def __init__(self, gallery_path=args.gallery_path, save_dir=args.save_dir):
print("==========\nArgs:{}\n==========".format(args))
self.gallery_path = gallery_path
self.save_dir=save_dir
print("gallery_path".format(gallery_path))
print("Initializing image data manager")
# ------------
self.galleryloader, self.gallery_datasets = self.create_galleryloader()
print("Initializing model: {}".format(args.arch))
self.model = models.init_model(name=args.arch, num_classes=args.num_train_pids, loss={'xent', 'htri'})
print("Model size: {:.3f} M".format(count_num_param(self.model)))
if args.load_weights and check_isfile(args.load_weights):
# load pretrained weights but ignore layers that don't match in size
map_location = 'cpu' if not use_gpu else None
checkpoint = torch.load(args.load_weights, map_location=map_location)
pretrain_dict = checkpoint['state_dict']
model_dict = self.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)
self.model.load_state_dict(model_dict)
print("Loaded pretrained weights from '{}'".format(args.load_weights))
if use_gpu:
self.model = nn.DataParallel(self.model).cuda()
self.gf = self.init_gallary_info()
def init_gallary_info(self):
batch_time = AverageMeter()
# init gallary info
self.model.eval()
with torch.no_grad():
gf, g_pids, g_camids = [], [], []
save_type = 'gpu' if use_gpu else 'cpu'
pkl_path = './data/market1501/market1501-%s.pkl' % save_type
if not check_isfile(pkl_path):
end = time.time()
for batch_idx, (imgs, _) in enumerate(self.galleryloader):
if use_gpu:
imgs = imgs.cuda()
end = time.time()
features = self.model(imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
gf.append(features)
gf = torch.cat(gf, 0)
# cache for CPU
with open(pkl_path, mode='wb') as fout:
pickle.dump(gf, fout)
else:
with open(pkl_path, mode='rb') as fin:
gf = pickle.load(fin)
print("Extracted features for gallery set, obtained {}-by-{} matrix".format(gf.size(0), gf.size(1)))
return gf
def resume(model):
if check_isfile(args.resume):
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
args.start_epoch = checkpoint['epoch']
rank1 = checkpoint['rank1']
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch, rank1))
def load_weights(model):
# load pretrained weights but ignore layers that don't match in size
if check_isfile(args.load_weights):
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))
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_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,
) #cuhk03_labeled: detected,labeled
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
#pdb.set_trace()
trainloader = DataLoader(
ImageDataset(dataset.train, transform=transform_train),
sampler=RandomIdentitySampler(dataset.train, args.train_batch,
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,
batchsize=args.test_batch,
loss={'xent', 'wcont', 'htri'})
print("Model size: {:.3f} M".format(count_num_param(model)))
criterion_xent = nn.CrossEntropyLoss()
criterion_htri = TripletLoss(margin=args.margin)
criterion_KA = KALoss(margin=args.margin,
same_margin=args.same_margin,
use_auto_samemargin=args.use_auto_samemargin)
cirterion_lifted = LiftedLoss(margin=args.margin)
cirterion_batri = BA_TripletLoss(margin=args.margin)
if args.use_auto_samemargin == True:
G_params = [{
'params': model.parameters(),
'lr': args.lr
}, {
'params': criterion_KA.auto_samemargin,
'lr': args.lr
}]
else:
G_params = [para for _, para in model.named_parameters()]
optimizer = init_optim(args.optim, G_params, args.lr, args.weight_decay)
if args.load_weights:
# load pretrained weights but ignore layers that don't match in size
if check_isfile(args.load_weights):
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:
if check_isfile(args.resume):
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
args.start_epoch = checkpoint['epoch']
rank1 = checkpoint['rank1']
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(
args.start_epoch, rank1))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
distmat = test(model,
queryloader,
galleryloader,
use_gpu,
return_distmat=True)
if args.vis_ranked_res:
visualize_ranked_results(
distmat,
dataset,
save_dir=osp.join(args.save_dir, 'ranked_results'),
topk=20,
)
return
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
adjust_learning_rate(optimizer, epoch)
train(epoch, model, cirterion_batri, cirterion_lifted, criterion_xent,
criterion_htri, criterion_KA, optimizer, trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
if (epoch + 1) > args.start_eval and args.eval_step > 0 and (
epoch + 1) % args.eval_step == 0 or (epoch +
1) == args.max_epoch:
rank1 = 0
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
print("==> Test")
sys.stdout.flush()
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("model saved")
print("==> Best Rank-1 {:.1%}, achieved at epoch {}".format(
best_rank1, best_epoch))
sys.stdout.flush()
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))
sys.stdout.flush()
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)
# print("Currently using GPU {}".format(args.gpu_devices))
# #cudnn.benchmark = False
# cudnn.deterministic = True
# torch.cuda.manual_seed_all(args.seed)
# torch.set_default_tensor_type('torch.DoubleTensor')
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(),
#T.Resize((args.height, args.width)),
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
if 'stanford' in args.dataset:
datasetLoader = ImageDataset_stanford
else:
datasetLoader = ImageDataset
if args.crop_img:
print("Using Cropped Images")
else:
print("NOT using cropped Images")
trainloader = DataLoader(
datasetLoader(dataset.train,
-1,
crop=args.crop_img,
transform=transform_train),
batch_size=args.train_batch,
shuffle=True,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
testloader = DataLoader(
datasetLoader(dataset.test,
-1,
crop=args.crop_img,
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', '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:
criterion = nn.CrossEntropyLoss()
else:
if args.label_smooth:
print("Using Label Smoothing")
criterion = AngularLabelSmooth(num_classes=dataset.num_train_pids,
use_gpu=use_gpu)
else:
criterion = AngleLoss()
optimizer = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
if args.scheduler != 0:
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,
list(model.classifier.parameters()) +
list(model.encoder.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.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,
testloader,
use_gpu,
args,
writer=None,
epoch=-1,
return_distmat=True,
draw_tsne=args.draw_tsne,
tsne_clusters=args.tsne_labels,
use_cosine=args.plot_deltaTheta)
if args.visualize_ranks:
visualize_ranked_results(
distmat,
dataset,
save_dir=osp.join(test_dir, 'ranked_results'),
topk=10,
)
if args.plot_deltaTheta:
plot_deltaTheta(distmat,
dataset,
save_dir=osp.join(test_dir, 'deltaTheta_results'),
min_rank=1)
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.test_rot:
print("Training only classifier for rotation")
model = models.init_model(name='rot_tester',
base_model=model,
inplanes=2048,
num_rot_classes=8)
criterion_rot = nn.CrossEntropyLoss()
optimizer_rot = init_optim(args.optim, model.fc_rot.parameters(),
args.fixbase_lr, args.weight_decay)
if use_gpu:
model = nn.DataParallel(model).cuda()
try:
best_epoch = 0
for epoch in range(0, args.max_epoch):
start_train_time = time.time()
train_rotTester(epoch, model, criterion_rot, optimizer_rot,
trainloader, use_gpu, writer, args)
train_time += round(time.time() - start_train_time)
if args.scheduler != 0:
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_rotTester(model,
criterion_rot,
queryloader,
galleryloader,
trainloader,
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 Cccuracy {:.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
except KeyboardInterrupt:
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, 'keyboardInterrupt_checkpoint_ep' +
str(epoch + 1) + '.pth.tar'))
return None, None
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 != 0:
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,
testloader,
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
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,
split_wild=args.split_wild)
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)))
use_autoTune = False
if not (args.use_angular):
if args.label_smooth:
print("Using Label Smoothing with epsilon", args.label_epsilon)
criterion = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids,
epsilon=args.label_epsilon,
use_gpu=use_gpu)
elif args.focal_loss:
print("Using Focal Loss with gamma=", args.focal_gamma)
criterion = FocalLoss(gamma=args.focal_gamma)
else:
print("Using Normal Cross-Entropy")
criterion = nn.CrossEntropyLoss()
if args.jsd:
print("Using JSD regularizer")
criterion = (criterion, JSD_loss(dataset.num_train_pids))
if args.auto_tune_mtl:
print("Using AutoTune")
use_autoTune = True
criterion = MultiHeadLossAutoTune(
list(criterion),
[args.lambda_xent, args.confidence_beta]).cuda()
else:
if args.confidence_penalty:
print("Using Confidence Penalty", args.confidence_beta)
criterion = (criterion, ConfidencePenalty())
if args.auto_tune_mtl and args.confidence_penalty:
print("Using AutoTune")
use_autoTune = True
criterion = MultiHeadLossAutoTune(
list(criterion),
[args.lambda_xent, -args.confidence_beta]).cuda()
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()
if use_autoTune:
optimizer = init_optim(
args.optim,
list(model.parameters()) + list(criterion.parameters()), args.lr,
args.weight_decay)
else:
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):
if use_autoTune:
optimizer_tmp = init_optim(
args.optim,
list(model.classifier.parameters()) +
list(criterion.parameters()), args.fixbase_lr,
args.weight_decay)
else:
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,
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.save_epoch > 0 and (epoch + 1) % args.save_epoch == 0) or
(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'))
is_best = False
rank1 = -1
if args.eval_step > 0:
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
val_set = ImageDataset(dataset.query + dataset.gallery, transform_test_flip)
val_loader = DataLoader(val_set,
batch_size=args.test_batch_size,
shuffle=False,
num_workers=args.workers,
collate_fn=val_collate_fn)
print(val_loader)
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
loss={'xent'},
use_gpu=use_gpu,
args=vars(args))
if args.load_weights and check_isfile(args.load_weights):
# load pretrained weights but ignore layers that don't match in size
try:
checkpoint = torch.load(args.load_weights)
except Exception as e:
print(e)
checkpoint = torch.load(args.load_weights, map_location={'cuda:0': 'cpu'})
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))
else:
print("no weight file!!!")
