netF = ResClassifier(class_num=args.class_num, extract=args.extract).cuda()
netF.apply(weights_init)
def get_L2norm_loss_self_driven(x):
l = (x.norm(p=2, dim=1).mean() - args.radius)**2
return args.weight_ring * l
def get_cls_loss(pred, gt):
cls_loss = F.nll_loss(F.log_softmax(pred), gt)
return cls_loss
opt_g = optim.SGD(netG.parameters(), lr=args.lr, weight_decay=0.0005)
opt_f = optim.SGD(netF.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=0.0005)
for epoch in range(1, args.pre_epoches + 1):
for i, (s_imgs, s_labels) in tqdm.tqdm(enumerate(source_loader)):
s_imgs = Variable(s_imgs.cuda())
s_labels = Variable(s_labels.cuda())
opt_g.zero_grad()
opt_f.zero_grad()
s_bottleneck = netG(s_imgs)
s_fc2_emb, s_logit = netF(s_bottleneck)
netG = ResBase50().cuda()
netF = ResClassifier(class_num=args.class_num, extract=args.extract, dropout_p=args.dropout_p).cuda()
netF.apply(weights_init)
def get_cls_loss(pred, gt):
cls_loss = F.nll_loss(F.log_softmax(pred), gt)
return cls_loss
def get_L2norm_loss_self_driven(x):
l = (x.norm(p=2, dim=1).mean() - args.radius) ** 2
return args.weight_L2norm * l
opt_g = optim.SGD(netG.parameters(), lr=args.lr, weight_decay=0.0005)
opt_f = optim.SGD(netF.parameters(), lr=args.lr, momentum=0.9, weight_decay=0.0005)
for epoch in range(1, args.pre_epoches + 1):
for i, (s_imgs, s_labels) in tqdm.tqdm(enumerate(source_loader)):
if s_imgs.size(0) != args.batch_size:
continue
s_imgs = Variable(s_imgs.cuda())
s_labels = Variable(s_labels.cuda())
opt_g.zero_grad()
opt_f.zero_grad()
s_bottleneck = netG(s_imgs)
s_fc2_emb, s_logit = netF(s_bottleneck)