def build_data_loader():
logger.info("build train dataset")
# dataset
dataset = init_dataset(cfg.TRAIN.DATASET, cfg.TRAIN.TXT)
sampler = RandomIdentitySampler(dataset.train, cfg.TRAIN.NUM_IDENTITIES)
train_loader = DataLoader(ImageData_train(dataset.train,
TrainTransformer()),
batch_size=cfg.TRAIN.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKERS,
pin_memory=True,
sampler=sampler)
query_loader = DataLoader(ImageData_val(dataset.probe, TestTransformer()),
batch_size=cfg.TRAIN.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKERS,
pin_memory=True,
shuffle=False)
gallery_loader = DataLoader(ImageData_val(dataset.gallery,
TestTransformer()),
batch_size=cfg.TRAIN.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKERS,
pin_memory=True,
shuffle=False)
return dataset, train_loader, query_loader, gallery_loader
def load_data(dataset, pin_memory):
dataloader = {}
if opt.loss == 'softmax':
trainloader = DataLoader(ImageData(dataset.train, TrainTransform()),
shuffle=True,
batch_size=opt.train_batch,
num_workers=8,
pin_memory=pin_memory,
drop_last=True)
else:
trainloader = DataLoader(ImageData(dataset.train, TrainTransform()),
sampler=RandomIdentitySampler(
dataset.train, opt.train_batch,
opt.num_instances),
batch_size=opt.train_batch,
num_workers=8,
pin_memory=pin_memory,
drop_last=True)
dataloader['train'] = trainloader
queryloader = DataLoader(ImageData(dataset.query, TestTransform()),
batch_size=opt.test_batch,
num_workers=8,
pin_memory=pin_memory)
dataloader['query'] = queryloader
galleryloader = DataLoader(ImageData(dataset.gallery, TestTransform()),
batch_size=opt.test_batch,
num_workers=8,
pin_memory=pin_memory)
dataloader['gallery'] = galleryloader
queryFliploader = DataLoader(ImageData(dataset.query, TestTransform(True)),
batch_size=opt.test_batch,
num_workers=8,
pin_memory=pin_memory)
dataloader['queryFlip'] = queryFliploader
galleryFliploader = DataLoader(ImageData(dataset.gallery,
TestTransform(True)),
batch_size=opt.test_batch,
num_workers=8,
pin_memory=pin_memory)
dataloader['galleryFlip'] = galleryFliploader
return dataloader
def train(**kwargs):
opt._parse(kwargs)
opt.model_name = 'bfe_test'
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
os.makedirs(opt.save_dir, exist_ok=True)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset, mode=opt.mode)
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
trainloader = DataLoader(ImageData(dataset.train,
TrainTransform(opt.datatype)),
sampler=RandomIdentitySampler(
dataset.train, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
queryloader = DataLoader(ImageData(dataset.query,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryloader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryFliploader = DataLoader(ImageData(dataset.query,
TestTransform(opt.datatype, True)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryFliploader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.datatype,
True)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
print('initializing model ...')
model = BFE(dataset.num_train_pids, 1.0, 0.33)
optim_policy = model.get_optim_policy()
if opt.pretrained_model:
state_dict = torch.load(opt.pretrained_model)['state_dict']
# state_dict = {k: v for k, v in state_dict.items() \
# if not ('reduction' in k or 'softmax' in k)}
model.load_state_dict(state_dict, False)
print('load pretrained model ' + opt.pretrained_model)
print('model size: {:.5f}M'.format(
sum(p.numel() for p in model.parameters()) / 1e6))
if use_gpu:
model = nn.DataParallel(model).cuda()
reid_evaluator = ResNetEvaluator(model)
if opt.evaluate:
reid_evaluator.evaluate(queryloader,
galleryloader,
queryFliploader,
galleryFliploader,
re_ranking=opt.re_ranking,
savefig=opt.savefig)
return
# xent_criterion = nn.CrossEntropyLoss()
xent_criterion = CrossEntropyLabelSmooth(dataset.num_train_pids)
if opt.loss == 'triplet':
embedding_criterion = TripletLoss(opt.margin)
elif opt.loss == 'lifted':
embedding_criterion = LiftedStructureLoss(hard_mining=True)
elif opt.loss == 'weight':
embedding_criterion = Margin()
def criterion(triplet_y, softmax_y, labels):
losses = [embedding_criterion(output, labels)[0] for output in triplet_y] + \
[xent_criterion(output, labels) for output in softmax_y]
loss = sum(losses)
return loss
# get optimizer
if opt.optim == "sgd":
optimizer = torch.optim.SGD(optim_policy,
lr=opt.lr,
momentum=0.9,
weight_decay=opt.weight_decay)
else:
optimizer = torch.optim.Adam(optim_policy,
lr=opt.lr,
weight_decay=opt.weight_decay)
start_epoch = opt.start_epoch
# get trainer and evaluator
reid_trainer = cls_tripletTrainer(opt, model, optimizer, criterion,
summary_writer)
def adjust_lr(optimizer, ep):
if ep < 10:
lr = opt.lr * 0.1 * (ep / 10.0) # warm_up
elif ep < 50:
lr = opt.lr * (ep // 5 + 1)
elif ep < 200:
lr = opt.lr * 10.0
elif ep < 300:
lr = opt.lr
else:
lr = opt.lr * 0.1
for p in optimizer.param_groups:
p['lr'] = lr
# start training
best_rank1 = opt.best_rank
best_epoch = 0
for epoch in range(start_epoch, opt.max_epoch):
if opt.adjust_lr:
adjust_lr(optimizer, epoch + 1)
reid_trainer.train(epoch, trainloader)
# skip if not save model
if opt.eval_step > 0 and (epoch + 1) % opt.eval_step == 0 or (
epoch + 1) == opt.max_epoch:
if opt.mode == 'class':
rank1 = test(model, queryloader)
else:
rank1 = reid_evaluator.evaluate(queryloader, galleryloader,
queryFliploader,
