def load(args):
add_args(args, sub_args)
net = LeNetCaffe()
net.build_gate(BBDropout)
net.build_gate(DBBDropout, argdicts={'kl_scale': args.kl_scale}, dep=True)
train_loader, test_loader = get_MNIST(args.batch_size)
base_params = []
dgate_params = []
for name, param in net.named_parameters():
if 'dgate' in name:
dgate_params.append(param)
elif 'base' in name:
base_params.append(param)
optimizer = optim.Adam([{
'params': dgate_params,
'lr': 1e-2
}, {
'params': base_params,
'lr': 1e-3,
'weight_decay': 1e-4
}])
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=[int(r * args.num_epochs) for r in [.5, .8]],
gamma=0.1)
return net, train_loader, test_loader, optimizer, scheduler
def load(args):
add_args(args, sub_args)
net = VGG(10)
net.build_gate(VIB)
train_loader, test_loader = get_CIFAR10(args.batch_size)
base_params = []
gate_params = []
for name, param in net.named_parameters():
if 'gate' in name:
gate_params.append(param)
else:
base_params.append(param)
optimizer = optim.Adam([{
'params': gate_params,
'lr': 1e-2
}, {
'params': base_params,
'lr': 1e-3,
'weight_decay': 1e-4
}])
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=[int(r * args.num_epochs) for r in [.5, .8]],
gamma=0.1)
return net, train_loader, test_loader, optimizer, scheduler
def load(args):
add_args(args, sub_args)
args.testfile = os.path.join(
benchmarks_path,
'kkanji_10_100_4.tar' if args.testfile is None else args.testfile)
args.clusterfile = os.path.join(
benchmarks_path, 'kkanji_10_400_12.tar'
if args.clusterfile is None else args.clusterfile)
return Model(args)
def load(args):
add_args(args, sub_args)
if args.testfile is None:
args.testfile = os.path.join(
os.path.join(benchmarks_path),
'emnist_phase{}_10_1000_4.tar'.format(args.phase))
return Model(args)
def load(args):
add_args(args, sub_args)
net = VGG(100)
train_loader, test_loader = get_CIFAR100(args.batch_size)
optimizer = optim.Adam(net.parameters(), lr=1e-3, weight_decay=1e-4)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=[int(r * args.num_epochs) for r in [0.5, 0.8]],
gamma=0.1)
return net, train_loader, test_loader, optimizer, scheduler
def load(args):
add_args(args, sub_args)
lamb = args.lambN / 60000.0
net = LeNetMLP()
net.build_gate(L0Reg, {'weight_decay':1e-4, 'lamb':lamb})
train_loader, test_loader = get_MNIST(args.batch_size)
optimizer = optim.Adam(net.parameters(), lr=1e-3)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=[int(r*args.num_epochs) for r in [.5, .8]],
gamma=0.1)
return net, train_loader, test_loader, optimizer, scheduler
def load(args):
add_args(args, sub_args)
if args.novel:
args.testfile = os.path.join(
benchmarks_path, 'emnist_paired_novel.tar'
if args.testfile is None else args.testfile)
else:
args.testfile = os.path.join(
benchmarks_path,
'emnist_paired.tar' if args.testfile is None else args.testfile)
return Model(args)
def load(args):
add_args(args, sub_args)
if args.novel:
args.testfile = os.path.join(benchmarks_path,
'mini_imagenet_novel_10_200_4.tar' if args.testfile is None else args.testfile)
args.clusterfile = os.path.join(benchmarks_path,
'mini_imagenet_novel_10_600_12.tar' if args.clusterfile is None else args.clusterfile)
else:
args.testfile = os.path.join(benchmarks_path,
'mini_imagenet_10_200_4.tar' if args.testfile is None else args.testfile)
args.clusterfile = os.path.join(benchmarks_path,
'mini_imagenet_10_600_12.tar' if args.clusterfile is None else args.clusterfile)
return Model(args)
def load(args):
add_args(args, sub_args)
net = VGG(100)
net.build_gate(L0Reg, {
'weight_decay': 1e-4,
'lamb': args.lamb,
'droprate_init': 0.2
})
train_loader, test_loader = get_CIFAR100(args.batch_size)
optimizer = optim.Adam(net.parameters(), lr=1e-3)
scheduler = optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=[int(r * args.num_epochs) for r in [.5, .8]],
gamma=0.1)
return net, train_loader, test_loader, optimizer, scheduler
def load(args):
add_args(args, sub_args)
return Model(args)
def load(args):
add_args(args, sub_args)
return AnchoredModel(args)
