model)
def cli():
parser = argparse.ArgumentParser(
prog='python3 -m openpifpaf.train',
usage='%(prog)s [options]',
description=__doc__,
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument(
'--version',
action='version',
version='OpenPifPaf {version}'.format(version=__version__))
parser.add_argument('-o', '--output', default=None, help='output file')
parser.add_argument('--disable-cuda',
action='store_true',
help='disable CUDA')
parser.add_argument('--ddp',
default=False,
action='store_true',
help='[experimental] DistributedDataParallel')
parser.add_argument('--local_rank',
default=None,
type=int,
help='[experimental] for torch.distributed.launch')
parser.add_argument('--no-sync-batchnorm',
dest='sync_batchnorm',
default=True,
action='store_false',
help='[experimental] in ddp, to not use syncbatchnorm')
logger.cli(parser)
network.Factory.cli(parser)
network.losses.Factory.cli(parser)
network.Trainer.cli(parser)
encoder.cli(parser)
optimize.cli(parser)
datasets.cli(parser)
show.cli(parser)
visualizer.cli(parser)
args = parser.parse_args()
logger.configure(args, LOG)
if args.log_stats:
logging.getLogger('openpifpaf.stats').setLevel(logging.DEBUG)
# DDP with SLURM
slurm_process_id = os.environ.get('SLURM_PROCID')
if args.ddp and slurm_process_id is not None:
if torch.cuda.device_count() > 1:
LOG.warning(
'Expected one GPU per SLURM task but found %d. '
'Try with "srun --gpu-bind=closest ...". Still trying.',
torch.cuda.device_count())
# if there is more than one GPU available, assume that other SLURM tasks
# have access to the same GPUs and assign GPUs uniquely by slurm_process_id
args.local_rank = (int(slurm_process_id) % torch.cuda.device_count()
if torch.cuda.device_count() > 0 else 0)
os.environ['RANK'] = slurm_process_id
if not os.environ.get('WORLD_SIZE') and os.environ.get('SLURM_NTASKS'):
os.environ['WORLD_SIZE'] = os.environ.get('SLURM_NTASKS')
LOG.info('found SLURM process id: %s', slurm_process_id)
LOG.info(
'distributed env: master=%s port=%s rank=%s world=%s, '
'local rank (GPU)=%d', os.environ.get('MASTER_ADDR'),
os.environ.get('MASTER_PORT'), os.environ.get('RANK'),
os.environ.get('WORLD_SIZE'), args.local_rank)
# add args.device
args.device = torch.device('cpu')
args.pin_memory = False
if not args.disable_cuda and torch.cuda.is_available():
args.device = torch.device('cuda')
args.pin_memory = True
LOG.info('neural network device: %s (CUDA available: %s, count: %d)',
args.device, torch.cuda.is_available(), torch.cuda.device_count())
# output
if args.output is None:
args.output = default_output_file(args)
os.makedirs('outputs', exist_ok=True)
network.Factory.configure(args)
network.losses.Factory.configure(args)
network.Trainer.configure(args)
encoder.configure(args)
datasets.configure(args)
show.configure(args)
visualizer.configure(args)
return args
def cli():
parser = argparse.ArgumentParser(
prog='python3 -m openpifpaf.predict',
usage='%(prog)s [options] images',
description=__doc__,
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument('--version', action='version',
version='OpenPifPaf {version}'.format(version=__version__))
decoder.cli(parser)
logger.cli(parser)
network.Factory.cli(parser)
show.cli(parser)
visualizer.cli(parser)
parser.add_argument('images', nargs='*',
help='input images')
parser.add_argument('--glob',
help='glob expression for input images (for many images)')
parser.add_argument('-o', '--image-output', default=None, nargs='?', const=True,
help='Whether to output an image, '
'with the option to specify the output path or directory')
parser.add_argument('--json-output', default=None, nargs='?', const=True,
help='Whether to output a json file, '
'with the option to specify the output path or directory')
parser.add_argument('--batch-size', default=1, type=int,
help='processing batch size')
parser.add_argument('--long-edge', default=None, type=int,
help='rescale the long side of the image (aspect ratio maintained)')
parser.add_argument('--loader-workers', default=None, type=int,
help='number of workers for data loading')
parser.add_argument('--precise-rescaling', dest='fast_rescaling',
default=True, action='store_false',
help='use more exact image rescaling (requires scipy)')
parser.add_argument('--disable-cuda', action='store_true',
help='disable CUDA')
args = parser.parse_args()
logger.configure(args, LOG) # logger first
# add args.device
args.device = torch.device('cpu')
args.pin_memory = False
if not args.disable_cuda and torch.cuda.is_available():
args.device = torch.device('cuda')
args.pin_memory = True
LOG.info('neural network device: %s (CUDA available: %s, count: %d)',
args.device, torch.cuda.is_available(), torch.cuda.device_count())
decoder.configure(args)
network.Factory.configure(args)
show.configure(args)
visualizer.configure(args)
if args.loader_workers is None:
args.loader_workers = args.batch_size
# glob
if args.glob:
args.images += glob.glob(args.glob)
if not args.images:
raise Exception("no image files given")
return args