def __init__(self,
datamanager,
model,
optimizer,
scheduler=None,
use_cpu=False,
label_smooth=True):
super(ImageSphereEngine, self).__init__(datamanager, model, optimizer,
scheduler, use_cpu)
self.criterion = AngleLoss(num_classes=self.datamanager.num_train_pids,
use_gpu=self.use_gpu,
label_smooth=label_smooth)
def main(args):
args = parser.parse_args(args)
#global best_rank1
best_rank1 = -np.inf
torch.manual_seed(args.seed)
# np.random.seed(args.seed)
# random.seed(args.seed)
if not args.use_avai_gpus:
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:
test_dir = args.save_dir
if args.save_dir == 'log':
if args.resume:
test_dir = os.path.dirname(args.resume)
else:
test_dir = os.path.dirname(args.load_weights)
sys.stdout = Logger(osp.join(test_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_imgreid_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.Random2DTranslation(args.height, args.width),
#T.Resize((args.height, args.width)),
T.RandomSizedEarser(),
T.RandomHorizontalFlip_custom(),
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,
return_path=args.draw_tsne),
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,
return_path=args.draw_tsne),
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', 'angular'} if args.use_angular else {'xent'},
use_gpu=use_gpu)
print("Model size: {:.3f} M".format(count_num_param(model)))
if not (args.use_angular):
if args.label_smooth:
print("Using Label Smoothing")
criterion = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
else:
print("Using Normal Cross-Entropy")
criterion = nn.CrossEntropyLoss()
if args.jsd:
print("Using JSD regularizer")
criterion = (criterion, JSD_loss(dataset.num_train_pids))
else:
if args.confidence_penalty:
print("Using Confidence Penalty")
criterion = (criterion, ConfidencePenalty())
else:
if args.label_smooth:
print("Using Angular Label Smoothing")
criterion = AngularLabelSmooth(num_classes=dataset.num_train_pids,
use_gpu=use_gpu)
else:
print("Using Angular Loss")
criterion = AngleLoss()
optimizer = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
if args.scheduler:
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma)
if args.fixbase_epoch > 0:
if hasattr(model, 'classifier') and isinstance(model.classifier,
nn.Module):
optimizer_tmp = init_optim(args.optim,
model.classifier.parameters(),
args.fixbase_lr, args.weight_decay)
else:
print(
"Warn: model has no attribute 'classifier' and fixbase_epoch is reset to 0"
)
args.fixbase_epoch = 0
if args.load_weights and check_isfile(args.load_weights):
# load pretrained weights but ignore layers that don't match in size
checkpoint = torch.load(args.load_weights)
pretrain_dict = checkpoint['state_dict']
model_dict = model.state_dict()
pretrain_dict = {
k: v
for k, v in pretrain_dict.items()
if k in model_dict and model_dict[k].size() == v.size()
}
model_dict.update(pretrain_dict)
model.load_state_dict(model_dict)
print("Loaded pretrained weights from '{}'".format(args.load_weights))
if args.resume and check_isfile(args.resume):
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
args.start_epoch = checkpoint['epoch'] + 1
best_rank1 = checkpoint['rank1']
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch,
best_rank1))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.single_folder != '':
extract_features(model,
use_gpu,
args,
transform_test,
return_distmat=False)
return
if args.evaluate:
print("Evaluate only")
test_dir = args.save_dir
if args.save_dir == 'log':
if args.resume:
test_dir = os.path.dirname(args.resume)
else:
test_dir = os.path.dirname(args.load_weights)
distmat = test(model,
queryloader,
galleryloader,
use_gpu,
args,
writer=None,
epoch=-1,
return_distmat=True,
draw_tsne=args.draw_tsne,
tsne_clusters=args.tsne_labels)
if args.visualize_ranks:
visualize_ranked_results(
distmat,
dataset,
save_dir=osp.join(test_dir, 'ranked_results'),
topk=10,
)
return
writer = SummaryWriter(log_dir=osp.join(args.save_dir, 'tensorboard'))
start_time = time.time()
train_time = 0
best_epoch = args.start_epoch
print("==> Start training")
if args.fixbase_epoch > 0:
print(
"Train classifier for {} epochs while keeping base network frozen".
format(args.fixbase_epoch))
for epoch in range(args.fixbase_epoch):
start_train_time = time.time()
train(epoch,
model,
criterion,
optimizer_tmp,
trainloader,
use_gpu,
writer,
args,
freeze_bn=True)
train_time += round(time.time() - start_train_time)
del optimizer_tmp
print("Now open all layers for training")
best_epoch = 0
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion, optimizer, trainloader, use_gpu, writer,
args)
train_time += round(time.time() - start_train_time)
if args.scheduler:
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:
if (epoch + 1) == args.max_epoch:
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint(
{
'state_dict': state_dict,
'rank1': -1,
'epoch': epoch,
}, False,
osp.join(
args.save_dir, 'beforeTesting_checkpoint_ep' +
str(epoch + 1) + '.pth.tar'))
print("==> Test")
rank1 = test(model,
queryloader,
galleryloader,
use_gpu,
args,
writer=writer,
epoch=epoch)
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))
return best_rank1, best_epoch