def parse_args():
parser = TrainingParser(
description='Distance metric learning using prototypical loss',
default_logfile='train_prototype.log',
default_model_prefix='prototype_model')
parser.add_argument(
'--batch-size',
type=int,
default=32,
help='Number of samples in a batch per device. Default is 32')
parser.add_argument(
'--nc',
type=int,
default=12,
help='Number of classes in each episode. Default is 12')
parser.add_argument(
'--nq',
type=int,
default=5,
help='Number of query examples in each episode. Default is 5.')
parser.add_argument(
'--ns',
type=int,
default=5,
help='Number of support examples in each episode. Default is 5.')
parser.add_argument('--epochs',
type=int,
default=30,
help='number of training epochs. default is 30.')
parser.add_argument('--iteration-per-epoch',
type=int,
default=100,
help='Number of iterations per epoch. Default is 100')
parser.add_argument('--lr',
type=float,
default=0.0001,
help='learning rate. default is 0.0001.')
parser.add_argument('--factor',
type=float,
default=0.5,
help='learning rate schedule factor. default is 0.5.')
parser.add_argument('--wd',
type=float,
default=0.00001,
help='weight decay rate. default is 0.00001.')
parser.add_argument(
'--steps',
type=str,
default='12,14,16,18',
help='epochs to update learning rate. default is 12,14,16,18.')
opt = parser.parse_args()
if opt.logfile.lower() != 'none':
logging.basicConfig(filename=append_postfix(opt.logfile,
opt.log_postfix),
level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
return opt
def parse_args():
parser = TrainingParser(
description='Distance metric learning using structured clusture loss',
default_logfile='train_clusterloss.log',
default_model_prefix='clusterloss_model')
parser.add_argument('--batch-size',
type=int,
default=120,
help='Number of samples in a batch. Default is 120')
parser.add_argument(
'--batch-k',
type=int,
default=5,
help=
'Number of images per class in a batch. Used only if iteration-per-epoch > 0. Default is 5.'
)
parser.add_argument('--epochs',
type=int,
default=25,
help='number of training epochs. default is 25.')
parser.add_argument('--epsilon',
type=float,
default=1e-7,
help='learning rate. default is 0.01.')
parser.add_argument('--lr',
type=float,
default=0.0001,
help='learning rate. default is 0.0001.')
parser.add_argument('--factor',
type=float,
default=0.5,
help='learning rate schedule factor. default is 0.5.')
parser.add_argument('--wd',
type=float,
default=0.00001,
help='weight decay rate. default is 0.00001.')
parser.add_argument(
'--steps',
type=str,
default='12,14,16,18',
help='epochs to update learning rate. default is 12,14,16,18.')
parser.add_argument(
'--iteration-per-epoch',
type=int,
default=0,
help='Number of iterations per epoch for iteration-based training')
parser.add_argument('--decrease-cnn-lr',
action="store_true",
help='Use a lower LR on the feature extractor')
opt = parser.parse_args()
if opt.logfile.lower() != 'none':
logging.basicConfig(filename=append_postfix(opt.logfile,
opt.log_postfix),
level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
return opt
def parse_args():
parser = TrainingParser(description='Distance metric learning using normproxies',
default_logfile='train_normproxy.log', default_model_prefix='normproxy_model')
parser.add_argument('--batch-size', type=int, default=75,
help='Number of samples in a batch. Default is 75')
parser.add_argument('--epochs', type=int, default=30,
help='number of training epochs. default is 60.')
parser.add_argument('--lr', type=float, default=0.0001,
help='learning rate. default is 0.0001.')
parser.add_argument('--factor', type=float, default=0.1,
help='learning rate schedule factor. default is 0.5.')
parser.add_argument('--wd', type=float, default=0.00001,
help='weight decay rate. default is 0.00001.')
parser.add_argument('--steps', type=str, default='15',
help='Epochs to update learning rate. Negative number represents periodic decrease.'