galleryFliploader)
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,
'epoch': epoch + 1
},
is_best=is_best,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
print('Best rank-1 {:.1%}, achived at epoch {}'.format(
best_rank1, best_epoch))
def train(**kwargs):
opt._parse(kwargs)
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
os.makedirs(opt.save_dir, exist_ok=True)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing tx_chanllege dataset')
dataset = Tx_dataset(file_list='train_list_new.txt').dataset
query_dataset = Tx_dataset(set='train_set',
file_list='val_query_list.txt').dataset
gallery_dataset = Tx_dataset(set='train_set',
file_list='val_gallery_list.txt').dataset
train_set = ImageDataset(dataset,
transform=build_transforms(opt, is_train=True))
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
if opt.sampler_new:
trainloader = DataLoader(
train_set,
sampler=RandomIdentitySampler_new(train_set, opt.train_batch,
opt.num_instances),
# sampler=RandomIdentitySampler(train_set, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
else:
trainloader = DataLoader(
train_set,
# sampler=RandomIdentitySampler_new(train_set, opt.train_batch, opt.num_instances),
sampler=RandomIdentitySampler(train_set, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
queryloader = DataLoader(ImageDataset(query_dataset,
transform=build_transforms(
opt, is_train=False)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryloader = DataLoader(ImageDataset(gallery_dataset,
transform=build_transforms(
opt, is_train=False)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryFliploader = DataLoader(ImageDataset(query_dataset,
transform=build_transforms(
opt,
is_train=False,
flip=True)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryFliploader = DataLoader(ImageDataset(gallery_dataset,
transform=build_transforms(
opt,
is_train=False,
flip=True)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryCenterloader = DataLoader(ImageDataset(query_dataset,
transform=build_transforms(
opt,
is_train=False,
crop='center')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryCenterloader = DataLoader(ImageDataset(gallery_dataset,
transform=build_transforms(
opt,
is_train=False,
crop='center')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryLtloader = DataLoader(ImageDataset(
query_dataset,
transform=build_transforms(opt, is_train=False, crop='lt')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryLtloader = DataLoader(ImageDataset(gallery_dataset,
transform=build_transforms(
opt,
is_train=False,
crop='lt')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryRtloader = DataLoader(ImageDataset(
query_dataset,
transform=build_transforms(opt, is_train=False, crop='rt')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryRtloader = DataLoader(ImageDataset(gallery_dataset,
transform=build_transforms(
opt,
is_train=False,
crop='rt')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryRbloader = DataLoader(ImageDataset(
query_dataset,
transform=build_transforms(opt, is_train=False, crop='rb')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryRbloader = DataLoader(ImageDataset(gallery_dataset,
transform=build_transforms(
opt,
is_train=False,
crop='rb')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryLbloader = DataLoader(ImageDataset(
query_dataset,
transform=build_transforms(opt, is_train=False, crop='lb')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryLbloader = DataLoader(ImageDataset(gallery_dataset,
transform=build_transforms(
opt,
is_train=False,
crop='lb')),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
print('initializing model ...')
model = build_model(opt)
optim_policy = model.get_optim_policy()
if opt.pretrained_choice == 'self':
state_dict = torch.load(opt.pretrained_model)['state_dict']
# state_dict = {k: v for k, v in state_dict.items() \
# if not ('reduction' in k or 'softmax' in k)}
model.load_state_dict(state_dict, False)
print('load pretrained model ' + opt.pretrained_model)
print('model size: {:.5f}M'.format(
sum(p.numel() for p in model.parameters()) / 1e6))
if use_gpu:
model = nn.DataParallel(model).cuda()
reid_evaluator = Evaluator(model,
norm=opt.norm,
eval_flip=opt.eval_flip,
re_ranking=opt.re_ranking)
if opt.use_center:
criterion = make_loss_with_center(opt)
else:
criterion = make_loss(opt)
# get optimizer
if opt.optim == "sgd":
optimizer = torch.optim.SGD(optim_policy,
lr=opt.lr,
momentum=0.9,
weight_decay=opt.weight_decay)
else:
optimizer = torch.optim.Adam(optim_policy,
lr=opt.lr,
weight_decay=opt.weight_decay)
start_epoch = opt.start_epoch
# get trainer and evaluator
reid_trainer = cls_tripletTrainer(opt, model, optimizer, criterion,
summary_writer)
# start training
best_rank1 = opt.best_rank
best_epoch = 0
for epoch in range(start_epoch, opt.max_epoch):
if opt.adjust_lr:
adjust_lr(optimizer, opt.lr, opt.model_name, epoch + 1)
reid_trainer.train(epoch, trainloader)
# skip if not save model
if opt.eval_step > 0 and (epoch + 1) % opt.eval_step == 0 or (
epoch + 1) == opt.max_epoch:
rank1 = reid_evaluator.validation(
queryloader, galleryloader, queryFliploader, galleryFliploader,
queryCenterloader, galleryCenterloader, queryLtloader,
galleryLtloader, queryRtloader, galleryRtloader, queryLbloader,
galleryLbloader, queryRbloader, galleryRbloader)
print('start re_ranking......')