def main():
global args, dropout_optimizer
torch.manual_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
log_name = 'log_test.txt' if args.evaluate else 'log_train.txt'
sys.stderr = sys.stdout = Logger(osp.join(args.save_dir, log_name))
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, however, GPU is highly recommended")
print("Initializing image data manager")
dm = ImageDataManager(use_gpu, **image_dataset_kwargs(args))
trainloader, testloader_dict = dm.return_dataloaders()
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dm.num_train_pids,
loss={'xent'},
use_gpu=use_gpu,
dropout_optimizer=dropout_optimizer,
args=vars(args))
print(model)
print("Model size: {:.3f} M".format(count_num_param(model)))
# criterion = WrappedCrossEntropyLoss(num_classes=dm.num_train_pids, use_gpu=use_gpu, label_smooth=args.label_smooth)
criterion, fix_criterion, switch_criterion, htri_param_controller = get_criterions(
dm.num_train_pids, use_gpu, args)
regularizer, reg_param_controller = get_regularizer(args.regularizer)
optimizer = init_optimizer(model.parameters(), **optimizer_kwargs(args))
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma)
if args.load_weights and check_isfile(args.load_weights):
# load pretrained weights but ignore layers that don't match in size
try:
checkpoint = torch.load(args.load_weights)
except Exception as e:
print(e)
checkpoint = torch.load(args.load_weights,
map_location={'cuda:0': 'cpu'})
# dropout_optimizer.set_p(checkpoint.get('dropout_p', 0))
# print(list(checkpoint.keys()), checkpoint['dropout_p'])
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)
state = model.state_dict()
state.update(checkpoint['state_dict'])
model.load_state_dict(state)
# args.start_epoch = checkpoint['epoch'] + 1
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch,
checkpoint['rank1']))
if use_gpu:
model = nn.DataParallel(
model,
device_ids=list(range(len(args.gpu_devices.split(','))))).cuda()
if args.evaluate:
print("Evaluate only")
for name in args.target_names:
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query'], testloader_dict[
name]['query_flip']
galleryloader = testloader_dict[name]['gallery'], testloader_dict[
name]['gallery_flip']
distmat = test(model,
queryloader,
galleryloader,
use_gpu,
return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(distmat,
dm.return_testdataset_by_name(name),
save_dir=osp.join(
args.save_dir, 'ranked_results',
name),
topk=20)
return
start_time = time.time()
ranklogger = RankLogger(args.source_names, args.target_names)
train_time = 0
print("==> Start training")
if os.environ.get('test_first') is not None:
for name in args.target_names:
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query'], testloader_dict[
name]['query_flip']
galleryloader = testloader_dict[name]['gallery'], testloader_dict[
name]['gallery_flip']
rank1 = test(model, queryloader, galleryloader, use_gpu)
if args.fixbase_epoch > 0:
oldenv = os.environ.get('sa', '')
os.environ['sa'] = ''
print(
"Train {} for {} epochs while keeping other layers frozen".format(
args.open_layers, args.fixbase_epoch))
initial_optim_state = optimizer.state_dict()
for epoch in range(args.fixbase_epoch):
start_train_time = time.time()
train(epoch,
model,
fix_criterion,
regularizer,
optimizer,
trainloader,
use_gpu,
fixbase=True)
train_time += round(time.time() - start_train_time)
print("Done. All layers are open to train for {} epochs".format(
args.max_epoch))
optimizer.load_state_dict(initial_optim_state)
os.environ['sa'] = oldenv
max_r1 = 0
for epoch in range(args.start_epoch, args.max_epoch):
dropout_optimizer.set_epoch(epoch)
reg_param_controller.set_epoch(epoch)
htri_param_controller.set_epoch(epoch)
dropout_optimizer.set_training(True)
start_train_time = time.time()
print(epoch, args.switch_loss)
print(criterion)
cond = args.switch_loss > 0 and epoch >= args.switch_loss
cond = cond or (args.switch_loss < 0
and args.switch_loss + args.max_epoch < epoch)
if cond:
print('Switch!')
criterion = switch_criterion
train(epoch,
model,
criterion,
regularizer,
optimizer,
trainloader,
use_gpu,
fixbase=False,
switch_loss=cond)
train_time += round(time.time() - start_train_time)
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint(
{
'state_dict': state_dict,
'rank1': 0,
'epoch': epoch,
'dropout_p': dropout_optimizer.p,
}, False,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
scheduler.step()
if (epoch + 1) > args.start_eval and args.eval_freq > 0 and (
epoch + 1) % args.eval_freq == 0 or (epoch +
1) == args.max_epoch:
print("==> Test")
dropout_optimizer.set_training(False) # IMPORTANT!
for name in args.target_names:
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query'], testloader_dict[
name]['query_flip']
galleryloader = testloader_dict[name][
'gallery'], testloader_dict[name]['gallery_flip']
print('!!!!!!!!FC!!!!!!!!')
os.environ['NOFC'] = ''
rank1 = test(model, queryloader, galleryloader, use_gpu)
ranklogger.write(name, epoch + 1, rank1)
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
if max_r1 < rank1:
print('Save!', max_r1, rank1)
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
'dropout_p': dropout_optimizer.p,
}, False, osp.join(args.save_dir,
'checkpoint_best.pth.tar'))
max_r1 = rank1
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))
ranklogger.show_summary()
def main():
global args
set_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
log_name = 'log_test.txt' if args.evaluate else 'log_train.txt'
sys.stdout = Logger(osp.join(args.save_dir, log_name))
print('==========\nArgs:{}\n=========='.format(args))
if use_gpu:
print('Currently using GPU {}'.format(args.gpu_devices))
cudnn.benchmark = True
else:
print('Currently using CPU, however, GPU is highly recommended')
print('Initializing video data manager')
dm = VideoDataManager(use_gpu, **video_dataset_kwargs(args))
trainloader, testloader_dict = dm.return_dataloaders()
print('Initializing model: {}'.format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dm.num_train_pids,
loss={'xent', 'htri'},
pretrained=not args.no_pretrained,
use_gpu=use_gpu)
print('Model size: {:.3f} M'.format(count_num_param(model)))
if args.load_weights and check_isfile(args.load_weights):
load_pretrained_weights(model, 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))
model = nn.DataParallel(model).cuda() if use_gpu else model
criterion = CrossEntropyLoss(num_classes=dm.num_train_pids,
use_gpu=use_gpu,
label_smooth=args.label_smooth)
criterion_htri = TripletLoss(margin=args.margin)
optimizer = init_optimizer(model, **optimizer_kwargs(args))
scheduler = init_lr_scheduler(optimizer, **lr_scheduler_kwargs(args))
if args.evaluate:
print('Evaluate only')
for name in args.target_names:
print('Evaluating {} ...'.format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
distmat = test(model,
queryloader,
galleryloader,
args.pool_tracklet_features,
use_gpu,
return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(distmat,
dm.return_testdataset_by_name(name),
save_dir=osp.join(
args.save_dir, 'ranked_results',
name),
topk=20)
return
start_time = time.time()
ranklogger = RankLogger(args.source_names, args.target_names)
train_time = 0
print('=> Start training')
if args.fixbase_epoch > 0:
print(
'Train {} for {} epochs while keeping other layers frozen'.format(
args.open_layers, args.fixbase_epoch))
initial_optim_state = optimizer.state_dict()
for epoch in range(args.fixbase_epoch):
start_train_time = time.time()
train(epoch,
model,
criterion_xent,
criterion_htri,
optimizer,
trainloader,
use_gpu,
fixbase=True)
train_time += round(time.time() - start_train_time)
print('Done. All layers are open to train for {} epochs'.format(
args.max_epoch))
optimizer.load_state_dict(initial_optim_state)
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, optimizer,
trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
scheduler.step()
if (epoch + 1) > args.start_eval and args.eval_freq > 0 and (
epoch + 1) % args.eval_freq == 0 or (epoch +
1) == args.max_epoch:
print('=> Test')
for name in args.target_names:
print('Evaluating {} ...'.format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
rank1 = test(model, queryloader, galleryloader,
args.pool_tracklet_features, use_gpu)
ranklogger.write(name, epoch + 1, rank1)
save_checkpoint(
{
'state_dict': model.state_dict(),
'rank1': rank1,
'epoch': epoch,
}, False,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
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))
ranklogger.show_summary()
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 = DatasetManager(dataset_dir=args.dataset, root=args.root)
transform_train = T.Compose_Keypt([
T.Random2DTranslation_Keypt((args.width, args.height)),
T.RandomHorizontalFlip_Keypt(),
T.ToTensor_Keypt(),
T.Normalize_Keypt(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
transform_test = T.Compose_Keypt([
T.Resize_Keypt((args.width, args.height)),
T.ToTensor_Keypt(),
T.Normalize_Keypt(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,
keyptaware=args.keyptaware,
heatmapaware=args.heatmapaware,