'Zero means no steps. Default is -1')
parser.add_argument('--binarize', action="store_true",
help='Thresholds the embedding into a binary vector')
parser.add_argument('--epsilon', type=float, default=1e-2,
help='Optimizer epsilon. default is 0.01.')
parser.add_argument('--label-smooth', type=float, default=0.1,
help='Label smoothing. Default is 0')
parser.add_argument('--temperature', type=float, default=0.05,
help='Sigma temperature constant. Default is 0.05')
parser.add_argument('--batch-k', type=int, default=0,
help='Number of images per class for episodic sampling. 0 Will turn it off.')
parser.add_argument('--start-epoch', type=int, default=1,
help='Epoch to start at, >1 means loading parameters')
parser.add_argument('--no-fc', action="store_true", help='Skips the fully-connected layer in the model.')
parser.add_argument('--static-proxies', action="store_true", help='Proxies will not be learned.')
parser.add_argument('--no-dropout', dest='dropout', action="store_false", help='Do not add dropout layer to the model.')
parser.add_argument('--similarity', type=str, choices=['euclidean', 'cosine'], default='euclidean')
parser.set_defaults(
embed_dim=2048,
lr=0.001,
wd=1e-4,
)
opt = parser.parse_args()
if opt.logfile.lower() != 'none':
logging.basicConfig(filename=append_postfix(opt.logfile, opt.log_postfix), level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
return opt
def parse_args():
parser = TrainingParser(
description=
'Distance metric learning using triplet loss with semihard mining',
default_logfile='train_triplet_semihard.log',
default_model_prefix='triplet_semihard_model')
parser.add_argument(
'--batch-size',
type=int,
default=128,
help='Number of samples in a batch per compute unit. Default is 128.')
parser.add_argument('--epochs',
type=int,
default=60,
help='number of training epochs. default is 60.')
parser.add_argument('--lr',
type=float,
default=0.0001,
help='learning rate. default is 0.0001.')
parser.add_argument('--factor',
type=float,
default=0.5,
help='learning rate schedule factor. default is 0.5.')
parser.add_argument('--wd',
type=float,
default=0.00001,
help='weight decay rate. default is 0.00001.')
parser.add_argument(
'--steps',
type=str,
default='20,30,40',
help='epochs to update learning rate. default is 20,30,40.')
opt = parser.parse_args()
if opt.logfile.lower() != 'none':
logging.basicConfig(filename=append_postfix(opt.logfile,
opt.log_postfix),
level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
return opt
def parse_args():
parser = TrainingParser(description='Deep Randomized Ensembles for Metric Learning',
default_logfile='train_dreml.log',
default_model_prefix='dreml_model')
parser.add_argument('--batch-size', type=int, default=128,
help='Number of samples in a batch. Default is 128')
parser.add_argument('--epochs', type=int, default=12,
help='number of training epochs. default is 20.')
parser.add_argument('--lr', type=float, default=0.01,
help='learning rate. default is 0.0001.')
parser.add_argument('--factor', type=float, default=0.1,
help='learning rate schedule factor. default is 0.1.')
parser.add_argument('--wd', type=float, default=5e-4,
help='weight decay rate. default is 5e-4.')
parser.add_argument('--steps', type=str, default='-4',
help='Epochs to update learning rate. Negative number represents periodic decrease.'
'Zero means no steps. Default is -4')
parser.add_argument('--loss', type=str, default='nca',
help='Which loss to use: [triplet, nca, xentropy]')
parser.add_argument('--epsilon', type=float, default=1e-2,
help='Optimizer epsilon. default is 0.01.')
parser.add_argument('--label-smooth', type=float, default=0,
help='Label smoothing. Default is 0')
parser.add_argument('--embedding-multiplier', type=float, default=3,
help='Multiplies normalized embeddings and proxies. Default is 3')
parser.add_argument('-L', '--number-of-ensembles', dest='L', type=int, default=48,
help='Number of ensembles.')
parser.add_argument('-D', '--meta-classes', dest='D', type=int, default=12,
help='Number of meta-classes.')
parser.add_argument('--static-proxies', action="store_true",
help='Do not learn proxies, but keep them fixed.')
parser.add_argument('--data-shape', type=int, default=224,
help='Input data size')
opt = parser.parse_args()
if opt.logfile.lower() != 'none':
logging.basicConfig(filename=append_postfix(opt.logfile, opt.log_postfix), level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
return opt
def parse_args():
parser = TrainingParser(
description='Distance metric learning using proxies',
default_logfile='train_proxy.log',
default_model_prefix='proxy_model')
parser.add_argument(
'--batch-size',
type=int,
default=32,
help='training batch size per device (CPU/GPU). default is 32.')
parser.add_argument(
'--batch-k',
type=int,
default=5,
help=
'Number of images per class in a batch. Used only if iteration-per-epoch > 0. Default is 5.'
