def main():
parser = argparse.ArgumentParser()
parser.add_argument('--batch_size', type=int, default=256)
parser.add_argument('--loss', default='softmax_loss',
choices=['softmax_loss', 'center_loss', 'sphere_face_loss', 'cos_face_loss', 'arc_face_loss'])
parser.add_argument('--viz', default='vizs')
parser.add_argument('--epochs', type=int, default=30)
parser.add_argument('--lr', type=float, default=0.001)
args = parser.parse_args()
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 8, 'pin_memory': True} if use_cuda else {}
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])
train_loader = DataLoader(
datasets.MNIST('../data', train=True, download=True, transform=transform),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = DataLoader(
datasets.MNIST('../data', train=False, transform=transform),
batch_size=512, shuffle=True, **kwargs)
model = Net().to(device)
if args.loss == 'center_loss':
criterion = CenterLoss().to(device)
center_optimizer = optim.SGD([criterion.centers], lr=args.lr, momentum=0.9)
elif args.loss == 'sphere_face_loss':
criterion = SphereFaceLoss().to(device)
elif args.loss == 'cos_face_loss':
criterion = CosFaceLoss(s=7, m=0.2).to(device)
elif args.loss == 'softmax_loss':
criterion = SoftmaxLoss().to(device)
elif args.loss == 'arc_face_loss':
criterion = ArcFaceLoss().to(device)
optimizer = optim.SGD([{'params': model.parameters()}, {'params': criterion.fc.parameters()}],
lr=args.lr, momentum=0.9)
for epoch in range(1, args.epochs + 1):
model.train()
embeddings = []
labels = []
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
if args.loss == 'center_loss':
center_optimizer.zero_grad()
embedding = model(data)
loss = criterion(embedding, target)
loss.backward()
optimizer.step()
if args.loss == 'center_loss':
center_optimizer.step()
embeddings.append(embedding)
labels.append(target)
if batch_idx % 100 == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
embeddings = torch.cat(embeddings, 0).cpu().detach().numpy()
labels = torch.cat(labels, 0).cpu().detach().numpy()
acc = val(model, criterion, device, test_loader)
visualize(args.viz, args.loss, embeddings, labels, epoch, acc)
print('Creating gif...')
create_gif('./%s/gifs/%s.gif' % (args.viz, args.loss),
'./%s/%s' % (args.viz, args.loss), 0.2)
print('Done')
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset)
transform_train = T.Compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
trainloader = DataLoader(
ImageDataset(dataset.train, transform=transform_train),
batch_size=args.train_batch,
shuffle=True,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
ImageDataset(dataset.query, transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
ImageDataset(dataset.gallery, transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
loss={'cent'})
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
criterion_cent = CenterLoss(num_classes=dataset.num_train_pids,
feat_dim=model.feat_dim,
use_gpu=use_gpu)
optimizer_model = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
optimizer_cent = torch.optim.SGD(criterion_cent.parameters(),
lr=args.lr_cent)
if args.stepsize > 0:
scheduler = lr_scheduler.StepLR(optimizer_model,
step_size=args.stepsize,
gamma=args.gamma)
start_epoch = args.start_epoch
if args.resume:
print("Loading checkpoint from '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
start_epoch = checkpoint['epoch']
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
test(model, queryloader, galleryloader, use_gpu)
return
start_time = time.time()
best_rank1 = -np.inf
for epoch in range(start_epoch, args.max_epoch):
print("==> Epoch {}/{}".format(epoch + 1, args.max_epoch))
train(model, criterion_xent, criterion_cent, optimizer_model,
optimizer_cent, trainloader, use_gpu)
if args.stepsize > 0: scheduler.step()
if args.eval_step > 0 and (epoch + 1) % args.eval_step == 0 or (
epoch + 1) == args.max_epoch:
print("==> Test")
rank1 = test(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best: best_rank1 = rank1
save_checkpoint(
{
'state_dict': model.state_dict(),
'rank1': rank1,
'epoch': epoch,
}, is_best,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
print("Finished. Total elapsed time (h:m:s): {}".format(elapsed))
def train_epoch(train_loader,
model,
loss_fn,
optimzer,
k,
n_K,
n_classes,
rm_zero,
global_loss=False,
train_embeddeds=None,
train_targets=None,
gamma=0.01,
center_sigma=-1.0,
method='kTriplet',
use_cross_entropy=False):
model.train()
losses_triplet = []
losses_inner_class = []
losses_center = []
losses_cross_entropy = []
losses = []
center_loss = CenterLoss(n_classes, mean_center=True).cuda()
cross_entropy = nn.CrossEntropyLoss().cuda()
if global_loss:
in_mat, out_mat = get_global_distance_map(train_embeddeds,
train_targets, k, n_K)
data_id = 0
for batch_idx, (data, target) in enumerate(
tqdm(train_loader, ncols=70, desc="Train")):
data = data.cuda()
target = target.cuda()
# !note, for Triplet, here just the embeddings, for Clas, it's predict
outputs = model(data)
embeddeds = outputs
if method == 'kTriplet':
anchor, positive, negative = generate_k_triplet(embeddeds,
target,
K=k,
B=n_K)
elif method == 'batchHardTriplet':
anchor, positive, negative = generate_batch_hard_triplet(
embeddeds, target)
elif method == 'batchAllTriplet' or method == 'batchSemiHardTriplet':
anchor, positive, negative = generate_all_triplet(
embeddeds, target)