segmentaware=args.segmentaware,
transform=transform_train,
imagesize=(args.width, args.height)),
sampler=RandomIdentitySampler(dataset.train, args.train_batch,
args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
ImageDataset(dataset.query,
keyptaware=args.keyptaware,
heatmapaware=args.heatmapaware,
segmentaware=args.segmentaware,
transform=transform_test,
imagesize=(args.width, args.height)),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
ImageDataset(dataset.gallery,
keyptaware=args.keyptaware,
heatmapaware=args.heatmapaware,
segmentaware=args.segmentaware,
transform=transform_test,
imagesize=(args.width, args.height)),
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_vids=dataset.num_train_vids,
num_vcolors=dataset.num_train_vcolors,
num_vtypes=dataset.num_train_vtypes,
keyptaware=args.keyptaware,
heatmapaware=args.heatmapaware,
segmentaware=args.segmentaware,
multitask=args.multitask)
print("Model size: {:.3f} M".format(count_num_param(model)))
if args.label_smooth:
criterion_xent_vid = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_vids, use_gpu=use_gpu)
criterion_xent_vcolor = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_vcolors, use_gpu=use_gpu)
criterion_xent_vtype = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_vtypes, use_gpu=use_gpu)
else:
criterion_xent_vid = nn.CrossEntropyLoss()
criterion_xent_vcolor = nn.CrossEntropyLoss()
criterion_xent_vtype = nn.CrossEntropyLoss()
criterion_htri = TripletLoss(margin=args.margin)
optimizer = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma)
if args.load_weights:
# load pretrained weights but ignore layers that don't match in size
if check_isfile(args.load_weights):
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:
if check_isfile(args.resume):
checkpoint = torch.load(args.resume)
model.load_state_dict(checkpoint['state_dict'])
args.start_epoch = checkpoint['epoch']
rank1 = checkpoint['rank1']
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(
args.start_epoch, rank1))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
distmat = test(model,
args.keyptaware,
args.multitask,
queryloader,
galleryloader,
use_gpu,
dataset.vcolor2label,
dataset.vtype2label,
return_distmat=True)
if args.vis_ranked_res:
visualize_ranked_results(
distmat,
dataset,
save_dir=osp.join(args.save_dir, 'ranked_results'),
topk=100,
)
return
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
print("==> Start training")
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, args.keyptaware, args.multitask,
criterion_xent_vid, criterion_xent_vcolor, criterion_xent_vtype,
criterion_htri, optimizer, trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
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, args.keyptaware, args.multitask, queryloader,
galleryloader, use_gpu, dataset.vcolor2label,
dataset.vtype2label)
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()
# rank1 = 1
# is_best = True
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 {:.2%}, 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))
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
logger_info = LoggerInfo()
sys.stdout = Logger(logger_info)
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.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
train_dataset = ImageFolder(os.path.join(args.data_dir, "train_all"),
transform=transform_train)
train_query_dataset = ImageFolder(os.path.join(args.data_dir, "val"),
transform=transform_train)
train_gallery_dataset = ImageFolder(os.path.join(args.data_dir, "train"),
transform=transform_train)
test_dataset = ImageFolder(os.path.join(args.data_dir, "probe"),
transform=transform_test)
query_dataset = ImageFolder(os.path.join(args.data_dir, "query"),
transform=transform_test)
gallery_dataset = ImageFolder(os.path.join(args.data_dir, "gallery"),
transform=transform_test)
# train_batch_sampler = VehicleIdBalancedBatchSampler(train_dataset, n_classes=8, n_samples=6)
# test_batch_sampler = VehicleIdBalancedBatchSampler(test_dataset, n_classes=8, n_samples=8)
train_batch_sampler = VehicleIdCCLBatchSampler(train_dataset,
n_classes=n_cls,
n_samples=n_samples)
trainloader = DataLoader(train_dataset,
batch_sampler=train_batch_sampler,
num_workers=args.workers,
pin_memory=pin_memory)
testloader = DataLoader(test_dataset,
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False)
# trainloader = DataLoader(
# ImageDataset(dataset.train, transform=transform_train),
# batch_sampler=train_batch_sampler, batch_size=args.train_batch,
# shuffle=True, num_workers=args.workers, pin_memory=pin_memory, drop_last=True
# )
queryloader = DataLoader(
query_dataset,
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
gallery_dataset,
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
train_query_loader = DataLoader(
train_query_dataset,
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
train_gallery_loader = DataLoader(
train_gallery_dataset,
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing model: {}".format(args.arch))
if args.evaluate:
model = models.init_model(name=args.arch,
num_classes=len(query_dataset.classes),
loss_type=args.loss_type)
else:
model = models.init_model(name=args.arch,
num_classes=len(train_dataset.classes),
loss_type=args.loss_type)
print("Model size: {:.3f} M".format(count_num_param(model)))
if args.label_smooth:
criterion = CrossEntropyLabelSmooth(num_classes=len(
train_dataset.classes),
use_gpu=use_gpu)
else:
if args.loss_type == 'xent':
criterion = nn.CrossEntropyLoss()
elif args.loss_type == 'angle':
criterion = AngleLoss()
elif args.loss_type == 'triplet':
# criterion = CoupledClustersLoss(margin=1., triplet_selector=RandomNegativeTripletSelector(margin=1.))
# criterion = OnlineTripletLoss(margin=1., triplet_selector=RandomNegativeTripletSelector(margin=1.))
# criterion = OnlineTripletLoss(margin=1., triplet_selector=HardestNegativeTripletSelector(margin=1.))
criterion = CoupledClustersLoss(margin=1.,
n_classes=n_cls,
n_samples=n_samples)
# criterion = OnlineTripletLoss(margin=1., triplet_selector=SemihardNegativeTripletSelector(margin=1.))
elif args.loss_type == 'xent_htri':
criterion = XentTripletLoss(
margin=1.,
triplet_selector=RandomNegativeTripletSelector(margin=1.))
else:
raise KeyError("Unsupported loss: {}".format(args.loss_type))
# model_param_list = [{'params': model.base.parameters(), 'lr': args.lr},
# {'params': model.classifier.parameters(), 'lr': args.lr * 10}]
# optimizer = init_optim(args.optim, model_param_list, lr=1.0, weight_decay=args.weight_decay)
optimizer = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
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:
# load pretrained weights but ignore layers that don't match in size
if check_isfile(args.load_weights):
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:
from functools import partial
import pickle
pickle.load = partial(pickle.load, encoding="latin1")
pickle.Unpickler = partial(pickle.Unpickler, encoding="latin1")
if check_isfile(args.resume):
checkpoint = torch.load(args.resume)
# checkpoint = torch.load(args.resume, map_location=lambda storage, loc: storage, pickle_module=pickle)
model.load_state_dict(checkpoint['state_dict'])
args.start_epoch = checkpoint['epoch'] + 1
rank1 = checkpoint['rank1']
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(
args.start_epoch, rank1))
if use_gpu:
model = nn.DataParallel(model).cuda()
# if args.evaluate:
# print("Evaluate only")
# distmat = test(model, queryloader, galleryloader, train_query_loader, train_gallery_loader,
# use_gpu, return_distmat=True)
# if args.vis_ranked_res:
# visualize_ranked_results(
# distmat, dataset,
# save_dir=osp.join(args.save_dir, 'ranked_results'),
# topk=20,
# )
# return
start_time = time.time()
train_time = 0
best_rank1 = -np.inf
best_epoch = 0
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,
freeze_bn=True)
train_time += round(time.time() - start_train_time)
del optimizer_tmp
print("Now open all layers for training")
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion, optimizer, trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
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, testloader, queryloader, galleryloader,
train_query_loader, train_gallery_loader, 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,
use_gpu_suo=True,
fpath=osp.join(
args.save_dir, 'checkpoint_ep' + str(epoch + 1) +
checkpoint_suffix + '.pth.tar'))
print("==> Best Rank-1 {:.2%}, 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))
def main():
global args
torch.manual_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
log_name = 'log_train_{}.txt'.format(time.strftime("%Y-%m-%d-%H-%M-%S"))
if args.evaluate:
log_name.replace('train', 'test')
sys.stdout = Logger(osp.join(args.save_dir, log_name))
print(' '.join(sys.argv))
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, however, GPU is highly recommended")
print("Initializing image data manager")
dm = ImageDataManager(use_gpu, **image_dataset_kwargs(args))
if hasattr(dm, 'lfw_dataset'):
lfw = dm.lfw_dataset
print('LFW dataset is used!')