)
parser.add_argument(
'--loss',
type=str,
default='triplet',
help='Which loss to use: [nca, triplet, proxymargin, xentropy]')
parser.add_argument('--epsilon',
type=float,
default=1e-2,
help='Optimizer epsilon. default is 0.01.')
parser.add_argument(
'--lr',
default=None,
type=float,
help=
'Learning rate for the whole model. Overwrites specific learning rates.'
)
parser.add_argument('--lr-embedding',
default=1e-5,
type=float,
help='Learning rate for embedding.')
parser.add_argument(
'--lr-inception',
default=1e-3,
type=float,
help='Learning rate for Inception, excluding embedding layer.')
parser.add_argument('--lr-proxynca',
default=1e-3,
type=float,
help='Learning rate for proxies of Proxy NCA.')
parser.add_argument('--wd',
type=float,
default=5e-4,
help='weight decay rate. default is 5e-4.')
parser.add_argument('--factor',
type=float,
default=1e-1,
help='learning rate schedule factor. default is 1e-1.')
parser.add_argument('--epochs',
type=int,
default=20,
help='number of training epochs. default is 20.')
parser.add_argument(
'--steps',
type=str,
default='3,10,16',
help=
'Epochs to update learning rate. Negative number represents periodic decrease.'
'Zero means no steps. Default is 3,10,16')
parser.add_argument(
'--iteration-per-epoch',
type=int,
default=0,
help='Number of iterations per epoch for iteration-based training')
parser.add_argument('--label-smooth',
type=float,
default=0,
help='Label smoothing. Default is 0')
parser.add_argument(
'--embedding-multiplier',
type=float,
default=3,
help='Multiplies normalized embeddings and proxies. Default is 3')
parser.add_argument(
'--temperature',
type=float,
default=1.0,
help='Temperature scaling for NCA and XEntropy losses. Default is 1')
opt = parser.parse_args()
if opt.logfile.lower() != 'none':
logging.basicConfig(filename=append_postfix(opt.logfile, opt.loss,
opt.log_postfix),
level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
return opt
def parse_args():
parser = TrainingParser(
description='Distance metric learning using angular loss',
default_logfile='train_angular.log',
default_model_prefix='angular_model')
parser.add_argument(
'--batch-size',
type=int,
default=128,
help='Number of samples in a batch, this equals to 2N. Default is 128')
parser.add_argument('--epochs',
type=int,
default=30,
help='number of training epochs. default is 25.')
parser.add_argument('--epoch-length',
type=int,
default=200,
help='Number of iterations per epoch. Default is 200.')
parser.add_argument('--lr',
type=float,
default=1e-5,
help='learning rate. default is 0.0001.')
parser.add_argument('--factor',
type=float,
default=0.5,
help='learning rate schedule factor. default is 0.5.')
parser.add_argument('--wd',
type=float,
default=0.00001,
help='weight decay rate. default is 0.00001.')
parser.add_argument('--alpha',
type=float,
default=45,
help='Alpha constant in degrees. Default is 45.')
parser.add_argument(
'--angular-lambda',
type=float,
default=0.0,
help=
'Angular loss factor used together with NPair loss. 0 turns of NPair loss. Default is 0.'
