def create_model(args):
model_1 = models.create(args.arch,
num_features=args.features,
dropout=args.dropout,
num_classes=1)
model_1_ema = models.create(args.arch,
num_features=args.features,
dropout=args.dropout,
num_classes=1)
model_1.cuda()
model_1_ema.cuda()
model_1 = nn.DataParallel(model_1)
model_1_ema = nn.DataParallel(model_1_ema)
if args.no_source:
print('No source pre-training')
else:
initial_weights = load_checkpoint(args.init_1)
copy_state_dict(initial_weights['state_dict'], model_1)
copy_state_dict(initial_weights['state_dict'], model_1_ema)
for param in model_1_ema.parameters():
param.detach_()
return model_1, model_1_ema
def create_model(args, classes, phase):
'''
创建模型，新建模型与保存的模型结构相同，以便可以进行载入权重
:param args:
:param classes:
:param phase:
:return:
'''
model_ema = models.create(args.arch,
num_features=args.features,
dropout=args.dropout,
num_classes=classes)
initial_weights = load_checkpoint(
osp.join(args.init, 'phase{}_model_best.pth.tar'.format(phase)))
if 'module.classifier.fc1.weight' in initial_weights['state_dict'].keys():
in_features = initial_weights['state_dict'][
'module.classifier.fc1.weight'].data.size(1)
out_features1 = initial_weights['state_dict'][
'module.classifier.fc1.weight'].data.size(0)
out_features2 = initial_weights['state_dict'][
'module.classifier.fc2.weight'].data.size(0)
new_fc = SplitCosineLinear(in_features,
out_features1,
out_features2,
sigma=True)
model_ema.classifier = new_fc
new_fc_max = copy.deepcopy(new_fc)
model_ema.classifier_max = new_fc_max
else:
in_features = initial_weights['state_dict'][
'module.classifier.weight'].data.size(1)
out_features = initial_weights['state_dict'][
'module.classifier.weight'].data.size(0)
new_fc = CosineLinear(in_features=in_features,
out_features=out_features,
sigma=True)
model_ema.classifier = new_fc
new_fc_max = copy.deepcopy(new_fc)
model_ema.classifier_max = new_fc_max
copy_state_dict(initial_weights['state_dict'], model_ema, strip='module.')
model_cur = copy.deepcopy(model_ema)
model_ref = copy.deepcopy(model_ema)
model_ema.cuda()
model_ema = nn.DataParallel(model_ema)
model_cur.cuda()
model_cur = nn.DataParallel(model_cur)
model_ref.cuda()
model_ref = nn.DataParallel(model_ref)
for param in model_ema.parameters():
param.detach_()
return model_cur, model_ema, model_ref
def main():
args = parser.parse_args()
if args.seed is not None:
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
cudnn.deterministic = True
########################################
cudnn.benchmark = True
sys.stdout = Logger(osp.join(args.logs_dir, 'log.txt'))
print("==========\nArgs:{}\n==========".format(args))
########处理dataset###########
dataset_target = get_data(args.dataset_target, args.data_dir)
# 创建测试使用的dataloader
test_loader_target = get_test_loader(dataset_target, args.height,
args.width, 256,
args.workers) #args.batch_size * 4
# 创建数据ID流
order_name = "./checkpoint/seed_{}_{}_order.pkl".format(
args.seed, args.dataset_target)
print("Order name:{}".format(order_name))
if os.path.exists(order_name):
print("Loading orders")
order = unpickle(order_name)
else:
print("Generating orders")
order = np.arange(dataset_target.num_train_pids)
np.random.shuffle(order)
savepickle(order, order_name)
