def cli():
parser = argparse.ArgumentParser(
description=__doc__,
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
logs.cli(parser)
nets.cli(parser)
losses.cli(parser)
encoder.cli(parser)
optimize.cli(parser)
datasets.train_cli(parser)
parser.add_argument('-o', '--output', default=None,
help='output file')
parser.add_argument('--stride-apply', default=1, type=int,
help='apply and reset gradients every n batches')
parser.add_argument('--epochs', default=75, type=int,
help='number of epochs to train')
parser.add_argument('--freeze-base', default=0, type=int,
help='number of epochs to train with frozen base')
parser.add_argument('--pre-lr', type=float, default=1e-4,
help='pre learning rate')
parser.add_argument('--update-batchnorm-runningstatistics',
default=False, action='store_true',
help='update batch norm running statistics')
parser.add_argument('--square-edge', default=401, type=int,
help='square edge of input images')
parser.add_argument('--crop-fraction', default=0.5, type=float,
help='crop fraction versus rescale')
parser.add_argument('--lambdas', default=[30.0, 2.0, 2.0, 50.0, 3.0, 3.0],
type=float, nargs='+',
help='prefactor for head losses')
parser.add_argument('--ema', default=1e-3, type=float,
help='ema decay constant')
parser.add_argument('--debug-without-plots', default=False, action='store_true',
help='enable debug but dont plot')
parser.add_argument('--profile', default=None,
help='enables profiling. specify path for chrome tracing file')
parser.add_argument('--disable-cuda', action='store_true',
help='disable CUDA')
args = parser.parse_args()
if args.output is None:
args.output = default_output_file(args)
if args.debug and 'skeleton' not in args.headnets:
raise Exception('add "skeleton" as last headnet to see debug output')
if args.debug_without_plots:
args.debug = True
# 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
return args
def cli():
parser = argparse.ArgumentParser(
description=__doc__,
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
nets.cli(parser)
decoder.cli(parser, force_complete_pose=True)
encoder.cli(parser)
parser.add_argument('--output',
default=None,
help='output filename without file extension')
parser.add_argument('-n', default=0, type=int, help='number of batches')
parser.add_argument('--skip-n', default=0, type=int, help='skip n batches')
parser.add_argument('--dataset',
choices=('val', 'test', 'test-dev'),
default='val',
help='dataset to evaluate')
parser.add_argument('--min-ann',
default=0,
type=int,
help='minimum number of truth annotations')
parser.add_argument('--batch-size', default=1, type=int, help='batch size')
parser.add_argument('--long-edge',
default=641,
type=int,
help='long edge of input images')
parser.add_argument('--loader-workers',
default=None,
type=int,
help='number of workers for data loading')
parser.add_argument('--skip-existing',
default=False,
action='store_true',
help='skip if output eval file exists already')
parser.add_argument('--disable-cuda',
action='store_true',
help='disable CUDA')
parser.add_argument('--write-predictions',
default=False,
action='store_true',
help='write a json and a zip file of the predictions')
parser.add_argument('--all-images',
default=False,
action='store_true',
help='run over all images irrespective of catIds')
group = parser.add_argument_group('logging')
group.add_argument('--debug',
default=False,
action='store_true',
help='print debug messages')
args = parser.parse_args()
logging.basicConfig()
log_level = logging.INFO if not args.debug else logging.DEBUG
logging.getLogger('openpifpaf').setLevel(log_level)
LOG.setLevel(log_level)
if args.loader_workers is None:
args.loader_workers = args.batch_size
if args.dataset == 'val':
args.image_dir = IMAGE_DIR_VAL
args.annotation_file = ANNOTATIONS_VAL
elif args.dataset == 'test':
args.image_dir = IMAGE_DIR_TEST
args.annotation_file = ANNOTATIONS_TEST
elif args.dataset == 'test-dev':
args.image_dir = IMAGE_DIR_TEST
args.annotation_file = ANNOTATIONS_TESTDEV
else:
raise Exception
if args.dataset in ('test', 'test-dev'
) and not args.write_predictions and not args.debug:
raise Exception('have to use --write-predictions for this dataset')
if args.dataset in ('test',
'test-dev') and not args.all_images and not args.debug:
raise Exception('have to use --all-images for this dataset')
# 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
# generate a default output filename
if args.output is None:
args.output = default_output_name(args)
return args
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(
description=__doc__,
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
logs.cli(parser)
nets.cli(parser)
losses.cli(parser)
encoder.cli(parser)
optimize.cli(parser)
datasets.train_cli(parser)
parser.add_argument('-o', '--output', default=None, help='output file')
parser.add_argument('--stride-apply',
default=1,
type=int,
help='apply and reset gradients every n batches')
parser.add_argument('--epochs',
default=75,
type=int,
help='number of epochs to train')
parser.add_argument('--freeze-base',
default=0,
type=int,
help='number of epochs to train with frozen base')
parser.add_argument('--pre-lr',
type=float,
default=1e-4,
help='pre learning rate')
parser.add_argument('--rescale-images',
type=float,
default=1.0,
help='overall image rescale factor')
parser.add_argument('--orientation-invariant',
default=False,
action='store_true',
help='augment with random orientations')
parser.add_argument('--update-batchnorm-runningstatistics',
default=False,
action='store_true',
help='update batch norm running statistics')
parser.add_argument('--square-edge',
default=401,
type=int,
help='square edge of input images')
parser.add_argument('--ema',
default=1e-3,
type=float,
help='ema decay constant')
parser.add_argument('--disable-cuda',
action='store_true',
help='disable CUDA')
parser.add_argument('--no-augmentation',
dest='augmentation',
default=True,
action='store_false',
help='do not apply data augmentation')
group = parser.add_argument_group('debug')
group.add_argument('--debug-pif-indices',
default=[],
nargs='+',
type=int,
help='indices of PIF fields to create debug plots for')
group.add_argument('--debug-paf-indices',
default=[],
nargs='+',
type=int,
help='indices of PAF fields to create debug plots for')
group.add_argument(
'--profile',
default=None,
help='enables profiling. specify path for chrome tracing file')
args = parser.parse_args()
if args.output is None:
args.output = default_output_file(args)
if args.debug and 'skeleton' not in args.headnets:
raise Exception('add "skeleton" as last headnet to see debug output')
if args.freeze_base and args.checkpoint:
raise Exception('remove --freeze-base when running from a checkpoint')
if args.debug_pif_indices or args.debug_paf_indices:
args.debug = True
# 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
return args