else:
lfw = None
trainloader, trainloader_dict, testloader_dict = dm.return_dataloaders()
num_train_pids = dm.num_train_pids
print("Initializing model: {}".format(args.arch))
model = models.init_model(
name=args.arch,
num_classes=num_train_pids,
loss={'xent', 'htri'},
pretrained=False if args.load_weights else 'imagenet',
grayscale=args.grayscale,
normalize_embeddings=args.normalize_embeddings,
normalize_fc=args.normalize_fc,
convbn=args.convbn)
print("Model size: {:.3f} M".format(count_num_param(model)))
count_flops(model, args.height, args.width, args.grayscale)
if args.load_weights and check_isfile(args.load_weights):
# load pretrained weights but ignore layers that don't match in size
load_weights(model, args.load_weights)
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
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch,
checkpoint['rank1']))
if use_gpu:
model = nn.DataParallel(model).cuda()
model = model.cuda()
if args.evaluate:
print("Evaluate only")
for name in args.target_names:
if not 'lfw' in name.lower():
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
distmat = test(args,
model,
queryloader,
galleryloader,
use_gpu,
return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(
distmat,
dm.return_testdataset_by_name(name),
save_dir=osp.join(args.save_dir, 'ranked_results',
name),
topk=20)
else:
model.eval()
same_acc, diff_acc, all_acc, auc, thresh = evaluate(
args,
dm.lfw_dataset,
model,
compute_embeddings_lfw,
args.test_batch_size,
verbose=False,
show_failed=args.show_failed)
log.info('Validation accuracy: {0:.4f}, {1:.4f}'.format(
same_acc, diff_acc))
log.info('Validation accuracy mean: {0:.4f}'.format(all_acc))
log.info('Validation AUC: {0:.4f}'.format(auc))
log.info('Estimated threshold: {0:.4f}'.format(thresh))
return
criterions = choose_losses(args, dm, model, use_gpu)
if not args.evaluate and len(criterions) == 0:
raise AssertionError('No loss functions were chosen!')
optimizer = init_optimizer(model.parameters(), **optimizer_kwargs(args))
if args.load_optim:
checkpoint = torch.load(args.load_weights)
optimizer.load_state_dict(checkpoint['optim'])
print("Loaded optimizer from '{}'".format(args.load_weights))
for param_group in optimizer.param_groups:
param_group['lr'] = args.lr
param_group['weight_decay'] = args.weight_decay
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma)
start_time = time.time()
ranklogger = RankLogger(args.source_names, args.target_names)
train_time = 0
train_writer = SummaryWriter(osp.join(args.save_dir, 'train_log'))
test_writer = SummaryWriter(osp.join(args.save_dir, 'test_log'))
print("=> Start training")
if args.fixbase_epoch > 0:
print(
"Train {} for {} epochs while keeping other layers frozen".format(
args.open_layers, args.fixbase_epoch))
initial_optim_state = optimizer.state_dict()
for epoch in range(args.fixbase_epoch):
start_train_time = time.time()
train(epoch,
model,
criterions,
optimizer,
trainloader,
use_gpu,
train_writer,
fixbase=True,
lfw=lfw)
train_time += round(time.time() - start_train_time)
for name in args.target_names:
if not 'lfw' in name.lower():
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
testloader = testloader_dict[name]['test']
criteria = None
rank1 = test(args,
model,
queryloader,
galleryloader,
use_gpu,
testloader=testloader,
criterions=criteria)
else:
model.eval()
same_acc, diff_acc, all_acc, auc, thresh = evaluate(
args,
dm.lfw_dataset,
model,
compute_embeddings_lfw,
args.test_batch_size,
verbose=False,
show_failed=args.show_failed)
print('Validation accuracy: {0:.4f}, {1:.4f}'.format(
same_acc, diff_acc))
print('Validation accuracy mean: {0:.4f}'.format(all_acc))
print('Validation AUC: {0:.4f}'.format(auc))
print('Estimated threshold: {0:.4f}'.format(thresh))
rank1 = all_acc
print("Done. All layers are open to train for {} epochs".format(
args.max_epoch))
optimizer.load_state_dict(initial_optim_state)
for epoch in range(args.start_epoch, args.max_epoch):
for criterion in criterions:
criterion.train_stats.reset()
start_train_time = time.time()
train(epoch,
model,
criterions,
optimizer,
trainloader,
use_gpu,
train_writer,
lfw=lfw)
train_time += round(time.time() - start_train_time)
scheduler.step()
if (epoch + 1) > args.start_eval and args.eval_freq > 0 and (
epoch + 1) % args.eval_freq == 0 or (epoch +
1) == args.max_epoch:
num_iter = (epoch + 1) * len(trainloader)
if not args.no_train_quality:
for name in args.source_names:
print(
"Measure quality on the {} train set...".format(name))
queryloader = trainloader_dict[name]['query']
galleryloader = trainloader_dict[name]['gallery']
rank1 = test(args, model, queryloader, galleryloader,
use_gpu)
train_writer.add_scalar('rank1/{}'.format(name), rank1,
num_iter)
print("=> Test")
for name in args.target_names:
if not 'lfw' in name.lower():
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
testloader = testloader_dict[name]['test']
criteria = criterions
if args.no_loss_on_val:
criteria = None
rank1 = test(args,
model,
queryloader,
galleryloader,
use_gpu,
testloader=testloader,
criterions=criteria)
test_writer.add_scalar('rank1/{}'.format(name), rank1,
num_iter)
if not args.no_loss_on_val:
for criterion in criterions:
test_writer.add_scalar(
'loss/{}'.format(criterion.name),
criterion.test_stats.avg, num_iter)
criterion.test_stats.reset()
ranklogger.write(name, epoch + 1, rank1)
else:
model.eval()
same_acc, diff_acc, all_acc, auc, thresh = evaluate(
args,
dm.lfw_dataset,
model,
compute_embeddings_lfw,
args.test_batch_size,
verbose=False,
show_failed=args.show_failed)
print('Validation accuracy: {0:.4f}, {1:.4f}'.format(
same_acc, diff_acc))
print('Validation accuracy mean: {0:.4f}'.format(all_acc))
print('Validation AUC: {0:.4f}'.format(auc))
print('Estimated threshold: {0:.4f}'.format(thresh))
test_writer.add_scalar('Accuracy/Val_same_accuracy',
same_acc, num_iter)
test_writer.add_scalar('Accuracy/Val_diff_accuracy',
diff_acc, num_iter)
test_writer.add_scalar('Accuracy/Val_accuracy', all_acc,
num_iter)
test_writer.add_scalar('Accuracy/AUC', auc, num_iter)
rank1 = all_acc
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_dict = {
'state_dict': state_dict,
'epoch': epoch,
'optim': optimizer.state_dict()
}
if len(args.target_names):
save_dict['rank1'] = rank1
save_checkpoint(
save_dict, False,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
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))
ranklogger.show_summary()
def main():
global args
set_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
log_name = 'log_test.txt' if args.evaluate else 'log_train.txt'
sys.stdout = Logger(osp.join(args.save_dir, log_name))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
else:
print("Currently using CPU, however, GPU is highly recommended")
print("Initializing MultiScale data manager")
assert args.train_batch_size % args.train_loss_batch_size == 0, "'{}' is not divisable by {}".format(args.train_loss_batch_size, args.train_loss_batch_size)
dm = ImageDataManager(use_gpu, scales=[224,160], **image_dataset_kwargs(args))
trainloader, testloader_dict = dm.return_dataloaders()
# sys.exit(0)
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch, num_classes=dm.num_train_pids, input_size=args.width, loss={'xent'}, use_gpu=use_gpu)
print("Model size: {:.3f} M".format(count_num_param(model)))
# print(model)
criterion = CrossEntropyLoss(num_classes=dm.num_train_pids, use_gpu=use_gpu, label_smooth=args.label_smooth)
optimizer = init_optimizer(model.parameters(), **optimizer_kwargs(args))
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=args.stepsize, gamma=args.gamma)
# # scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=3, verbose=True, threshold=1e-04)
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
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch, checkpoint['rank1']))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
for name in args.target_names:
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
test_set = dm.return_testdataset_by_name(name)
rank1, mAP = test(model, test_set, name, queryloader, galleryloader, use_gpu, visualize=args.visualize_ranks)
