correct += predicted.eq(targets.data).cpu().sum()
test_acc = 100. * correct / total
print('* Test results : Acc@1 = %.2f%%' % (test_acc))
record.write('\nTest Acc: %f\n' % test_acc)
record.flush()
record = open('./checkpoint/' + args.id + '.txt', 'w')
record.write('learning rate: %f\n' % args.lr)
record.write('batch size: %f\n' % args.batch_size)
record.write('start iter: %d\n' % args.start_iter)
record.write('mid iter: %d\n' % args.mid_iter)
record.flush()
#
loader = cifar_dataloader(args.dataset,r=args.noise_rate,noise_mode='instance',batch_size=args.batch_size,num_workers=5,\
root_dir='./')
val_loader = loader.run('eval_train')
test_loader = loader.run('test')
train_loader = loader.run('warmup')
#loader = dataloader.clothing_dataloader(batch_size=args.batch_size,num_workers=5,shuffle=True)
#train_loader,val_loader,test_loader = loader.run()
best = 0
init = True
# Model
print('\nModel setup')
print('| Building net')
##
##
model = ResNet32(num_classes=args.num_class)
# from resnet_imselar import resnet18
# model = resnet18(num_classes=args.num_class)
model = model.cuda()
return model
stats_log=open('./checkpoint/%s_%.1f_%s_%s'%(args.dataset,args.r,args.noise_mode,args.sess)+'_stats.txt','w')
test_log=open('./checkpoint/%s_%.1f_%s_%s'%(args.dataset,args.r,args.noise_mode,args.sess)+'_acc.txt','w')
if args.dataset=='cifar10':
warm_up = 10
elif args.dataset=='cifar100':
warm_up = 30
loader = dataloader.cifar_dataloader(args.dataset,r=args.r,noise_mode=args.noise_mode,batch_size=args.batch_size,num_workers=5,\
root_dir=args.data_path,log=stats_log,noise_file='%s/%.1f_%s.json'%(args.data_path,args.r,args.noise_mode))
print('| Building net')
net1 = create_model()
net2 = create_model()
cudnn.benchmark = True
criterion = SemiLoss()
# optimizer1 = optim.SGD(net1.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
# optimizer2 = optim.SGD(net2.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
optimizer1 = optim.SGD(net1.params(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
optimizer2 = optim.SGD(net2.params(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
CE = nn.CrossEntropyLoss(reduction='none')
CEloss = nn.CrossEntropyLoss()
if args.noise_mode=='asym':
return torch.mean(torch.sum(probs.log() * probs, dim=1))
def create_model(devices=[0]):
model = ResNet18(num_classes=args.num_class)
model = model.cuda()
model = torch.nn.DataParallel(model, device_ids=devices).cuda()
return model
loader = dataloader.cifar_dataloader(
dataset=args.dataset,
r=args.r,
noise_mode=args.noise_mode,
batch_size=args.batch_size,
warmup_batch_size=args.warmup_batch_size,
num_workers=args.num_workers,
root_dir=args.data_path,
noise_file=f"{args.checkpoint_path}/saved/labels.json",
preaug_file=(
f"{args.checkpoint_path}/saved/{args.preset}_preaugdata.pth.tar"
if args.preaugment
else ""
),
augmentation_strategy=args,
)
print("| Building net")
devices = range(torch.cuda.device_count())
net1 = create_model(devices)
net2 = create_model(devices)
cudnn.benchmark = True
criterion = SemiLoss()