order_list = list(order)
print(order_list)
init_class_num = dataset_target.num_train_pids - (args.num_phase -
1) * args.nb_cl
start_phase = 1
for phase in range(start_phase, args.num_phase):
global best_mAP
best_mAP = 0
if phase > 0:
proto_dataset = unpickle(
osp.join(args.logs_dir, 'phase{}_proto.pkl'.format(phase - 1)))
else:
proto_dataset = []
old_class_num = get_old_class_num(proto_dataset)
print('\n\n phase {} have {} old classes'.format(phase, old_class_num))
# 每个周期都需要载入最佳模型模型。
model_cur, model_ema, model_ref = create_model(args, 100, phase)
evaluator_ema = Evaluator(model_ema)
# 更新 input dataset
if phase == 0:
iters = 200
model_ref = None
input_id = order_list[0:init_class_num]
input_dataset = [(fname, pid, cam)
for (fname, pid, cam) in dataset_target.train
if pid in input_id]
print('phase:{} input id:{},input image:{}'.format(
phase, init_class_num, len(input_dataset)))
else:
iters = args.iters
input_id = order_list[init_class_num +
(phase - 1) * args.nb_cl:init_class_num +
phase * args.nb_cl]
input_dataset = [(fname, pid, cam)
for (fname, pid, cam) in dataset_target.train
if pid in input_id]
print('phase:{} input id:{},input image:{}'.format(
phase, args.nb_cl, len(input_dataset)))
tar_cluster_loader = get_test_loader(dataset_target,
args.height,
args.width,
256,
workers=args.workers,
testset=sorted(input_dataset))
for epoch in range(args.epochs):
dict_f, _, dic_logit = extract_features(model_ema,
tar_cluster_loader,
print_freq=40)
cf = torch.stack(list(dict_f.values())) #已经经过normalize
agent_sim = torch.stack(list(dic_logit.values()))
agent_sim = agent_sim[:, :
old_class_num] # probs dim=1,是否与old_num相同。
agent_sim = F.softmax(agent_sim, dim=1)
agent_sim_dist = torch.cdist(agent_sim, agent_sim, p=1) / 2
agent_sim_dist = agent_sim_dist.numpy()
rerank_dist = compute_jaccard_dist(
cf, use_gpu=True).numpy() # 经过rerank距离在0到1之间 args.rr_gpu
lambda_a = 0.3 if phase == 0 else 0
total_dist = (1 -
lambda_a) * rerank_dist + lambda_a * agent_sim_dist
if (epoch == 0):
#DBSCAN cluster
tri_mat = np.triu(rerank_dist, 1) # tri_mat.dim=2
tri_mat = tri_mat[np.nonzero(tri_mat)] # tri_mat.dim=1
tri_mat = np.sort(tri_mat, axis=None)
rho = 2e-3
top_num = np.round(rho * tri_mat.size).astype(int)
eps = tri_mat[:top_num].mean()
print('eps for cluster: {:.3f}'.format(eps))
cluster = DBSCAN(eps=eps,
min_samples=4,
metric='precomputed',
n_jobs=-1)
print('Clustering and labeling...')
labels = cluster.fit_predict(total_dist)
num_ids = len(set(labels)) - (1 if -1 in labels else 0)
args.num_clusters = num_ids
print('\n Clustered into {} classes \n'.format(args.num_clusters))
# generate new dataset and calculate cluster centers
new_dataset = []
cluster_centers = collections.defaultdict(list)
for i, ((fname, _, cid),
label) in enumerate(zip(sorted(input_dataset), labels)):
if label == -1:
continue
new_dataset.append((fname, label + old_class_num, cid))
if label == -1:
print('error')
cluster_centers[label].append(cf[i])
cluster_centers = [
torch.stack(cluster_centers[idx]).mean(0)
for idx in sorted(cluster_centers.keys())
]
cluster_centers = torch.stack(cluster_centers)
cluster_centers_avg = F.normalize(cluster_centers[:, :2048],
dim=1).float().cuda()
cluster_centers_max = F.normalize(cluster_centers[:, 2048:],
dim=1).float().cuda()
# 新的类别为num_ids，开始更新模型分类器系数 consineLinear
# 在每个epoch中不断更新 cur 和 ema的 new_fc层
in_features = model_ema.module.classifier.in_features
if phase == 0:
out_features = args.num_clusters
#创建new_fc_avg 为 model_cur和model_ema的avg赋值。
new_fc_avg = CosineLinear(in_features=in_features,
out_features=out_features,
sigma=True).cuda()
print("in_features:", in_features, "out_features:",
out_features)
# Evaluator
new_fc_avg.weight.data.copy_(cluster_centers_avg)
#new_fc.sigma.data = (model_ema.module.classifier.sigma.data+model_ema.module.classifier_max.sigma.data)/2 后面sigma 按20来处理
model_ema.module.classifier = new_fc_avg
model_cur.module.classifier = copy.deepcopy(new_fc_avg)
# 创建new_fc_max 为 model_cur和model_ema的classifier_max赋值。
new_fc_max = CosineLinear(in_features=in_features,
out_features=out_features,
sigma=True).cuda()
new_fc_max.weight.data.copy_(cluster_centers_max)
model_ema.module.classifier_max = new_fc_max
model_cur.module.classifier_max = copy.deepcopy(new_fc_max)
cur_lamda = 0
elif phase == 1:
############################################################
# increment classe
if epoch == 0:
out_features = model_ema.module.classifier.out_features
print("in phase 1:epoch 0 : in_features:", in_features,
"out_features:", out_features)
new_fc_cur = SplitCosineLinear(in_features, out_features,
args.num_clusters).cuda()
new_fc_cur.fc1.weight.data = model_cur.module.classifier.weight.data #shiyong id()lai panduan shifou tongyidizhi
new_fc_cur.fc2.weight.data.copy_(cluster_centers_avg)
model_cur.module.classifier = new_fc_cur
new_fc_cur_max = SplitCosineLinear(
in_features, out_features, args.num_clusters).cuda()
new_fc_cur_max.fc1.weight.data = model_cur.module.classifier_max.weight.data
new_fc_cur_max.fc2.weight.data.copy_(cluster_centers_max)
model_cur.module.classifier_max = new_fc_cur_max
new_fc_ema = SplitCosineLinear(in_features, out_features,
args.num_clusters).cuda()
new_fc_ema.fc1.weight.data = model_ema.module.classifier.weight.data
new_fc_ema.fc2.weight.data.copy_(cluster_centers_avg)
model_ema.module.classifier = new_fc_ema
new_fc_ema_max = SplitCosineLinear(
in_features, out_features, args.num_clusters).cuda()
new_fc_ema_max.fc1.weight.data = model_ema.module.classifier_max.weight.data
new_fc_ema_max.fc2.weight.data.copy_(cluster_centers_max)
model_ema.module.classifier_max = new_fc_ema_max
else:
out_features = model_ema.module.classifier.fc1.out_features
new_ema_fc = CosineLinear(in_features,
args.num_clusters,
sigma=False).cuda()
new_ema_fc.weight.data.copy_(cluster_centers_avg)
model_ema.module.classifier.fc2 = new_ema_fc
new_fc_cur = copy.deepcopy(new_ema_fc)
model_cur.module.classifier.fc2 = new_fc_cur
new_ema_fc_max = CosineLinear(in_features,
args.num_clusters,
sigma=False).cuda()
new_ema_fc_max.weight.data.copy_(cluster_centers_max)
model_ema.module.classifier_max.fc2 = new_ema_fc_max
new_fc_cur_max = copy.deepcopy(new_ema_fc_max)
model_cur.module.classifier_max.fc2 = new_fc_cur_max
lamda_mult = out_features * 1.0 / args.num_clusters # class_old / class_new
cur_lamda = args.lamda * math.sqrt(lamda_mult)
print("###############################")