_ = reid_evaluator.validation(queryloader,
galleryloader,
queryFliploader,
galleryFliploader,
queryCenterloader,
galleryCenterloader,
queryLtloader,
galleryLtloader,
queryRtloader,
galleryRtloader,
queryLbloader,
galleryLbloader,
queryRbloader,
galleryRbloader,
re_ranking=True)
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,
'epoch': epoch + 1
},
is_best=is_best,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
print('Best rank-1 {:.1%}, achived at epoch {}'.format(
best_rank1, best_epoch))
def train(**kwargs):
opt._parse(kwargs)
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset)
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
if 'triplet' in opt.model_name:
trainloader = DataLoader(
ImageData(dataset.train, TrainTransform(opt.height, opt.width)),
sampler=RandomIdentitySampler(dataset.train, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
else:
trainloader = DataLoader(ImageData(
dataset.train, TrainTransform(opt.height, opt.width)),
batch_size=opt.train_batch,
shuffle=True,
num_workers=opt.workers,
pin_memory=pin_memory)
queryloader = DataLoader(ImageData(dataset.query,
TestTransform(opt.height, opt.width)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryloader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.height, opt.width)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
print('initializing model ...')
if opt.model_name == 'softmax' or opt.model_name == 'softmax_triplet':
model, optim_policy = get_baseline_model(dataset.num_train_pids)
elif opt.model_name == 'triplet':
model, optim_policy = get_baseline_model(num_classes=None)
print('model size: {:.5f}M'.format(
sum(p.numel() for p in model.parameters()) / 1e6))
# xent_criterion = nn.CrossEntropyLoss()
xent_criterion = CrossEntropyLabelSmooth(dataset.num_train_pids)
tri_criterion = TripletLoss(opt.margin)
def cls_criterion(cls_scores, targets):
cls_loss = xent_criterion(cls_scores, targets)
return cls_loss
def triplet_criterion(feat, targets):
triplet_loss, _, _ = tri_criterion(feat, targets)
return triplet_loss
def cls_tri_criterion(cls_scores, feat, targets):
cls_loss = xent_criterion(cls_scores, targets)
triplet_loss, _, _ = tri_criterion(feat, targets)
loss = cls_loss + triplet_loss
return loss
# get optimizer
optimizer = torch.optim.Adam(optim_policy,
lr=opt.lr,
weight_decay=opt.weight_decay)
def adjust_lr(optimizer, ep):
if ep < 20:
lr = 1e-4 * (ep + 1) / 2
elif ep < 80:
lr = 1e-3 * opt.num_gpu
elif ep < 180:
lr = 1e-4 * opt.num_gpu
elif ep < 300:
lr = 1e-5 * opt.num_gpu
elif ep < 320:
lr = 1e-5 * 0.1**((ep - 320) / 80) * opt.num_gpu
elif ep < 400:
lr = 1e-6
elif ep < 480:
lr = 1e-4 * opt.num_gpu
else:
lr = 1e-5 * opt.num_gpu
for p in optimizer.param_groups:
p['lr'] = lr
start_epoch = opt.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
# get trainer and evaluator
if opt.model_name == 'softmax':
reid_trainer = clsTrainer(opt, model, optimizer, cls_criterion,
summary_writer)
elif opt.model_name == 'softmax_triplet':
reid_trainer = cls_tripletTrainer(opt, model, optimizer,
cls_tri_criterion, summary_writer)
elif opt.model_name == 'triplet':
reid_trainer = tripletTrainer(opt, model, optimizer, triplet_criterion,
summary_writer)
reid_evaluator = ResNetEvaluator(model)
# start training
best_rank1 = -np.inf
best_epoch = 0
for epoch in range(start_epoch, opt.max_epoch):
if opt.step_size > 0:
adjust_lr(optimizer, epoch + 1)
reid_trainer.train(epoch, trainloader)
# skip if not save model
if opt.eval_step > 0 and (epoch + 1) % opt.eval_step == 0 or (
epoch + 1) == opt.max_epoch:
rank1 = reid_evaluator.evaluate(queryloader, galleryloader)
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,
'epoch': epoch + 1,
},
is_best=is_best,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
print('Best rank-1 {:.1%}, achived at epoch {}'.format(
best_rank1, best_epoch))
def train(**kwargs):
opt._parse(kwargs)
#opt.lr=0.00002
opt.model_name = 'AlignedReid'
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
os.makedirs(opt.save_dir, exist_ok=True)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset, mode=opt.mode)
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
trainloader = DataLoader(ImageData(dataset.train,
TrainTransform(opt.datatype)),
sampler=RandomIdentitySampler(
dataset.train, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
queryloader = DataLoader(ImageData(dataset.query,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryloader = DataLoader(ImageData(dataset.gallery,
TestTransform(opt.datatype)),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
queryFliploader = queryloader
galleryFliploader = galleryloader
# queryFliploader = DataLoader(
# ImageData(dataset.query, TestTransform(opt.datatype, True)),
# batch_size=opt.test_batch, num_workers=opt.workers,
# pin_memory=pin_memory
# )
#
# galleryFliploader = DataLoader(
# ImageData(dataset.gallery, TestTransform(opt.datatype, True)),
# batch_size=opt.test_batch, num_workers=opt.workers,
# pin_memory=pin_memory
# )
print('initializing model ...')