else:
raise ValueError
triplet_loss = loss_fn(anchor, positive, negative)
if method == 'batchSemiHardTriplet':
semi = torch.nonzero((triplet_loss <= loss_fn.margin)
& (triplet_loss > 0))
triplet_loss.index_select(dim=0, index=semi.squeeze())
if rm_zero:
non_zero = torch.nonzero(triplet_loss.cpu().data).size(0)
if non_zero == 0:
loss_triplet = triplet_loss.mean()
else:
loss_triplet = (triplet_loss / non_zero).sum()
else:
loss_triplet = triplet_loss.mean()
if gamma > 0:
loss_inner_class = torch.log1p(
(anchor - positive).pow(2).sum(1)).mean()
loss = loss_triplet + gamma * loss_inner_class
elif gamma == 0:
loss_inner_class = np.mean(
np.sum(np.power(
np.log1p((anchor.cpu().detach().numpy() -
positive.cpu().detach().numpy())), 2),
axis=1))
loss_inner_class = torch.tensor(loss_inner_class)
loss = loss_triplet
else:
loss_inner_class = torch.tensor(0)
loss = loss_triplet
if center_sigma > 0:
closs = center_sigma * center_loss(embeddeds, target)
loss += closs
losses_center.append(closs.item())
if global_loss:
g_pos, g_neg = get_global_data(
train_loader.dataset, in_mat, out_mat,
np.array(range(data_id, data_id + train_loader.batch_size)))
data_id += train_loader.batch_size
g_pos = g_pos.cuda()
g_neg = g_neg.cuda()
out_p = model(g_pos)
out_n = model(g_neg)
anchor, positive, negative = makeup_global_triplet(
embeddeds, out_p, out_n, k, n_K)
global_loss = loss_fn(anchor, positive, negative)
loss += global_loss
# TODO
# add cross entropy
if use_cross_entropy:
#ce_loss = cross_entropy(, target)
#loss = loss + ce_loss
print('NOt Implement!')
sys.exit(-1)
pass
else:
ce_loss = torch.tensor(0)
optimzer.zero_grad()
loss.backward()
optimzer.step()
losses_triplet.append(loss_triplet.item())
losses_inner_class.append(loss_inner_class.item())
losses_cross_entropy.append(ce_loss.item())
losses.append(loss.item())
average_triplet_loss = sum(losses_triplet) / len(train_loader)
average_inner_class = sum(losses_inner_class) / len(train_loader)
average_center_loss = sum(losses_center) / len(train_loader)
average_cross_entropy = sum(losses_cross_entropy) / len(train_loader)
total_loss = sum(losses) / len(train_loader)
return total_loss, average_triplet_loss, average_inner_class, average_center_loss, average_cross_entropy
def main():
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_img_dataset(
root=args.root,
name=args.dataset,
split_id=args.split_id,
cuhk03_labeled=args.cuhk03_labeled,
cuhk03_classic_split=args.cuhk03_classic_split,
)
transform_train = T.Compose([
T.Resize((args.height, args.width)),
T.RandomHorizontalFlip(p=0.5),
T.Pad(10),
T.RandomCrop([args.height, args.width]),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
torchvision.transforms.RandomErasing(p=0.5,
scale=(0.02, 0.4),
ratio=(0.3, 3.33),
value=(0.4914, 0.4822, 0.4465))
# T.RandomErasing(probability=0.5, sh=0.4, mean=(0.4914, 0.4822, 0.4465)),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = True if use_gpu else False
trainloader = DataLoader(
ImageDataset(dataset.train, transform=transform_train),
sampler=RandomIdentitySampler2(dataset.train,
batch_size=args.train_batch,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
ImageDataset(dataset.query, transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
ImageDataset(dataset.gallery, transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
loss={'xent', 'htri', 'cent'})
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
criterion_htri = TripletLoss(margin=args.margin)
criterion_cent = CenterLoss(num_classes=dataset.num_train_pids,
feat_dim=model.feat_dim,
use_gpu=use_gpu)
optimizer_model = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
optimizer_cent = torch.optim.SGD(criterion_cent.parameters(),
lr=args.lr_cent)
'''only the optimizer_model use learning rate schedule'''
# if args.stepsize > 0:
# scheduler = lr_scheduler.StepLR(optimizer_model, step_size=args.stepsize, gamma=args.gamma)
'''------Modify lr_schedule here------'''
current_schedule = init_lr_schedule(schedule=args.schedule,
warm_up_epoch=args.warm_up_epoch,
half_cos_period=args.half_cos_period,
lr_milestone=args.lr_milestone,
gamma=args.gamma,
stepsize=args.stepsize)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer_model,
lr_lambda=current_schedule)
'''------Please refer to the args.xxx for details of hyperparams------'''
#embed()
start_epoch = args.start_epoch
if args.resume:
print("Loading checkpoint from '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
start_epoch = checkpoint['epoch']
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
test(model, queryloader, galleryloader, use_gpu)
return
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
for epoch in range(start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, criterion_cent,
optimizer_model, optimizer_cent, trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
if args.schedule: scheduler.step()
if (epoch + 1) > args.start_eval and args.eval_step > 0 and (
epoch + 1) % args.eval_step == 0 or (epoch +
1) == args.max_epoch:
print("==> Test")
rank1 = test(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, is_best,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(
best_rank1, best_epoch))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
train_time = str(datetime.timedelta(seconds=train_time))
print(
"Finished. Total elapsed time (h:m:s): {}. Training time (h:m:s): {}.".