print("Lamda for less forget is set to ", cur_lamda)
print("###############################")
assert model_ema.module.classifier.fc1.weight.data.size(
0) == old_class_num
assert model_ema.module.classifier.fc2.weight.data.size(
0) == args.num_clusters
else:
if epoch == 0:
out_features1 = model_ema.module.classifier.fc1.out_features
out_features2 = model_ema.module.classifier.fc2.out_features
out_features = out_features1 + out_features2
print("in_features:", in_features, "out_features1:", \
out_features1, "out_features2:", out_features2)
new_fc_cur = SplitCosineLinear(in_features, out_features,
args.num_clusters).cuda()
new_fc_cur.fc1.weight.data[:out_features1].copy_(
model_cur.module.classifier.fc1.weight.data)
new_fc_cur.fc1.weight.data[out_features1:].copy_(
model_cur.module.classifier.fc2.weight.data)
new_fc_cur.fc2.weight.data.copy_(cluster_centers_avg)
#new_fc_cur.sigma.data = model_cur.module.classifier.sigma.data
model_cur.module.classifier = new_fc_cur
new_fc_cur_max = SplitCosineLinear(
in_features, out_features, args.num_clusters).cuda()
new_fc_cur_max.fc1.weight.data[:out_features1].copy_(
model_cur.module.classifier_max.fc1.weight.data)
new_fc_cur_max.fc1.weight.data[out_features1:].copy_(
model_cur.module.classifier_max.fc2.weight.data)
new_fc_cur_max.fc2.weight.data.copy_(cluster_centers_max)
#new_fc_cur_max.sigma.data = model_cur.module.classifier_max.sigma.data
model_cur.module.classifier_max = new_fc_cur_max
new_fc_ema = SplitCosineLinear(in_features, out_features,
args.num_clusters).cuda()
new_fc_ema.fc1.weight.data[:out_features1].copy_(
model_ema.module.classifier.fc1.weight.data)
new_fc_ema.fc1.weight.data[out_features1:].copy_(
model_ema.module.classifier.fc2.weight.data)
new_fc_ema.fc2.weight.data.copy_(cluster_centers_avg)
#new_fc_ema.sigma.data = model_ema.module.classifier.sigma.data
model_ema.module.classifier = new_fc_ema
new_fc_ema_max = SplitCosineLinear(
in_features, out_features, args.num_clusters).cuda()
new_fc_ema_max.fc1.weight.data[:out_features1].copy_(
model_ema.module.classifier_max.fc1.weight.data)
new_fc_ema_max.fc1.weight.data[out_features1:].copy_(
model_ema.module.classifier_max.fc2.weight.data)
new_fc_ema_max.fc2.weight.data.copy_(cluster_centers_max)
#new_fc_ema_max.sigma.data = model_ema.module.classifier_max.sigma.data
model_ema.module.classifier_max = new_fc_ema_max
else:
out_features = model_ema.module.classifier.fc1.out_features
new_ema_fc = CosineLinear(in_features,
args.num_clusters,
sigma=False).cuda()
new_ema_fc.weight.data.copy_(cluster_centers_avg)
model_ema.module.classifier.fc2 = new_ema_fc
new_fc_cur = copy.deepcopy(new_ema_fc)
model_cur.module.classifier.fc2 = new_fc_cur
new_ema_fc_max = CosineLinear(in_features,
args.num_clusters,
sigma=False).cuda()
new_ema_fc_max.weight.data.copy_(cluster_centers_max)
model_ema.module.classifier_max.fc2 = new_ema_fc_max
new_fc_cur_max = copy.deepcopy(new_ema_fc_max)
model_cur.module.classifier_max.fc2 = new_fc_cur_max
lamda_mult = (out_features) * 1.0 / (args.nb_cl)
cur_lamda = args.lamda * math.sqrt(lamda_mult)
print("###############################")
print("Lamda for less forget is set to ", cur_lamda)
print("###############################")
assert model_ema.module.classifier.fc1.weight.data.size(
0) == old_class_num
assert model_ema.module.classifier.fc2.weight.data.size(
0) == args.num_clusters
# 生成相应的trainloader & optimizer
# 设置optimizer.
params = []
for key, value in model_cur.named_parameters():
if not value.requires_grad:
continue
if key == 'module.classifier.sigma' or key == 'module.classifier_max.sigma': # 需要更改温度系数的weight_decay, cosine loss中使用weight decay=0.1
params += [{
"params": [value],
"lr": args.lr,
"weight_decay": args.weight_decay * 100
}]
print('key: {} ,weight decay : {}, value : {}'.format(
key, args.weight_decay * 100, value))
elif key == 'module.classifier.fc1.weight' or key == 'module.classifier_max.fc1.weight':
params += [{
"params": [value],
"lr": args.lr,
"weight_decay": args.weight_decay
}] #从零改变3.5e-4
print('lr of {} is 0'.format(key))
else:
params += [{
"params": [value],
"lr": args.lr,
"weight_decay": args.weight_decay
}]
optimizer = torch.optim.Adam(params)
# 设置dataloader与dataset
if phase > 0:
example_dataset = new_dataset + proto_dataset
train_loader_examplar = get_train_loader(dataset_target,
args.height,
args.width,
args.batch_size,
args.workers,
args.num_instances,
iters,
trainset=example_dataset)
train_loader_target = get_train_loader(dataset_target,
args.height,
args.width,
args.batch_size,
args.workers,
args.num_instances,
iters,
trainset=new_dataset)
# Trainer
trainer = UsicTrainer_E(model_cur=model_cur,
model_ema=model_ema,
model_ref=model_ref,
old_class_num=old_class_num,
new_class_num=args.num_clusters,
alpha=args.alpha)
train_loader_target.new_epoch()
train_loader_examplar.new_epoch(
) # 为了统一，可以在外面生成train_loader_examplar，在train()不使用
trainer.train(phase,
epoch,
train_loader_target,
train_loader_examplar,
optimizer,
cur_lamda,
ce_soft_weight=args.soft_ce_weight,
print_freq=args.print_freq,
train_iters=iters)
def save_model(model_ema, is_best, best_mAP, phase):
save_checkpoint(
{
'state_dict': model_ema.state_dict(),
'phase': phase,
'epoch': epoch + 1,
'best_mAP': best_mAP,
},
is_best,
fpath=osp.join(
args.logs_dir,
'phase{}_model_checkpoint.pth.tar'.format(phase)))
if ((epoch + 1) % args.eval_step == 0
or (epoch == args.epochs - 1)):
mAP_1 = evaluator_ema.evaluate(test_loader_target,
dataset_target.query,
dataset_target.gallery,
cmc_flag=False)
is_best = mAP_1 > best_mAP
best_mAP = max(mAP_1, best_mAP)
save_model(model_ema, is_best, best_mAP, phase + 1)
print(
'\n * Finished phase {:3d} epoch {:3d} model no.1 mAP: {:5.1%} best: {:5.1%}{}\n'
.format(phase, epoch, mAP_1, best_mAP,
' *' if is_best else ''))
# 更新examplar
# 载入最佳模型到model_ema
print('update proto_dataset')
best_weights = load_checkpoint(
osp.join(args.init,
'phase{}_model_best.pth.tar'.format(phase + 1)))
copy_state_dict(best_weights['state_dict'], model_ema)
# 提取特征，并进行聚类
dict_f, _, dic_logit = extract_features(model_ema,
tar_cluster_loader,
print_freq=40)
cf = torch.stack(list(dict_f.values())) # 已经经过normalize
agent_sim = torch.stack(list(dic_logit.values()))
agent_sim = agent_sim[:, :old_class_num] # probs dim=1,是否与old_num相同。
agent_sim = F.softmax(agent_sim, dim=1)
agent_sim_dist = torch.cdist(agent_sim, agent_sim, p=1) / 2
agent_sim_dist = agent_sim_dist.numpy()
rerank_dist = compute_jaccard_dist(
cf, use_gpu=True).numpy() # 经过rerank距离在0到1之间 args.rr_gpu
lambda_a = 0.3 if phase == 0 else 0
rerank_dist = (1 - lambda_a) * rerank_dist + lambda_a * agent_sim_dist
# dict_f, _,_= extract_features(model_ema, tar_cluster_loader, print_freq=40)
# cf = torch.stack(list(dict_f.values()))
# rerank_dist = compute_jaccard_dist(cf, use_gpu=args.rr_gpu).numpy()
#generate DBSCAN
if epoch == 0:
tri_mat = np.triu(rerank_dist, 1) # tri_mat.dim=2
tri_mat = tri_mat[np.nonzero(tri_mat)] # tri_mat.dim=1
tri_mat = np.sort(tri_mat, axis=None)
rho = 2e-3
top_num = np.round(rho * tri_mat.size).astype(int)
eps = tri_mat[:top_num].mean()
print('eps for cluster: {:.3f}'.format(eps))
cluster = DBSCAN(eps=eps,
min_samples=4,
metric='precomputed',
n_jobs=-1)
print('Clustering and labeling...')