)
parser.add_argument(
'--l2reg-weight',
type=float,
default=0.0005,
help='Weight of L2 regularization for feature vectors. '
'Default is 0.25 * 0.002. Used only in NL&AL.')
parser.add_argument(
'--steps',
type=str,
default='10,20',
help='epochs to update learning rate. default is 12,14,16,18.')
parser.add_argument(
'--same-image-sampling',
type=float,
default=0.1,
help='Chance to sample both items from the same image. Default is 0.1')
parser.add_argument('--symmetric-loss',
action="store_true",
help='Use symmetric loss. Used only in NL&AL.')
parser.add_argument('--decrease-cnn-lr',
action="store_true",
help='Use a lower LR on the feature extractor')
opt = parser.parse_args()
if opt.logfile.lower() != 'none':
logging.basicConfig(filename=append_postfix(opt.logfile,
opt.log_postfix),
level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
return opt
def parse_args():
parser = TrainingParser(
description=
'Distance metric learning using ranked list loss with semihard mining',
default_logfile='train_rankedlist.log',
default_model_prefix='rankedlist_model')
parser.add_argument('--batch-size',
type=int,
default=180,
help='Number of samples in a batch. Default is 180')
parser.add_argument('--epochs',
type=int,
default=60,
help='number of training epochs. default is 60.')
parser.add_argument('--lr',
type=float,
default=0.0001,
help='learning rate. default is 0.0001.')
parser.add_argument('--factor',
type=float,
default=0.5,
help='learning rate schedule factor. default is 0.5.')
parser.add_argument('--wd',
type=float,
default=0.00001,
help='weight decay rate. default is 0.00001.')
parser.add_argument(
'--steps',
type=str,
default='20,30,40',
help='epochs to update learning rate. default is 20,30,40.')
parser.add_argument('--batch-k',
type=int,
default=3,
help='Number of images per class')
parser.add_argument('--alpha',
type=float,
default=1.2,
help='Margin for negatives')
parser.add_argument('--margin',
type=float,
default=0.4,
help='Margin for positives')
parser.add_argument('--temperature',
type=float,
default=10,
help='Temperature for negatives')
parser.add_argument(
'--iteration-per-epoch',
type=int,
default=200,
help='Number of iterations per epoch for iteration-based training')
parser.add_argument('--bottleneck-layers', type=str, default='')
opt = parser.parse_args()
assert opt.batch_size % opt.batch_k == 0, 'Batch size must be divisible by batch-k'
if opt.logfile.lower() != 'none':
logging.basicConfig(filename=append_postfix(opt.logfile,
opt.log_postfix),
level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
return opt
def parse_args():
parser = TrainingParser(
description=
'Distance metric learning with marginloss and distance-weighted sampling.',
default_logfile='train_margin.log',
default_model_prefix='margin_loss_model')
parser.add_argument(
'--batch-size',
type=int,
default=125,
help='Number of samples in a batch per compute unit. Default is 125.'
'Must be divisible with batch-k.')
parser.add_argument(
'--batch-k',
type=int,
default=5,
help='number of images per class in a batch. default is 5.')
parser.add_argument('--epochs',
type=int,
default=20,
help='number of training epochs. default is 20.')
parser.add_argument('--lr',
type=float,
default=0.0001,
help='learning rate. default is 0.0001.')
parser.add_argument(
'--lr-beta',
type=float,
default=0.1,
help='learning rate for the beta in margin based loss. default is 0.1.'
)
parser.add_argument(
'--margin',
type=float,
default=0.2,
help='margin for the margin based loss. default is 0.2.')
parser.add_argument('--beta',
type=float,
default=1.2,
help='initial value for beta. default is 1.2.')
parser.add_argument(
'--nu',
type=float,
default=0.0,
help='regularization parameter for beta. default is 0.0.')
parser.add_argument('--factor',
type=float,
default=0.5,
help='learning rate schedule factor. default is 0.5.')
parser.add_argument(
'--steps',
type=str,
default='12,14,16,18',
help='epochs to update learning rate. default is 12,14,16,18.')
parser.add_argument('--wd',
type=float,
default=0.00001,
help='weight decay rate. default is 0.00001.')
parser.add_argument('--iteration-per-epoch',
type=int,
default=200,
help='Number of iteration per epoch. default=200.')
opt = parser.parse_args()
if opt.logfile.lower() != 'none':
logging.basicConfig(filename=append_postfix(opt.logfile,
opt.log_postfix),
level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler())
return opt