model = AlignedResNet50(num_classes=dataset.num_train_pids,
loss={'softmax', 'metric'},
aligned=True,
use_gpu=use_gpu)
optim_policy = model.get_optim_policy()
if opt.pretrained_model:
state_dict = torch.load(opt.pretrained_model)['state_dict']
# state_dict = {k: v for k, v in state_dict.items() \
# if not ('reduction' in k or 'softmax' in k)}
model.load_state_dict(state_dict, False)
print('load pretrained model ' + opt.pretrained_model)
print('model size: {:.5f}M'.format(
sum(p.numel() for p in model.parameters()) / 1e6))
if use_gpu:
model = nn.DataParallel(model).cuda()
reid_evaluator = AlignedEvaluator(model)
if opt.evaluate:
#rank1 = test(model, queryloader, galleryloader, use_gpu)
reid_evaluator.evaluate(queryloader,
galleryloader,
queryFliploader,
galleryFliploader,
re_ranking=opt.re_ranking,
savefig=opt.savefig,
test_distance='global')
return
# xent_criterion = nn.CrossEntropyLoss()
xent_criterion = CrossEntropyLabelSmooth(dataset.num_train_pids,
use_gpu=use_gpu)
embedding_criterion = TripletLossAlignedReID(margin=opt.margin)
# def criterion(triplet_y, softmax_y, labels):
# losses = [embedding_criterion(output, labels)[0] for output in triplet_y] + \
# [xent_criterion(output, labels) for output in softmax_y]
# loss = sum(losses)
# return loss
def criterion(outputs, features, local_features, labels):
if opt.htri_only:
if isinstance(features, tuple):
global_loss, local_loss = DeepSupervision(
embedding_criterion, features, labels)
else:
global_loss, local_loss = embedding_criterion(
features, labels, local_features)
else:
if isinstance(outputs, tuple):
xent_loss = DeepSupervision(xent_criterion, outputs, labels)
else:
xent_loss = xent_criterion(outputs, labels)
if isinstance(features, tuple):
global_loss, local_loss = DeepSupervision(
embedding_criterion, features, labels)
else:
global_loss, local_loss = embedding_criterion(
features, labels, local_features)
loss = xent_loss + global_loss + local_loss
return loss
# get optimizer
if opt.optim == "sgd":
optimizer = torch.optim.SGD(optim_policy,
lr=opt.lr,
momentum=0.9,
weight_decay=opt.weight_decay)
else:
optimizer = torch.optim.Adam(optim_policy,
lr=opt.lr,
weight_decay=opt.weight_decay)
start_epoch = opt.start_epoch
# get trainer and evaluator
reid_trainer = AlignedTrainer(opt, model, optimizer, criterion,
summary_writer)
def adjust_lr(optimizer, ep):
if ep < 50:
lr = opt.lr * (ep // 5 + 1)
elif ep < 200:
lr = opt.lr * 10
elif ep < 300:
lr = opt.lr
else:
lr = opt.lr * 0.1
for p in optimizer.param_groups:
p['lr'] = lr
# start training
best_rank1 = opt.best_rank
best_epoch = 0
print('start train......')
for epoch in range(start_epoch, opt.max_epoch):
if opt.adjust_lr:
adjust_lr(optimizer, epoch + 1)
reid_trainer.train(epoch, trainloader)
# skip if not save model
if opt.eval_step > 0 and (epoch + 1) % opt.eval_step == 0 or (
epoch + 1) == opt.max_epoch:
# just avoid out of memory during eval,and can't save the model
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'epoch': epoch + 1
},
is_best=0,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
if opt.mode == 'class':
rank1 = test(model, queryloader)
else:
rank1 = reid_evaluator.evaluate(queryloader, galleryloader,
queryFliploader,
galleryFliploader)
#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()
if is_best:
save_checkpoint({
'state_dict': state_dict,
'epoch': epoch + 1
},
is_best=is_best,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
print('Best rank-1 {:.1%}, achived at epoch {}'.format(
best_rank1, best_epoch))
def train(**kwargs):
opt._parse(kwargs)
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
# os.makedirs(opt.save_dir, exist_ok=True)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
print(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset, mode=opt.mode)
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
trainloader = DataLoader(
ImageData(dataset.train, TrainTransform(opt.datatype)),
sampler=RandomIdentitySampler(dataset.train, opt.num_instances),
batch_size=opt.train_batch, num_workers=opt.workers,
pin_memory=pin_memory, drop_last=True
)
queryloader = DataLoader(
ImageData(dataset.query, TestTransform(opt.datatype)),
batch_size=opt.test_batch, num_workers=opt.workers,
pin_memory=pin_memory
)
galleryloader = DataLoader(
ImageData(dataset.gallery, TestTransform(opt.datatype)),
batch_size=opt.test_batch, num_workers=opt.workers,
pin_memory=pin_memory
)
queryFliploader = DataLoader(
ImageData(dataset.query, TestTransform(opt.datatype, True)),
batch_size=opt.test_batch, num_workers=opt.workers,
pin_memory=pin_memory
)
galleryFliploader = DataLoader(
ImageData(dataset.gallery, TestTransform(opt.datatype, True)),
batch_size=opt.test_batch, num_workers=opt.workers,
pin_memory=pin_memory
)
print('initializing model ...')