format(elapsed, train_time))
fc = torch.nn.Linear(2, 2).cuda()
layer = Membership_norm(2, 4,
init_c=-5 * torch.ones((2, 4), dtype=torch.float),
init_lamda=4 * torch.ones((2, 4), dtype=torch.float)).cuda()
# x = torch.tensor([[[0.9, 0.1], [0.9, 0.1]], [[-0.9, 0.1], [-0.1, -2.5]]], dtype=torch.float, requires_grad=True)
# x2 = x ** 2
# print(x2.requires_grad)
# print(x.shape)
# print(layer(x))
# print(layer.c)
# print(x.shape)
# loss_focal = torch.nn.MSELoss()
loss_focal = FocalLoss()
loss_center = CenterLoss()
para = [
{"params": fc.parameters(), "lr": 1e-3},
{"params": layer.c, "lr": 1e-3},
{"params": layer.lamda, "lr": 1e-3},
]
# optim = torch.optim.SGD(para)
optim = torch.optim.Adam(para)
# bestloss = 1e5
# bestnetweightfc = []
# bestnetweightlayer = []
for i in range(0, 100000):
h = fc(x_in_tensor).unsqueeze(2)
y = layer(h)
def __init__(self, config):
super(Trainer, self).__init__(config)
self.datamanager = DataManger(config["data"])
# model
self.model = Baseline(
num_classes=self.datamanager.datasource.get_num_classes("train")
)
# summary model
summary(
self.model,
input_size=(3, 256, 128),
batch_size=config["data"]["batch_size"],
device="cpu",
)
# losses
cfg_losses = config["losses"]
self.criterion = Softmax_Triplet_loss(
num_class=self.datamanager.datasource.get_num_classes("train"),
margin=cfg_losses["margin"],
epsilon=cfg_losses["epsilon"],
use_gpu=self.use_gpu,
)
self.center_loss = CenterLoss(
num_classes=self.datamanager.datasource.get_num_classes("train"),
feature_dim=2048,
use_gpu=self.use_gpu,
)
# optimizer
cfg_optimizer = config["optimizer"]
self.optimizer = torch.optim.Adam(
self.model.parameters(),
lr=cfg_optimizer["lr"],
weight_decay=cfg_optimizer["weight_decay"],
)
self.optimizer_centerloss = torch.optim.SGD(
self.center_loss.parameters(), lr=0.5
)
# learing rate scheduler
cfg_lr_scheduler = config["lr_scheduler"]
self.lr_scheduler = WarmupMultiStepLR(
self.optimizer,
milestones=cfg_lr_scheduler["steps"],
gamma=cfg_lr_scheduler["gamma"],
warmup_factor=cfg_lr_scheduler["factor"],
warmup_iters=cfg_lr_scheduler["iters"],
warmup_method=cfg_lr_scheduler["method"],
)
# track metric
self.train_metrics = MetricTracker("loss", "accuracy")
self.valid_metrics = MetricTracker("loss", "accuracy")
# save best accuracy for function _save_checkpoint
self.best_accuracy = None
# send model to device
self.model.to(self.device)
self.scaler = GradScaler()
# resume model from last checkpoint
if config["resume"] != "":
self._resume_checkpoint(config["resume"])
aux_loss = PerpetualOrthogonalProjectionLoss(
num_classes=100,
feat_dim=embedding_dim[args.net],
no_norm=False,
use_attention=False)
params = list(net.parameters()) + list(aux_loss.parameters())
else:
aux_loss = OrthogonalProjectionLoss(no_norm=False, use_attention=False)
if args.hnc:
hnc_loss = cam_loss_kd_topk()
else:
hnc_loss = None
if args.cl:
center_loss = CenterLoss(num_classes=100, feat_dim=2048, use_gpu=True)
params = list(net.parameters()) + list(center_loss.parameters())
else:
optimizer = optim.SGD(params=params,
lr=args.lr,
momentum=0.9,
weight_decay=5e-4)
train_scheduler = optim.lr_scheduler.MultiStepLR(
optimizer, milestones=settings.MILESTONES,
gamma=0.2) # learning rate decay
iter_per_epoch = len(training_loader)
warmup_scheduler = WarmUpLR(optimizer, iter_per_epoch * args.warm)
if args.resume:
if args.pth is not None: