def main(args):
reset_seed(args.seed)
prepare_logger(args)
logger.info("These are the hyper-parameters you want to tune:\n%s", pprint.pformat(vars(args)))
device = 'cuda' if torch.cuda.is_available() else 'cpu'
train_loader, test_loader = data_preprocess(args)
model = models.__dict__[args.model]()
model.to(device)
criterion = nn.CrossEntropyLoss()
if args.optimizer == 'adam':
optimizer = optim.Adam(model.parameters(), lr=args.initial_lr, weight_decay=args.weight_decay)
else:
if args.optimizer == 'sgd':
optimizer_cls = optim.SGD
elif args.optimizer == 'rmsprop':
optimizer_cls = optim.RMSprop
optimizer = optimizer_cls(model.parameters(), lr=args.initial_lr, momentum=args.momentum, weight_decay=args.weight_decay)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs, eta_min=args.ending_lr)
for epoch in range(1, args.epochs + 1):
train(model, train_loader, criterion, optimizer, scheduler, args, epoch, device)
top1, _ = test(model, test_loader, criterion, args, epoch, device)
logger.info("Final accuracy is: %.6f", top1)
def main(args):
reset_seed(args.seed)
prepare_logger(args)
logger.info("These are the hyper-parameters you want to tune:\n%s",
pprint.pformat(vars(args)))
device = 'cuda' if torch.cuda.is_available() else 'cpu'
train_loader, test_loader = data_preprocess(args)
# model = models.__dict__[args.model](num_classes=10)
model = CNN(32, 3, args.channels, 10, args.layers)
model.to(device)
criterion = nn.CrossEntropyLoss()
if args.optimizer == 'adam':
optimizer = optim.Adam(model.parameters(),
lr=args.initial_lr,
weight_decay=args.weight_decay)
else:
if args.optimizer == 'sgd':
optimizer_cls = optim.SGD
elif args.optimizer == 'rmsprop':
optimizer_cls = optim.RMSprop
optimizer = optimizer_cls(model.parameters(),
lr=args.initial_lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.lr_scheduler == 'cosin':
scheduler = optim.lr_scheduler.CosineAnnealingLR(
optimizer, args.epochs, eta_min=args.ending_lr)
elif args.lr_scheduler == 'linear':
scheduler = optim.lr_scheduler.StepLR(optimizer,
step_size=15,
gamma=0.1)
trainer = DartsTrainer(
model,
loss=criterion,
metrics=lambda output, target: accuracy(output, target),
optimizer=optimizer,
num_epochs=args.epochs,
dataset_train=train_loader,
dataset_valid=test_loader,
batch_size=args.batch_size,
log_frequency=args.log_frequency,
unrolled=args.unrolled,
callbacks=[
LRSchedulerCallback(scheduler),
ArchitectureCheckpoint("./checkpoints_layer5")
])
if args.visualization:
trainer.enable_visualization()
trainer.train()
def main(args):
args = nni.get_next_parameter()
reset_seed(args['seed'])
prepare_logger(args)
logger.info("These are the hyper-parameters you want to tune:\n%s",
pprint.pformat(args))
device = 'cuda' if torch.cuda.is_available() else 'cpu'
train_loader, test_loader = data_preprocess(args)
model = models.__dict__[args['model']]()
model.to(device)
criterion = nn.CrossEntropyLoss()
if args['optimizer'] == 'adam':
optimizer = optim.Adam(model.parameters(),
lr=args['initial_lr'],
weight_decay=args['weight_decay'])
else:
if args['optimizer'] == 'sgd':
optimizer_cls = optim.SGD
elif args['optimizer'] == 'rmsprop':
optimizer_cls = optim.RMSprop
optimizer = optimizer_cls(model.parameters(),
lr=args['initial_lr'],
momentum=args['momentum'],
weight_decay=args['weight_decay'])
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer,
args['epochs'],
eta_min=args['ending_lr'])
for epoch in range(1, args['epochs'] + 1):
train(model, train_loader, criterion, optimizer, scheduler, args,
epoch, device)
top1, _ = test(model, test_loader, criterion, args, epoch, device)
nni.report_intermediate_result(_)
nni.report_final_result(top1)
logger.info("Final accuracy is: %.6f", top1)
def main():
global mainlogger
global bm
global lm
global fm
loggerlevel = 10
refine = False
if (len(os.sys.argv) > 1):
loggerlevel = os.sys.argv[1]
if (len(os.sys.argv) > 2):
refine = os.sys.argv[2]
mainlogger = utils.prepare_logger(__name__, int(loggerlevel))
mainlogger.info("Logger ready")
mainlogger.info("Building up bluetooth connection")
bm = BluetoothManager.BlueManager(mainlogger.getChild("BluetoothManager"))
lm = LedManager.LedManager(mainlogger.getChild("LedManager"))
fm = FDCManager.FDCManager(mainlogger.getChild("FDCManager"), refine)
socket = bm.build_socket()
client = None
try:
while 1:
client = connectionloop(socket)
try:
dataloop(client)
except Exception as e:
if e == UnboundLocalError:
mainlogger.error(traceback.format_exc())
elif e == ConnectionResetError:
mainlogger.warning("Connection reset")
mainlogger.warning("Connection reset error")
continue
elif e == Exception:
print(e)
mainlogger.error(traceback.format_exc())
mainlogger.error("Unknown/handled exception")
break
else:
mainlogger.warning("Error unhandeld. Doing nothing")
mainlogger.warning(traceback.format_exc())
finally:
lm.colorwipe()
bm.clean_up(client, socket)
def main(args):
reset_seed(args.seed)
prepare_logger(args)
logger.info("These are the hyper-parameters you want to tune:\n%s", pprint.pformat(vars(args)))
device = 'cuda' if torch.cuda.is_available() else 'cpu'
train_loader, test_loader = data_preprocess(args)
# model = models.__dict__[args.model]()
if args.model == 'vgg16':
model = VGG('VGG16')
elif args.model == 'vgg19':
model = VGG('VGG19')
elif args.model == 'resnet18':
model = ResNet18()
elif args.model == 'resnet34':
model = ResNet34()
elif args.model == 'senet18':
model = SENet18()
elif args.model == 'densenet121':
model = densenet_cifar()
# if args['model'] == 'resnet50':
# model = models.resnet50(pretrained=True)
# if args['model'] == 'googlenet':
# model = GoogLeNet()
# if args['model'] == 'mobilenet':
# model = MobileNet()
# if args['model'] == 'dpn92':
# model = DPN92()
# if args['model'] == 'shufflenetg2':
# model = ShuffleNetG2()
model.to(device)
criterion = nn.CrossEntropyLoss()
if args.optimizer == 'SGD':
optimizer = optim.SGD(model.parameters(), lr=args.initial_lr, momentum=args.momentum, weight_decay=args.weight_decay)
elif args.optimizer == 'Adadelta':
optimizer = optim.Adadelta(model.parameters(), lr=args.initial_lr)
elif args.optimizer == 'Adagrad':
optimizer = optim.Adagrad(model.parameters(), lr=args.initial_lr)
elif args.optimizer == 'Adam':
optimizer = optim.Adam(model.parameters(), lr=args.initial_lr)
elif args.optimizer == 'Adamax':
optimizer = optim.Adamax(model.parameters(), lr=args.initial_lr)
elif args.optimizer == 'RMSprop':
optimizer = optim.RMSprop(model.parameters(), lr=args.initial_lr, momentum=args.momentum, weight_decay=args.weight_decay)
# if args.optimizer == 'adam':
# optimizer = optim.Adam(model.parameters(), lr=args.initial_lr, weight_decay=args.weight_decay)
# else:
# if args.optimizer == 'sgd':
# optimizer_cls = optim.SGD
# elif args.optimizer == 'rmsprop':
# optimizer_cls = optim.RMSprop
# optimizer = optimizer_cls(model.parameters(), lr=args.initial_lr, momentum=args.momentum, weight_decay=args.weight_decay)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs, eta_min=args.ending_lr)
best_top1 = 0
for epoch in range(1, args.epochs + 1):
train(model, train_loader, criterion, optimizer, scheduler, args, epoch, device)
top1, _ = test(model, test_loader, criterion, args, epoch, device)
if top1 > best_top1:
best_top1 = top1
nni.report_intermediate_result(top1)
logger.info("Final accuracy is: %.6f", top1)
nni.report_final_result(best_top1)
def main(args):
reset_seed(args.seed)
prepare_logger(args)
logger.info("These are the hyper-parameters you want to tune:\n%s",
pprint.pformat(vars(args)))
if args.model == 'nas':
logger.info("Using NAS.\n")
if args.fix_arch:
if not os.path.exists(args.arc_checkpoint):
print(args.arc_checkpoint,
'does not exist, don not fix archetect')
args.fix_arch = False
device = 'cuda' if torch.cuda.is_available() else 'cpu'
if args.model == 'nas':
if not args.fix_arch:
model = CNN(32, 3, args.channels, 10, args.layers)
trainset, testset = data_preprocess(args)
else:
model = CNN(32, 3, args.channels, 10, args.layers)
apply_fixed_architecture(model, args.arc_checkpoint)
model.to(device)
train_loader, test_loader = data_preprocess(args)
else:
train_loader, test_loader = data_preprocess(args)
model = models.__dict__[args.model]()
model.to(device)
criterion = nn.CrossEntropyLoss()
if args.optimizer == 'adam':
optimizer = optim.Adam(model.parameters(),
lr=args.initial_lr,
weight_decay=args.weight_decay)
else:
if args.optimizer == 'sgd':
optimizer_cls = optim.SGD
elif args.optimizer == 'rmsprop':
optimizer_cls = optim.RMSprop
optimizer = optimizer_cls(model.parameters(),
lr=args.initial_lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer,
args.epochs,
eta_min=args.ending_lr)
if args.model == 'nas' and not args.fix_arch:
trainer = DartsTrainer(model,
loss=criterion,
metrics=lambda output, target: accuracyTopk(
output, target, topk=(1, )),
optimizer=optimizer,
num_epochs=args.epochs,
dataset_train=trainset,
dataset_valid=testset,
batch_size=args.batch_size,
log_frequency=args.log_frequency,
unrolled=args.unrolled,
callbacks=[
LRSchedulerCallback(scheduler),
ArchitectureCheckpoint("./checkpoints")
])
if args.visualization:
trainer.enable_visualization()
trainer.train()
trainer.export("final_arch.json")
else:
for epoch in range(1, args.epochs + 1):
train(model, train_loader, criterion, optimizer, scheduler, args,
epoch, device)
top1, _ = test(model, test_loader, criterion, args, epoch, device)
nni.report_intermediate_result(top1)
logger.info("Final accuracy is: %.6f", top1)
nni.report_final_result(top1)