if opt.model_name == 'softmax' or opt.model_name == 'softmax_triplet':
model = ResNetBuilder(dataset.num_train_pids, 1, True)
elif opt.model_name == 'triplet':
model = ResNetBuilder(None, 1, True)
elif opt.model_name == 'CBDB':
if opt.datatype == "person":
model = CBDBdataset.num_train_pids, 1.0, 0.33)
else:
model = CBDB(dataset.num_train_pids, 0.5, 0.5)
elif opt.model_name == 'ide':
model = IDE(dataset.num_train_pids)
elif opt.model_name == 'resnet':
model = Resnet(dataset.num_train_pids)
optim_policy = model.get_optim_policy()
if opt.pretrained_model:
state_dict = torch.load(opt.pretrained_model)['state_dict']
#state_dict = {k: v for k, v in state_dict.items() \
# if not ('reduction' in k or 'softmax' in k)}
model.load_state_dict(state_dict, False)
print('load pretrained model ' + opt.pretrained_model)
print('model size: {:.5f}M'.format(sum(p.numel() for p in model.parameters()) / 1e6))
if use_gpu:
model = nn.DataParallel(model).cuda()
reid_evaluator = ResNetEvaluator(model)
if opt.evaluate:
reid_evaluator.evaluate(queryloader, galleryloader,
queryFliploader, galleryFliploader, re_ranking=opt.re_ranking, savefig=opt.savefig)
return
def train(**kwargs):
opt._parse(kwargs)
#opt.lr=0.00002
opt.model_name='PCB'
# set random seed and cudnn benchmark
torch.manual_seed(opt.seed)
os.makedirs(opt.save_dir, exist_ok=True)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset, mode=opt.mode)
tgt_dataset = data_manager.init_dataset(name=opt.tgt_dataset,mode=opt.mode)
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
trainloader = DataLoader(
ImageData(dataset.train, TrainTransform(opt.datatype)),
batch_size=opt.train_batch, num_workers=opt.workers,
pin_memory=pin_memory, drop_last=True
)
tgt_trainloader = DataLoader(
ImageData(tgt_dataset.train, TrainTransform(opt.datatype)),
batch_size=opt.train_batch,num_workers=opt.workers,
pin_memory=pin_memory,drop_last=True
)
tgt_queryloader = DataLoader(
ImageData(tgt_dataset.query, TestTransform(opt.datatype)),
batch_size=opt.test_batch, num_workers=opt.workers,
pin_memory=pin_memory
)
tgt_galleryloader = DataLoader(
ImageData(tgt_dataset.gallery, TestTransform(opt.datatype)),
batch_size=opt.test_batch, num_workers=opt.workers,
pin_memory=pin_memory
)
tgt_queryFliploader = DataLoader(
ImageData(tgt_dataset.query, TestTransform(opt.datatype, True)),
batch_size=opt.test_batch, num_workers=opt.workers,
pin_memory=pin_memory
)
tgt_galleryFliploader = DataLoader(
ImageData(tgt_dataset.gallery, TestTransform(opt.datatype, True)),
batch_size=opt.test_batch, num_workers=opt.workers,
pin_memory=pin_memory
)
print('initializing model ...')
model = PCB(dataset.num_train_pids)
optim_policy = model.get_optim_policy()
start_epoch = opt.start_epoch
if opt.pretrained_model:
checkpoint = torch.load(opt.pretrained_model)
state_dict = checkpoint['state_dict']
# state_dict = {k: v for k, v in state_dict.items() \
# if not ('reduction' in k or 'softmax' in k)}
try:
model.load_state_dict(state_dict, False)
print('load pretrained model ' + opt.pretrained_model)
except:
RuntimeError('please keep the same size with source dataset..')
else:
raise RuntimeError('please load a pre-trained model...')
print('model size: {:.5f}M'.format(sum(p.numel() for p in model.parameters()) / 1e6))
if use_gpu:
model = nn.DataParallel(model).cuda()
reid_evaluator = ResNetEvaluator(model)
if opt.evaluate:
print('transfer directly....... ')
reid_evaluator.evaluate(tgt_queryloader, tgt_galleryloader,
tgt_queryFliploader, tgt_galleryFliploader, re_ranking=opt.re_ranking, savefig=opt.savefig)
return
#xent_criterion = CrossEntropyLabelSmooth(dataset.num_train_pids)
embedding_criterion = SelfTraining_TripletLoss(margin=0.5,num_instances=4)
# def criterion(triplet_y, softmax_y, labels):
# losses = [embedding_criterion(output, labels)[0] for output in triplet_y] + \
# [xent_criterion(output, labels) for output in softmax_y]
# loss = sum(losses)
# return loss
def criterion(triplet_y, softmax_y, labels):
#losses = [torch.sum(torch.stack([xent_criterion(logits, labels) for logits in softmax_y]))]
losses = [torch.sum(torch.stack([embedding_criterion(output,labels) for output in triplet_y]))]
loss = sum(losses)
return loss
# get optimizer
if opt.optim == "sgd":
optimizer = torch.optim.SGD(optim_policy, lr=opt.lr, momentum=0.9, weight_decay=opt.weight_decay)
else:
optimizer = torch.optim.Adam(optim_policy, lr=opt.lr, weight_decay=opt.weight_decay)
# get trainer and evaluator
reid_trainer = PCBTrainer(opt, model, optimizer, criterion, summary_writer)
def adjust_lr(optimizer, ep):
if ep < 50:
lr = opt.lr * (ep // 5 + 1)
elif ep < 200:
lr = opt.lr*10
elif ep < 300:
lr = opt.lr
else:
lr = opt.lr*0.1
for p in optimizer.param_groups:
p['lr'] = lr
# start training
best_rank1 = opt.best_rank
best_epoch = 0
print('transfer directly.....')
reid_evaluator.evaluate(tgt_queryloader, tgt_galleryloader,
tgt_queryFliploader, tgt_galleryFliploader, re_ranking=opt.re_ranking, savefig=opt.savefig)
for iter_n in range(start_epoch,opt.max_epoch):
if opt.lambda_value == 0:
source_features = 0
else:
# get source datas' feature
print('Iteration {}: Extracting Source Dataset Features...'.format(iter_n + 1))
source_features, _ = extract_pcb_features(model, trainloader)
# extract training images' features
print('Iteration {}: Extracting Target Dataset Features...'.format(iter_n + 1))
target_features, _ = extract_pcb_features(model, tgt_trainloader)
# synchronization feature order with dataset.train
# calculate distance and rerank result
print('Calculating feature distances...')
target_features = target_features.numpy()
rerank_dist = re_ranking(
source_features, target_features, lambda_value=opt.lambda_value)
if iter_n == 0:
# DBSCAN cluster
tri_mat = np.triu(rerank_dist, 1) # tri_mat.dim=2 取上三角
tri_mat = tri_mat[np.nonzero(tri_mat)] # tri_mat.dim=1
tri_mat = np.sort(tri_mat, axis=None)
top_num = np.round(opt.rho * tri_mat.size).astype(int)
eps = tri_mat[:top_num].mean() # DBSCAN聚类半径
print('eps in cluster: {:.3f}'.format(eps))
cluster = DBSCAN(eps=eps, min_samples=4, metric='precomputed', n_jobs=8)
# select & cluster images as training set of this epochs
print('Clustering and labeling...')
labels = cluster.fit_predict(rerank_dist)
del(rerank_dist)
del(source_features)
del(target_features)
try:
gc.collect()
except:
print('cannot collect')
num_ids = len(set(labels)) - 1
print('Iteration {} have {} training ids'.format(iter_n + 1, num_ids))
# generate new dataset
new_dataset = []
for (fname, _, _), label in zip(tgt_dataset.train, labels):
if label == -1:
continue
# dont need to change codes in trainer.py _parsing_input function and sampler function after add 0
new_dataset.append((fname, label, 0))
print('Iteration {} have {} training images'.format(iter_n + 1, len(new_dataset)))
selftrain_loader = DataLoader(
ImageData(new_dataset, TrainTransform(opt.datatype)),
sampler=RandomIdentitySampler(new_dataset, opt.num_instances),
batch_size=opt.train_batch, num_workers=opt.workers,
pin_memory=pin_memory, drop_last=True
)
# train model with new generated dataset
trainer = PCBTrainer(opt, model, optimizer, criterion, summary_writer)
reid_evaluator = ResNetEvaluator(model)
# Start training
for epoch in range(opt.selftrain_iterations):
trainer.train(epoch, selftrain_loader)
# skip if not save model
if opt.eval_step > 0 and (iter_n + 1) % opt.eval_step == 0 or (iter_n + 1) == opt.max_epoch:
# just avoid out of memory during eval,and can't save the model
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({'state_dict': state_dict, 'epoch': iter_n + 1},
is_best=0, save_dir=opt.save_dir,
filename='checkpoint_ep' + str(iter_n + 1) + '.pth.tar')
if (iter_n + 1) % (opt.eval_step*4) == 0:
if opt.mode == 'class':
rank1 = test(model, tgt_queryloader)
else:
rank1 = reid_evaluator.evaluate(tgt_queryloader, tgt_galleryloader, tgt_queryFliploader,
tgt_galleryFliploader)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = iter_n + 1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
if is_best:
save_checkpoint({'state_dict': state_dict, 'epoch': iter_n + 1},
is_best=is_best, save_dir=opt.save_dir,
filename='checkpoint_ep' + str(iter_n + 1) + '.pth.tar')
print('Best rank-1 {:.1%}, achived at epoch {}'.format(best_rank1, best_epoch))
from datasets.data_loader import ImageData
from datasets.samplers import RandomIdentitySampler
from utils.serialization import Logger, save_checkpoint
from utils.transforms import TestTransform, TrainTransform
use_gpu = True
if __name__ == '__main__':
dataset = data_manager.init_dataset(name=opt.dataset, mode=opt.mode)
pin_memory = True if use_gpu else False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
trainloader = DataLoader(ImageData(dataset.train,
TrainTransform(opt.datatype)),
sampler=RandomIdentitySampler(
dataset.train, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True)
model = PlateNet(batch_size=opt.train_batch, n_class=66)
optim_policy = model.get_optim_policy()
if use_gpu:
model = nn.DataParallel(model).cuda()
print('dddddddddddd')
criterion = nn.CTCLoss()
# get optimizer
if opt.optim == "sgd":
def train(**kwargs):
opt._parse(kwargs)
# set random seed and cudnn benchmark
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
print('initializing dataset {}'.format(opt.dataset))
dataset = data_manager.init_dataset(name=opt.dataset)
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
if 'triplet' in opt.model_name:
trainloader = DataLoader(
ImageData(dataset.train, TrainTransform(opt.height, opt.width)),
sampler=RandomIdentitySampler(dataset.train, opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
last_batch='discard')
else:
trainloader = DataLoader(
ImageData(dataset.train, TrainTransform(opt.height, opt.width)),
batch_size=opt.train_batch,
shuffle=True,
num_workers=opt.workers,
)
queryloader = DataLoader(
ImageData(dataset.query, TestTransform(opt.height, opt.width)),
batch_size=opt.test_batch,
num_workers=opt.workers,
)
galleryloader = DataLoader(
ImageData(dataset.gallery, TestTransform(opt.height, opt.width)),
batch_size=opt.test_batch,
num_workers=opt.workers,
)
print('initializing model ...')
model = get_baseline_model(dataset.num_train_pids, mx.gpu(0),
opt.pretrained_model)
print('model size: {:.5f}M'.format(
sum(p.data().size for p in model.collect_params().values()) / 1e6))
xent_criterion = gluon.loss.SoftmaxCrossEntropyLoss()
tri_criterion = TripletLoss(opt.margin)
def cls_criterion(cls_scores, feat, targets):
cls_loss = xent_criterion(cls_scores, targets)
return cls_loss
def triplet_criterion(cls_scores, feat, targets):
triplet_loss, dist_ap, dist_an = tri_criterion(feat, targets)
return triplet_loss
def cls_tri_criterion(cls_scores, feat, targets):
cls_loss = xent_criterion(cls_scores, targets)
triplet_loss, dist_ap, dist_an = tri_criterion(feat, targets)
loss = cls_loss + triplet_loss
return loss
# get optimizer
optimizer = gluon.Trainer(model.collect_params(), opt.optim, {
'learning_rate': opt.lr,
'wd': opt.weight_decay
})
def adjust_lr(optimizer, ep):
if ep < 20:
lr = 1e-4 * (ep + 1) / 2
elif ep < 80:
lr = 1e-3 * opt.num_gpu
elif ep < 180:
lr = 1e-4 * opt.num_gpu
elif ep < 300:
lr = 1e-5 * opt.num_gpu
elif ep < 320:
lr = 1e-5 * 0.1**((ep - 320) / 80) * opt.num_gpu
elif ep < 400:
lr = 1e-6
elif ep < 480:
lr = 1e-4 * opt.num_gpu
else:
lr = 1e-5 * opt.num_gpu
optimizer.set_learning_rate(lr)
start_epoch = opt.start_epoch
# get trainer and evaluator
use_criterion = None
if opt.model_name == 'softmax':
use_criterion = cls_criterion
elif opt.model_name == 'softmax_triplet':
use_criterion = cls_tri_criterion
elif opt.model_name == 'triplet':
use_criterion = triplet_criterion
reid_trainer = reidTrainer(opt, model, optimizer, use_criterion,
summary_writer, mx.gpu(0))
reid_evaluator = reidEvaluator(model, mx.gpu(0))
# start training
best_rank1 = -np.inf
best_epoch = 0
for epoch in range(start_epoch, opt.max_epoch):
if opt.step_size > 0:
adjust_lr(optimizer, epoch + 1)
reid_trainer.train(epoch, trainloader)
# skip if not save model
if opt.eval_step > 0 and (epoch + 1) % opt.eval_step == 0 or (
epoch + 1) == opt.max_epoch:
rank1 = reid_evaluator.evaluate(queryloader, galleryloader)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
state_dict = {'model': model, 'epoch': epoch}
save_checkpoint(state_dict,
is_best=is_best,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.params')
print('Best rank-1 {:.1%}, achived at epoch {}'.format(
best_rank1, best_epoch))
def train(**kwargs):
#### Part 1 : Initialization
opt._parse(kwargs)
torch.backends.cudnn.deterministic = True
# set random seed and cudnn benchmark
#torch.manual_seed(opt.seed)
#random.seed(opt.seed)
#np.random.seed(opt.seed)
use_gpu = torch.cuda.is_available()
sys.stdout = Logger(osp.join(opt.save_dir, 'log_train.txt'))
print('=========user config==========')
pprint(opt._state_dict())
print('============end===============')
if use_gpu:
print('currently using GPU')
cudnn.benchmark = True
torch.cuda.manual_seed_all(opt.seed)
else:
print('currently using cpu')
#### Part 2 : Preparing Data
print('initializing train dataset {}'.format(opt.trainset))
train_dataset = data_manager.init_dataset(name=opt.trainset)
print('initializing test dataset {}'.format(opt.testset))
test_dataset = data_manager.init_dataset(name=opt.testset)
pin_memory = True if use_gpu else False
pin_memory = False
summary_writer = SummaryWriter(osp.join(opt.save_dir, 'tensorboard_log'))
collateFn = NormalCollateFn()
if opt.sampler == "randomidentity":
trainloader = DataLoader(
data_manager.init_datafolder(
opt.trainset,
train_dataset.train,
TrainTransform(opt.height,
opt.width,
random_erase=opt.with_randomerase),
if_train=True),
sampler=RandomIdentitySampler(train_dataset.train,
opt.num_instances),
batch_size=opt.train_batch,
num_workers=opt.workers,
pin_memory=pin_memory,
drop_last=True,
collate_fn=collateFn,
)
elif opt.sampler == "randomidentitycamera":
trainloader = DataLoader(
data_manager.init_datafolder(
opt.trainset,
train_dataset.train,
TrainTransform(opt.height,
opt.width,
random_erase=opt.with_randomerase),
if_train=True),
batch_sampler=RandomIdentityCameraSampler(train_dataset.train,
opt.num_instances,
opt.train_batch),
num_workers=opt.workers,
pin_memory=pin_memory,
collate_fn=collateFn,
)
queryloader = DataLoader(data_manager.init_datafolder(opt.testset,
test_dataset.query,
TestTransform(
opt.height,
opt.width),
if_train=False),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
galleryloader = DataLoader(data_manager.init_datafolder(
opt.testset,
test_dataset.gallery,
TestTransform(opt.height, opt.width),
if_train=False),
batch_size=opt.test_batch,
num_workers=opt.workers,
pin_memory=pin_memory)
#### Part 3 : Preparing Backbone Network
print('initializing model ...')
if opt.model_name in ['triplet', 'distance']:
model, optim_policy = get_baseline_model(num_classes=None,
model='triplet')
elif opt.model_name in ["softmax"]:
model, optim_policy = get_baseline_model(train_dataset.num_train_pids,
model='softmax',
drop_prob=opt.drop)
else:
assert False, "unknown model name"
if (not opt.model_path == 'zero') and 'tar' in opt.model_path:
print('load pretrain reid model......' + opt.model_path)
ckpt = torch.load(opt.model_path)
# remove classifer
tmp = dict()
for k, v in ckpt['state_dict'].items():
if opt.keep_layer:
for i in opt.keep_layer:
if 'layer' + str(i) in k:
#print(k+" skip....")
continue
if opt.keepfc or ('fc' not in k and 'classifier' not in k):
tmp[k] = v
ckpt['state_dict'] = tmp
model.load_state_dict(ckpt['state_dict'], strict=False)
print('model size: {:.5f}M'.format(
sum(p.numel() for p in model.parameters()) / 1e6))
#### Part 4: Preparing Loss Functions
if opt.margin1 is not None:
distance_loss = DistanceLoss(margin=(opt.margin1, opt.margin2))
else:
distance_loss = DistanceLoss()
tri_loss = TripletLoss(margin=opt.margin)
xent_loss = nn.CrossEntropyLoss()
vis = dict()
vis['tri_acc1'] = AverageMeter()
vis['tri_acc2'] = AverageMeter()
vis['cls_accuracy'] = AverageMeter()
vis['cls_loss'] = AverageMeter()
def dist_criterion(feat,
targets,
cameras,
model=None,
paths=None,
epoch=0):
dis_loss, tri_acc1, tri_acc2 = distance_loss(feat,
targets,
cameras,
model,
paths,
epoch=epoch)
vis['tri_acc1'].update(float(tri_acc1))
vis['tri_acc2'].update(float(tri_acc2))
return dis_loss
def triplet_criterion(feat, targets):
triplet_loss, tri_accuracy, _, _ = tri_loss(feat, targets)
vis['tri_acc1'].update(float(tri_accuracy))
return triplet_loss
def cls_criterion(cls_scores, targets):
cls_loss = xent_loss(cls_scores, targets)
_, preds = torch.max(cls_scores.data, 1)
corrects = float(torch.sum(preds == targets.data))
vis['cls_accuracy'].update(float(corrects / opt.train_batch))
vis['cls_loss'].update(float(cls_loss))
return cls_loss
#### Part 5: Preparing Optimizer and Trainer
optimizer, adjust_lr = get_optimizer_strategy(opt.model_name, optim_policy,
opt)
start_epoch = opt.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
#model=model.cuda()
# get trainer and evaluatori
if opt.model_name == "distance":
reid_trainer = tripletTrainer(opt,
model,
optimizer,
dist_criterion,
summary_writer,
need_cam=True)
elif opt.model_name == 'triplet' or opt.model_name == 'triplet_fc':
reid_trainer = tripletTrainer(opt, model, optimizer, triplet_criterion,
summary_writer)
elif opt.model_name == 'softmax':
reid_trainer = clsTrainer(opt, model, optimizer, cls_criterion,
summary_writer)
else:
print("Error: Unknown model name {}".format(opt.model_name))
reid_evaluator = evaluator_manager.init_evaluator(opt.testset,
model,
flip=True)
#### Part 6 : Training
best_rank1 = -np.inf
best_epoch = 0
for epoch in range(start_epoch, opt.max_epoch):
if opt.step_size > 0:
current_lr = adjust_lr(optimizer, epoch)
reid_trainer.train(epoch, trainloader)
for k, v in vis.items():
print("{}:{}".format(k, v.mean))
v.reset()
if (epoch + 1) == opt.max_epoch:
if use_gpu and opt.num_gpu > 1:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'epoch': epoch + 1,
},
is_best=False,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
# skip if not save model
if (opt.eval_step > 0 and
(epoch + 1) % opt.eval_step == 0 and epoch >= 0
or (epoch + 1) == opt.max_epoch):
#print('Test on '+opt.testset)
#rank1 = reid_evaluator.evaluate(queryloader, galleryloader,normalize=opt.with_normalize)
print('Test on ' + opt.trainset)
if use_gpu and opt.num_gpu > 1:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
'epoch': epoch + 1,
},
is_best=False,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) +
'.pth.tar')
rank1, mAP = reid_evaluator.evaluate(queryloader,
galleryloader,
normalize=opt.with_normalize)
is_best = rank1 > best_rank1
if is_best:
best_rank1 = rank1
best_epoch = epoch + 1
save_checkpoint(
{
'state_dict': state_dict,
'epoch': epoch + 1,
},
is_best=False,
save_dir=opt.save_dir,
filename='checkpoint_ep' + str(epoch + 1) + '.pth.tar')
print('Best rank-1 {:.1%}, achieved at epoch {}'.format(
best_rank1, best_epoch))