return
start_time = time.time()
ranklogger = RankLogger(args.source_names, args.target_names)
maplogger = RankLogger(args.source_names, args.target_names)
train_time = 0
# Tensorboard
writer = SummaryWriter(log_dir=osp.join('runs', args.save_dir))
print("=> Start training")
if args.fixbase_epoch > 0:
print("Train {} for {} epochs while keeping other layers frozen".format(args.open_layers, args.fixbase_epoch))
initial_optim_state = optimizer.state_dict()
for epoch in range(args.fixbase_epoch):
start_train_time = time.time()
loss, prec1 = train(epoch, model, criterion, optimizer, trainloader, writer, use_gpu, fixbase=True)
writer.add_scalar('train/loss', loss, epoch+1)
writer.add_scalar('train/prec1', prec1, epoch+1)
print('Epoch: [{:02d}] [Average Loss:] {:.4f}\t [Average Prec.:] {:.2%}'.format(epoch+1, loss, prec1))
train_time += round(time.time() - start_train_time)
print("Done. All layers are open to train for {} epochs".format(args.max_epoch))
optimizer.load_state_dict(initial_optim_state)
args.start_epoch += args.fixbase_epoch
args.max_epoch += args.fixbase_epoch
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
loss, prec1 = train(epoch, model, criterion, optimizer, trainloader, writer, use_gpu)
writer.add_scalar('train/loss', loss, epoch+1)
writer.add_scalar('train/prec1', prec1, epoch+1)
print('Epoch: [{:02d}] [Average Loss:] {:.4f}\t [Average Prec.:] {:.2%}'.format(epoch+1, loss, prec1))
train_time += round(time.time() - start_train_time)
scheduler.step()
if (epoch + 1) > args.start_eval and args.eval_freq > 0 and (epoch + 1) % args.eval_freq == 0 or (epoch + 1) == args.max_epoch:
print("=> Test")
for name in args.target_names:
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
test_set = dm.return_testdataset_by_name(name)
if epoch+1 == args.max_epoch:
rank1, mAP = test(model, test_set, name, queryloader, galleryloader, use_gpu, visualize=True)
else:
rank1, mAP = test(model, test_set, name, queryloader, galleryloader, use_gpu)
writer.add_scalar(name + '_test/top1', rank1, epoch+1)
writer.add_scalar(name + '_test/mAP', mAP, epoch+1)
ranklogger.write(name, epoch + 1, rank1)
maplogger.write(name, epoch + 1, mAP)
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,
}, False, osp.join(args.save_dir, 'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
# save last checkpoint
save_checkpoint({
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
}, False, osp.join(args.save_dir, 'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
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))
ranklogger.show_summary()
maplogger.show_summary()
def main():
global args, dropout_optimizer
torch.manual_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
log_name = 'log_test.txt' if args.evaluate else 'log_train.txt'
sys.stderr = sys.stdout = Logger(osp.join(args.save_dir, log_name))
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, however, GPU is highly recommended")
print("Initializing image data manager")
dm = ImageDataManager(use_gpu, **image_dataset_kwargs(args))
trainloader, testloader_dict = dm.return_dataloaders()
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch, num_classes=dm.num_train_pids, loss={'xent'}, use_gpu=use_gpu,
dropout_optimizer=dropout_optimizer)
print(model)
print("Model size: {:.3f} M".format(count_num_param(model)))
# criterion = WrappedCrossEntropyLoss(num_classes=dm.num_train_pids, use_gpu=use_gpu, label_smooth=args.label_smooth)
criterion, fix_criterion, switch_criterion, htri_param_controller = get_criterions(dm.num_train_pids, use_gpu, args)
regularizer, reg_param_controller = get_regularizer(args.regularizer)
optimizer = init_optimizer(model.parameters(), **optimizer_kwargs(args))
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=args.stepsize, gamma=args.gamma)
if args.load_weights and check_isfile(args.load_weights):
# load pretrained weights but ignore layers that don't match in size
try:
checkpoint = torch.load(args.load_weights)
except Exception as e:
print(e)
checkpoint = torch.load(args.load_weights, map_location={'cuda:0': 'cpu'})
# dropout_optimizer.set_p(checkpoint.get('dropout_p', 0))
# print(list(checkpoint.keys()), checkpoint['dropout_p'])
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)
state = model.state_dict()
state.update(checkpoint['state_dict'])
model.load_state_dict(state)
# args.start_epoch = checkpoint['epoch'] + 1
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch, checkpoint['rank1']))
if use_gpu:
model = nn.DataParallel(model, device_ids=list(range(len(args.gpu_devices.split(','))))).cuda()
extract_train_info(model, trainloader)
def main():
global args
if not args.evaluate:
raise RuntimeError('Test only!')
torch.manual_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
log_name = 'log_test.txt' if args.evaluate else 'log_train.txt'
log_fn = osp.join(args.save_dir, log_name)
sys.stderr = sys.stdout = Logger(log_fn)
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, however, GPU is highly recommended")
print("Initializing image data manager")
dm = ImageDataManager(use_gpu, **image_dataset_kwargs(args))
trainloader, testloader_dict = dm.return_dataloaders()
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dm.num_train_pids,
loss={'xent'},
use_gpu=use_gpu,
args=vars(args))
print(model)
print("Model size: {:.3f} M".format(count_num_param(model)))
if args.load_weights and check_isfile(args.load_weights):
# load pretrained weights but ignore layers that don't match in size
try:
checkpoint = torch.load(args.load_weights)
except Exception as e:
print(e)
checkpoint = torch.load(args.load_weights,
map_location={'cuda:0': 'cpu'})
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)
state = model.state_dict()
state.update(checkpoint['state_dict'])
model.load_state_dict(state)
# args.start_epoch = checkpoint['epoch'] + 1
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch,
checkpoint['rank1']))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
for name in args.target_names:
print("Evaluating {} ...".format(name))
distmat = test(model,
testloader_dict[name],
use_gpu,
return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(distmat,
dm.return_testdataset_by_name(name),
save_dir=osp.join(
args.save_dir, 'ranked_results',
name),
topk=20)
return
def choose_losses(args, data_manager, model, use_gpu):
criteria = []
if args.xent_loss == 'am':
criteria.append(Criterion('am', AMSoftmaxLoss(), args.weight_xent))
elif args.xent_loss is not None:
criteria.append(
Criterion(
'xent',
CrossEntropyLoss(num_classes=data_manager.num_train_pids,
use_gpu=use_gpu,
label_smooth=args.label_smooth),
args.weight_xent))
if args.euclid_loss:
if args.euclid_loss == 'triplet':
loss = TripletLoss(margin=args.margin)
elif args.euclid_loss == 'lifted':
loss = LiftedLoss(margin=args.margin)
else:
raise KeyError('Unknown euclidean loss: {}'.format(
args.euclid_loss))
criteria.append(Criterion('euclid', loss, args.weight_euclid))
if args.face_loss:
input_channels = 3
if args.grayscale:
input_channels = 1
features = model(
torch.zeros(1, input_channels, args.height, args.width))
feature_vector_size = features.shape[1]
if args.face_loss == 'arc':
loss = ArcFace(feature_vector_size, data_manager.num_train_pids)
elif args.face_loss == 'cos':
loss = CosFace(feature_vector_size, data_manager.num_train_pids)
elif args.face_loss == 'sphere':
loss = SphereFace(feature_vector_size, data_manager.num_train_pids)
elif args.face_loss == 'am':
loss = AMSoftmaxLoss()
else:
raise KeyError('Unknown face loss: {}'.format(args.face_loss))
if use_gpu:
loss = loss.cuda()
criteria.append(Criterion('face', loss, args.weight_face))
if args.teacher_arch:
print("Initializing teacher model: {}".format(args.teacher_arch))
teacher_model = models.init_model(
name=args.teacher_arch,
num_classes=data_manager.num_train_pids,
loss={'xent', 'htri'},
pretrained=False)
if not args.load_teacher_weights or not check_isfile(
args.load_teacher_weights):
print('Teacher model checkpoint wasn\'t provided!')
return
load_weights(teacher_model, args.load_teacher_weights)
print("Loaded pretrained weights from '{}' for teacher model".format(
args.load_teacher_weights))
if use_gpu:
teacher_model = nn.DataParallel(teacher_model).cuda()
criteria.append(
Criterion('distill', DistillLoss(teacher_model),
args.weight_distill))
return criteria
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
logger_info = LoggerInfo()
sys.stdout = Logger(logger_info)
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("\nInitializing dataset {}".format(args.dataset_plt))
dataset_plt = data_manager.init_imgreid_dataset(root=args.root,
name=args.dataset_plt)
print("\nInitializing dataset {}".format(args.dataset_vecl))
dataset_vecl = data_manager.init_imgreid_dataset(root=args.root,
name=args.dataset_vecl)
transform_test_plt = T.Compose([
T.Resize((args.height_plt, args.width_plt)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
# transform_flip_test_plt = T.Compose([
# T.Resize((args.height_plt, args.width_plt)),
# functional.hflip,
# T.ToTensor(),
# T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
# ])
transform_test_vecl = T.Compose([
T.Resize((args.height_vecl, args.width_vecl)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
# transform_flip_test_vecl = T.Compose([
# T.Resize((args.height_vecl, args.width_vecl)),
# functional.hflip,
# 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
queryloader_plt = DataLoader(
ImageDatasetV2(dataset_plt.query, transform=transform_test_plt),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
# queryloader_flip_plt = DataLoader(
# ImageDatasetV2(dataset_plt.query, transform=transform_flip_test_plt),
# batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
# pin_memory=pin_memory, drop_last=False,
# )
# queryloader_plt = [queryloader_plt, queryloader_flip_plt]
queryloader_plt = [queryloader_plt]
galleryloader_plt = DataLoader(
ImageDatasetV2(dataset_plt.gallery, transform=transform_test_plt),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
# galleryloader_flip_plt = DataLoader(
# ImageDatasetV2(dataset_plt.gallery, transform=transform_flip_test_plt),
# batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
# pin_memory=pin_memory, drop_last=False,
# )
# galleryloader_plt = [galleryloader_plt, galleryloader_flip_plt]
galleryloader_plt = [galleryloader_plt]
queryloader_vecl = DataLoader(
ImageDatasetWGL(dataset_vecl.query,
transform=transform_test_vecl,
with_image_name=True),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
# queryloader_flip_vecl = DataLoader(
# ImageDatasetV2(dataset_vecl.query, transform=transform_flip_test_vecl),
# batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
# pin_memory=pin_memory, drop_last=False,
# )
# queryloader_vecl = [queryloader_vecl, queryloader_flip_vecl]
queryloader_vecl = [queryloader_vecl]
galleryloader_vecl = DataLoader(
ImageDatasetWGL(dataset_vecl.gallery,
transform=transform_test_vecl,
with_image_name=True),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
# galleryloader_flip_vecl = DataLoader(
# ImageDatasetV2(dataset_vecl.gallery, transform=transform_flip_test_vecl),
# batch_size=args.test_batch, shuffle=False, num_workers=args.workers,
# pin_memory=pin_memory, drop_last=False,
# )
# galleryloader_vecl = [galleryloader_vecl, galleryloader_flip_vecl]
galleryloader_vecl = [galleryloader_vecl]
print("\nInitializing model: {}".format(args.arch))
model_plt = models.init_model(name=args.arch_plt,
num_classes=dataset_plt.num_train_pids,
loss_type=args.loss_type)
model_vecl = models.init_model(name=args.arch_vecl,
num_classes=dataset_vecl.num_train_pids,
loss_type=args.loss_type)
print("Plate model size: {:.3f} M".format(count_num_param(model_plt)))
print("Vehicle model size: {:.3f} M".format(count_num_param(model_vecl)))
if args.loss_type == 'xent':
criterion = nn.CrossEntropyLoss()
else:
raise KeyError("Unsupported loss: {}".format(args.loss_type))
if args.resm_plt and args.resm_vecl:
if check_isfile(args.resm_plt) and check_isfile(args.resm_vecl):
ckpt_plt = torch.load(args.resm_plt)
pre_dic_plt = ckpt_plt['state_dict']
model_dic_plt = model_plt.state_dict()
pre_dic_plt = {
k: v
for k, v in pre_dic_plt.items()
if k in model_dic_plt and model_dic_plt[k].size() == v.size()
}
model_dic_plt.update(pre_dic_plt)
model_plt.load_state_dict(model_dic_plt)
args.start_epoch_plt = ckpt_plt['epoch']
rank1_plt = ckpt_plt['rank1']
ckpt_vecl = torch.load(args.resm_vecl)
pre_dic_vecl = ckpt_vecl['state_dict']
model_dic_vecl = model_vecl.state_dict()
pre_dic_vecl = {
k: v
for k, v in pre_dic_vecl.items() if k in model_dic_vecl
and model_dic_vecl[k].size() == v.size()
}
model_dic_vecl.update(pre_dic_vecl)
model_vecl.load_state_dict(model_dic_vecl)
args.start_epoch_vecl = ckpt_vecl['epoch']
rank1_vecl = ckpt_vecl['rank1']
print("\nLoaded checkpoint from '{}' \nand '{}".format(
args.resm_plt, args.resm_vecl))
print("Plate model: start_epoch: {}, rank1: {}".format(
args.start_epoch_plt, rank1_plt))
print("Vehicle model: start_epoch: {}, rank1: {}".format(
args.start_epoch_vecl, rank1_vecl))
if use_gpu:
model_plt = nn.DataParallel(model_plt).cuda()
model_vecl = nn.DataParallel(model_vecl).cuda()
if args.evaluate:
print("\nEvaluate only")
test(model_plt, model_vecl, queryloader_plt, queryloader_vecl,
galleryloader_plt, galleryloader_vecl, use_gpu)
return
def main():
global args, criterion, testloader_dict, trainloader, use_gpu
set_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
log_name = 'test.log' if args.evaluate else 'train.log'
sys.stdout = Logger(osp.join(args.save_dir, log_name))
print('==========\nArgs:{}\n=========='.format(args))
if use_gpu:
print('Currently using GPU {}'.format(args.gpu_devices))
cudnn.benchmark = True
else:
warnings.warn('Currently using CPU, however, GPU is highly recommended')
print('Initializing image data manager')
dm = ImageDataManager(use_gpu, **image_dataset_kwargs(args))
trainloader, testloader_dict = dm.return_dataloaders()
print('Initializing model: {}'.format(args.arch))
model = models.init_model(name=args.arch, num_classes=dm.num_train_pids, loss={'xent'}, pretrained=not args.no_pretrained, use_gpu=use_gpu)
print('Model size: {:.3f} M'.format(count_num_param(model)))
if args.load_weights and check_isfile(args.load_weights):
load_pretrained_weights(model, args.load_weights)
model = nn.DataParallel(model).cuda() if use_gpu else model
criterion = CrossEntropyLoss(num_classes=dm.num_train_pids, use_gpu=use_gpu, label_smooth=args.label_smooth)
if args.resume and check_isfile(args.resume):
args.start_epoch = resume_from_checkpoint(args.resume, model, optimizer=None)
resumed = True
else:
resumed = False
if args.evaluate:
print('Evaluate only')
for name in args.target_names:
print('Evaluating {} ...'.format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
distmat = test(model, queryloader, galleryloader, use_gpu, return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(
distmat, dm.return_testdataset_by_name(name),
save_dir=osp.join(args.save_dir, 'ranked_results', name),
topk=20
)
return
time_start = time.time()
# ranklogger = RankLogger(args.source_names, args.target_names)
print('=> Start training')
if not resumed:
train_base(model)
train_RRI(model, 7)
elapsed = round(time.time() - time_start)
elapsed = str(datetime.timedelta(seconds=elapsed))
print('Elapsed {}'.format(elapsed))
def main():
global args
torch.manual_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
log_name = 'log_test.txt' if args.evaluate else 'log_train.txt'
sys.stdout = Logger(osp.join(args.save_dir, log_name))
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, however, GPU is highly recommended")
print("Initializing image data manager")
dm = ImageDataManager(use_gpu, **image_dataset_kwargs(args))
trainloader, testloader_dict = dm.return_dataloaders()
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch, num_classes=dm.num_train_pids, loss={'xent', 'htri'})
print("Model size: {:.3f} M".format(count_num_param(model)))
criterion_xent = CrossEntropyLoss(num_classes=dm.num_train_pids, use_gpu=use_gpu, label_smooth=args.label_smooth)
criterion_htri = TripletLoss(margin=args.margin)
optimizer = init_optimizer(model.parameters(), **optimizer_kwargs(args))
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=args.stepsize, gamma=args.gamma)
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
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch, checkpoint['rank1']))
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
for name in args.target_names:
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
distmat = test(model, queryloader, galleryloader, use_gpu, return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(
distmat, dm.return_testdataset_by_name(name),
save_dir=osp.join(args.save_dir, 'ranked_results', name),
topk=20
)
return
start_time = time.time()
ranklogger = RankLogger(args.source_names, args.target_names)
train_time = 0
print("=> Start training")
if args.fixbase_epoch > 0:
print("Train {} for {} epochs while keeping other layers frozen".format(args.open_layers, args.fixbase_epoch))
initial_optim_state = optimizer.state_dict()
for epoch in range(args.fixbase_epoch):
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, optimizer, trainloader, use_gpu, fixbase=True)
train_time += round(time.time() - start_train_time)
print("Done. All layers are open to train for {} epochs".format(args.max_epoch))
optimizer.load_state_dict(initial_optim_state)
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, optimizer, trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
scheduler.step()
if (epoch + 1) > args.start_eval and args.eval_freq > 0 and (epoch + 1) % args.eval_freq == 0 or (epoch + 1) == args.max_epoch:
print("=> Test")
for name in args.target_names:
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
rank1 = test(model, queryloader, galleryloader, use_gpu)
ranklogger.write(name, epoch + 1, rank1)
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,
}, False, osp.join(args.save_dir, 'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
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))
ranklogger.show_summary()
def main():
global args, best_rank1
torch.manual_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:
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_vidreid_dataset(root=args.root,
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
# decompose tracklets into images for image-based training
new_train = []
for img_paths, pid, camid in dataset.train:
for img_path in img_paths:
new_train.append((img_path, pid, camid))
trainloader = DataLoader(
ImageDataset(new_train, transform=transform_train),
sampler=RandomIdentitySampler(new_train, args.train_batch,
args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
VideoDataset(dataset.query,
seq_len=args.seq_len,
sample='evenly',
transform=transform_test),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
seq_len=args.seq_len,
sample='evenly',
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'})
print("Model size: {:.3f} M".format(count_num_param(model)))
if args.label_smooth:
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
else:
criterion_xent = nn.CrossEntropyLoss()
criterion_htri = TripletLoss(margin=args.margin)
optimizer = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma)
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.evaluate:
print("Evaluate only")
distmat = test(model,
queryloader,
galleryloader,
args.pool,
use_gpu,
return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(
distmat,
dataset,
save_dir=osp.join(args.save_dir, 'ranked_results'),
topk=20,
)
return
start_time = time.time()
train_time = 0
best_epoch = args.start_epoch
print("==> Start training")
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion_xent, criterion_htri, optimizer,
trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
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, args.pool, 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))
def main():
global use_apex
global args
torch.manual_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
log_name = 'log_test.txt' if args.evaluate else 'log_train.txt'
sys.stderr = sys.stdout = Logger(osp.join(args.save_dir, log_name))
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, however, GPU is highly recommended")
print("Initializing image data manager")
dm = ImageDataManager(use_gpu, **image_dataset_kwargs(args))
trainloader, testloader_dict = dm.return_dataloaders()
print("Initializing model: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dm.num_train_pids,
loss={'xent'},
use_gpu=use_gpu,
args=vars(args))
print(model)
print("Model size: {:.3f} M".format(count_num_param(model)))
if use_gpu:
print("using gpu")
model = model.cuda()
print("criterion===>")
criterion = get_criterion(dm.num_train_pids, use_gpu, args)
print(criterion)
print("regularizer===>")
regularizer = get_regularizer(vars(args))
print(regularizer)
print("optimizer===>")
optimizer = init_optimizer(model.parameters(), **optimizer_kwargs(args))
print(optimizer)
print("scheduler===>")
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
'max',
factor=0.1,
patience=5,
verbose=True)
print(scheduler)
if args.load_weights and check_isfile(args.load_weights):
# load pretrained weights but ignore layers that don't match in size
try:
checkpoint = torch.load(args.load_weights)
except Exception as e:
print(e)
checkpoint = torch.load(args.load_weights,
map_location={'cuda:0': 'cpu'})
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))
max_r1 = 0
if args.resume and check_isfile(args.resume):
checkpoint = torch.load(args.resume)
state = model.state_dict()
state.update(checkpoint['state_dict'])
model.load_state_dict(state)
optimizer.load_state_dict(checkpoint['optimizer'])
args.start_epoch = checkpoint['epoch'] + 1
max_r1 = checkpoint['rank1']
print("Loaded checkpoint from '{}'".format(args.resume))
print("- start_epoch: {}\n- rank1: {}".format(args.start_epoch,
checkpoint['rank1']))
if use_apex:
print("using apex")
model, optimizer = amp.initialize(model, optimizer, opt_level="O0")
if args.evaluate:
print("Evaluate only")
for name in args.target_names:
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query'], testloader_dict[
name]['query_flip']
galleryloader = testloader_dict[name]['gallery'], testloader_dict[
name]['gallery_flip']
distmat = test(model,
queryloader,
galleryloader,
use_gpu,
return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(distmat,
dm.return_testdataset_by_name(name),
save_dir=osp.join(
args.save_dir, 'ranked_results',
name),
topk=20)
return
start_time = time.time()
ranklogger = RankLogger(args.source_names, args.target_names)
train_time = 0
print("==> Start training")
if args.fixbase_epoch > 0:
oldenv = os.environ.get('sa', '')
os.environ['sa'] = ''
print(
"Train {} for {} epochs while keeping other layers frozen".format(
args.open_layers, args.fixbase_epoch))
initial_optim_state = optimizer.state_dict()
for epoch in range(args.fixbase_epoch):
start_train_time = time.time()
train(epoch,
model,
criterion,
regularizer,
optimizer,
trainloader,
use_gpu,
fixbase=True)
train_time += round(time.time() - start_train_time)
print("Done. All layers are open to train for {} epochs".format(
args.max_epoch))
optimizer.load_state_dict(initial_optim_state)
os.environ['sa'] = oldenv
for epoch in range(args.start_epoch, args.max_epoch):
auto_reset_learning_rate(optimizer, args)
print(
f"===========================start epoch {epoch + 1} {now()}==========================================="
)
print(f"lr:{optimizer.param_groups[0]['lr']}")
loss = train(epoch,
model,
criterion,
regularizer,
optimizer,
trainloader,
use_gpu,
fixbase=False)
train_time += round(time.time() - start_train_time)
state_dict = model.state_dict()
rank1 = 0
if (epoch + 1) > args.start_eval and args.eval_freq > 0 and (
epoch + 1) % args.eval_freq == 0 or (epoch +
1) == args.max_epoch:
print("==> Test")
for name in args.target_names:
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query'], testloader_dict[
name]['query_flip']
galleryloader = testloader_dict[name][
'gallery'], testloader_dict[name]['gallery_flip']
rank1 = test(model, queryloader, galleryloader, use_gpu)
ranklogger.write(name, epoch + 1, rank1)
if max_r1 < rank1:
print('Save!', max_r1, rank1)
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
'optimizer': optimizer.state_dict(),
}, False, osp.join(args.save_dir,
'checkpoint_best.pth.tar'))
max_r1 = rank1
save_checkpoint(
{
'state_dict': state_dict,
'rank1': rank1,
'epoch': epoch,
'optimizer': optimizer.state_dict(),
}, False,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
scheduler.step(rank1)
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))
ranklogger.show_summary()
def main():
global args
torch.manual_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
log_name = 'log_test.txt'
sys.stdout = Logger(osp.join(args.save_dir, log_name))
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, however, GPU is highly recommended")
print("Initializing image data manager")
if not args.convert_to_onnx: # and not args.infer:
dm = ImageDataManager(use_gpu, **image_dataset_kwargs(args))
trainloader, trainloader_dict, testloader_dict = dm.return_dataloaders(
)
num_train_pids = 100
print("Initializing model: {}".format(args.arch))
model = models.init_model(
name=args.arch,
num_classes=num_train_pids,
loss={'xent', 'htri'},
pretrained=False if args.load_weights else 'imagenet',
grayscale=args.grayscale,
ceil_mode=not args.convert_to_onnx,
infer=True,
bits=args.bits,
normalize_embeddings=args.normalize_embeddings,
normalize_fc=args.normalize_fc,
convbn=args.convbn)
print("Model size: {:.3f} M".format(count_num_param(model)))
if args.load_weights and check_isfile(args.load_weights):
# load pretrained weights but ignore layers that don't match in size
load_weights(model, args.load_weights)
print("Loaded pretrained weights from '{}'".format(args.load_weights))
if args.absorb_bn:
search_absorbed_bn(model)
if args.quantization or args.save_quantized_model:
from gap_quantization.quantization import ModelQuantizer
from gap_quantization.dump_utils import dump_quant_params, remove_extra_dump, remove_cat_files
if args.quant_data_dir is None:
raise AttributeError('quant-data-dir argument is required.')
num_channels = 1 if args.grayscale else 3
cfg = {
"bits":
args.bits, # number of bits to store weights and activations
"accum_bits":
32, # number of bits to store intermediate convolution result
"signed": True, # use signed numbers
"save_folder": args.save_dir, # folder to save results
"data_source": args.
quant_data_dir, # folder with images to collect dataset statistics
"use_gpu": False, # use GPU for inference
"batch_size": 1,
"num_workers": 0, # number of workers for PyTorch dataloader
"verbose": True,
"save_params": args.
save_quantized_model, # save quantization parameters to the file
"quantize_forward":
True, # replace usual convs, poolings, ... with GAP-like ones
"num_input_channels": num_channels,
"raw_input": args.no_normalize,
"double_precision":
args.double_precision # use double precision convolutions
}
model = model.cpu()
quantizer = ModelQuantizer(
model, cfg, dm.transform_test
) # transform test is OK if we use args.no_normalize
quantizer.quantize_model(
) # otherwise we need to add QuantizeInput operation
if args.infer:
if args.image_path == '':
raise AttributeError('Image for inference is required')
quantizer.dump_activations(args.image_path,
dm.transform_test,
save_dir=os.path.join(
args.save_dir, 'activations_dump'))
dump_quant_params(args.save_dir, args.convbn)
if args.convbn:
remove_extra_dump(
os.path.join(args.save_dir, 'activations_dump'))
remove_cat_files(args.save_dir)
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
for name in args.target_names:
if not 'lfw' in name.lower():
print("Evaluating {} ...".format(name))
queryloader = testloader_dict[name]['query']
galleryloader = testloader_dict[name]['gallery']
distmat = test(args,
model,
queryloader,
galleryloader,
use_gpu,
return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(
distmat,
dm.return_testdataset_by_name(name),
save_dir=osp.join(args.save_dir, 'ranked_results',
name),
topk=20)
else:
model.eval()
same_acc, diff_acc, all_acc, auc, thresh = evaluate(
args,
dm.lfw_dataset,
model,
compute_embeddings_lfw,
args.test_batch_size,
verbose=False,
show_failed=args.show_failed,
load_embeddings=args.load_embeddings)
log.info('Validation accuracy: {0:.4f}, {1:.4f}'.format(
same_acc, diff_acc))
log.info('Validation accuracy mean: {0:.4f}'.format(all_acc))
log.info('Validation AUC: {0:.4f}'.format(auc))
log.info('Estimated threshold: {0:.4f}'.format(thresh))
#roc_auc(model, '/home/maxim/data/lfw/pairsTest.txt', '/media/slow_drive/cropped_lfw', args, use_gpu)
return
def main():
torch.manual_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:
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_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.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={'xent'},
use_gpu=use_gpu)
print("Model size: {:.3f} M".format(count_num_param(model)))
if args.label_smooth:
criterion = CrossEntropyLabelSmooth(num_classes=dataset.num_train_pids,
use_gpu=use_gpu)
else:
criterion = nn.CrossEntropyLoss()
optimizer = init_optim(args.optim, model.parameters(), args.lr,
args.weight_decay)
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.evaluate:
print("Evaluate only")
distmat = test(model,
queryloader,
galleryloader,
use_gpu,
return_distmat=True)
if args.vis_ranked_res:
visualize_ranked_results(
distmat,
dataset,
save_dir=osp.join(args.save_dir, 'ranked_results'),
topk=20,
)
return
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,
freeze_bn=True)
train_time += round(time.time() - start_train_time)
del optimizer_tmp
print("Now open all layers for training")
for epoch in range(args.start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch, model, criterion, optimizer, trainloader, use_gpu)
train_time += round(time.time() - start_train_time)
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))
def main():
global best_rank1, best_mAP
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_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'.format(time.strftime('-%Y-%m-%d-%H-%M-%S'))))
else:
sys.stdout = Logger(
osp.join(
args.save_dir,
'log_test{}.txt'.format(time.strftime('-%Y-%m-%d-%H-%M-%S'))))
writer = SummaryWriter(log_dir=args.save_dir, comment=args.arch)
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = False if 'resnet3dt' in args.arch else 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_vidreid_dataset(root=args.root,
name=args.dataset,
split_id=args.split_id,
use_pose=args.use_pose)
transform_train = list()
print('Transform:')
if args.misalign_aug:
print('+ Misalign Augmentation')
transform_train.append(T.GroupMisAlignAugment())
if args.rand_crop:
print('+ Random Crop')
transform_train.append(T.GroupRandomCrop(size=(240, 120)))
print('+ Resize to ({} x {})'.format(args.height, args.width))
transform_train.append(T.GroupResize((args.height, args.width)))
if args.flip_aug:
print('+ Random HorizontalFlip')
transform_train.append(T.GroupRandomHorizontalFlip())
print('+ ToTensor')
transform_train.append(T.GroupToTensor())
print(
'+ Normalize with mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]'
)
transform_train.append(
T.GroupNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224,
0.225]))
if args.rand_erase:
print('+ Random Erasing')
transform_train.append(T.GroupRandomErasing())
transform_train = T.Compose(transform_train)
transform_test = T.Compose([
T.GroupResize((args.height, args.width)),
T.GroupToTensor(),
T.GroupNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224,
0.225]),
])
pin_memory = True if use_gpu else False
trainloader = DataLoader(
VideoDataset(dataset.train,
seq_len=args.seq_len,
sample=args.train_sample,
transform=transform_train,
training=True,
pose_info=dataset.process_poses,
num_split=args.num_split,
num_parts=args.num_parts,
num_scale=args.num_scale,
pyramid_part=args.pyramid_part,
enable_pose=args.use_pose),
sampler=eval(args.train_sampler)(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(
VideoDataset(dataset.query,
seq_len=args.seq_len,
sample=args.test_sample,
transform=transform_test,
pose_info=dataset.process_poses,
num_split=args.num_split,
num_parts=args.num_parts,
num_scale=args.num_scale,
pyramid_part=args.pyramid_part,
enable_pose=args.use_pose),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
seq_len=args.seq_len,
sample=args.test_sample,
transform=transform_test,
pose_info=dataset.process_poses,
num_split=args.num_split,
num_parts=args.num_parts,
num_scale=args.num_scale,
pyramid_part=args.pyramid_part,
enable_pose=args.use_pose),
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'},
last_stride=args.last_stride,
num_parts=args.num_parts,
num_scale=args.num_scale,
num_split=args.num_split,
pyramid_part=args.pyramid_part,
num_gb=args.num_gb,
use_pose=args.use_pose,
learn_graph=args.learn_graph,
consistent_loss=args.consistent_loss,
bnneck=args.bnneck,
save_dir=args.save_dir)
input_size = sum(calc_splits(
args.num_split)) if args.pyramid_part else args.num_split
input_size *= args.num_scale * args.seq_len
num_params, flops = compute_model_complexity(
model,
input=[
torch.randn(1, args.seq_len, 3, args.height, args.width),
torch.ones(1, input_size, input_size)
],
verbose=True,
only_conv_linear=False)
print('Model complexity: params={:,} flops={:,}'.format(num_params, flops))
if args.label_smooth:
criterion_xent = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
else:
criterion_xent = nn.CrossEntropyLoss()
criterion_htri = TripletLoss(margin=args.margin, soft=args.soft_margin)
param_groups = model.parameters()
optimizer = init_optim(args.optim, param_groups, args.lr,
args.weight_decay)
scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma)
if args.warmup:
scheduler = lr_scheduler.WarmupMultiStepLR(optimizer,
milestones=args.stepsize,
gamma=args.gamma,
warmup_iters=10,
warmup_factor=0.01)
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):
print("Loaded checkpoint from '{}'".format(args.resume))
from functools import partial
import pickle
pickle.load = partial(pickle.load, encoding="latin1")
pickle.Unpickler = partial(pickle.Unpickler, encoding="latin1")
checkpoint = torch.load(args.resume,
map_location=lambda storage, loc: storage,
pickle_module=pickle)
print('Loaded model weights')
model.load_state_dict(checkpoint['state_dict'])
if optimizer is not None and 'optimizer' in checkpoint:
print('Loaded optimizer')
optimizer.load_state_dict(checkpoint['optimizer'])
if use_gpu:
for state in optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.cuda()
start_epoch = checkpoint['epoch'] + 1
print('- start_epoch: {}'.format(start_epoch))
best_rank1 = checkpoint['rank1']
print("- rank1: {}".format(best_rank1))
if 'mAP' in checkpoint:
best_mAP = checkpoint['mAP']
print("- mAP: {}".format(best_mAP))
else:
start_epoch = 0
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
distmat = test(model,
queryloader,
galleryloader,
args.pool,
use_gpu,
return_distmat=True)
if args.visualize_ranks:
visualize_ranked_results(
distmat,
dataset,
save_dir=osp.join(args.save_dir, 'ranked_results'),
topk=20,
)
return
start_time = time.time()
train_time = 0
best_epoch = start_epoch
print("==> Start training")
for epoch in range(start_epoch, args.max_epoch):
start_train_time = time.time()
train(epoch,
model,
criterion_xent,
criterion_htri,
optimizer,
trainloader,
use_gpu,
writer=writer)
train_time += round(time.time() - start_train_time)
if epoch >= args.zero_wd > 0:
set_wd(optimizer, 0)
for group in optimizer.param_groups:
assert group['weight_decay'] == 0, '{} is not zero'.format(
group['weight_decay'])
scheduler.step(epoch)
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, mAP = test(model, queryloader, galleryloader, args.pool,
use_gpu)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_mAP = mAP
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,
'optimizer': optimizer.state_dict(),
'rank1': rank1,
'mAP': mAP,
'epoch': epoch,
}, False,
osp.join(args.save_dir,
'checkpoint_ep' + str(epoch + 1) + '.pth.tar'))
writer.add_scalar(tag='acc/rank1',
scalar_value=rank1,
global_step=epoch + 1)
writer.add_scalar(tag='acc/mAP',
scalar_value=mAP,
global_step=epoch + 1)
print("==> Best Rank-1 {:.2%}, mAP: {:.2%}, achieved at epoch {}".format(
best_rank1, best_mAP, 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))