labels = cluster.fit_predict(rerank_dist)
num_ids = len(set(labels)) - (1 if -1 in labels else 0)
args.num_clusters = num_ids
print('phase {}， Clustered into {} example classes '.format(
phase, args.num_clusters))
# 计算每个类的所有特征
class_features = collections.defaultdict(list)
image_index = collections.defaultdict(list)
for i, label in enumerate(labels):
if label == -1:
continue
class_features[label].append(cf[i])
image_index[label].append(i)
class_features = [
torch.stack(class_features[idx])
for idx in sorted(class_features.keys())
] # list，其中元素为每个类特征组成的张量
image_index = [image_index[idx] for idx in sorted(image_index.keys())]
tmp_dataset = sorted(input_dataset)
#example_dataset, class_remain = select_proto(class_features, image_index, tmp_dataset, old_class_num=old_class_num)
example_dataset, class_remain = select_proto(
class_features,
image_index,
tmp_dataset,
old_class_num=old_class_num,
mode='herd',
delete_ratio=0.05)
nmi = eval_nmi(example_dataset)
print('NMI of phase{} is {} '.format(phase, nmi))
# 保存此时的proto_dataset
proto_dataset.extend(example_dataset)
proto_dataset_save = osp.join(args.logs_dir,
'phase{}_proto.pkl'.format(phase))
savepickle(proto_dataset, proto_dataset_save)
# 更新最新模型FC参数
class_features = torch.stack(
[torch.mean(features, dim=0) for features in class_remain])
class_features_avg = F.normalize(class_features[:, :2048],
dim=1).float().cuda()
class_features_max = F.normalize(class_features[:, 2048:],
dim=1).float().cuda()
in_features = model_ema.module.classifier.in_features
if isinstance(model_ema.module.classifier, CosineLinear):
new_fc_avg = CosineLinear(in_features,
len(class_features),
sigma=True)
new_fc_avg.weight.data.copy_(class_features_avg)
# new_fc.sigma.data = model_ema.module.classifier.sigma.data
model_ema.module.classifier = new_fc_avg
new_fc_max = CosineLinear(in_features,
len(class_features),
sigma=True)
new_fc_max.weight.data.copy_(class_features_max)
# new_fc_max.sigma.data = model_ema.module.classifier_max.sigma.data
model_ema.module.classifier_max = new_fc_max
else:
new_fc_avg = CosineLinear(in_features,
len(class_features),
sigma=False)
new_fc_avg.weight.data.copy_(class_features_avg)
model_ema.module.classifier.fc2 = new_fc_avg
new_fc_max = CosineLinear(in_features,
len(class_features),
sigma=False)
new_fc_max.weight.data.copy_(class_features_max)
model_ema.module.classifier_max.fc2 = new_fc_max
state = {
'state_dict': model_ema.state_dict(),
'phase': best_weights['phase'],
'epoch': best_weights['epoch'],
'best_mAP': best_weights['best_mAP'],
}
torch.save(
state,
osp.join(args.logs_dir,
'phase{}_model_best.pth.tar'.format(phase + 1)))
def main_worker(args):
global start_epoch, best_mAP
cudnn.benchmark = True
sys.stdout = Logger(osp.join(args.logs_dir, 'log.txt'))
print("==========\nArgs:{}\n==========".format(args))
# Create data loaders
iters = args.iters if (args.iters>0) else None
dataset_target = get_data(args.dataset_target, args.data_dir)
ori_train = dataset_target.train
if not args.no_source:
dataset_source = get_data(args.dataset_source, args.data_dir)
test_loader_target = get_test_loader(dataset_target, args.height, args.width, args.batch_size, args.workers)
# Create model
model_1, model_1_ema = create_model(args)
# Evaluator
evaluator_1_ema = Evaluator(model_1_ema)
best_mAP = 0
for nc in range(args.epochs):
cluster_loader = get_test_loader(dataset_target, args.height, args.width, args.batch_size, args.workers,
testset=dataset_target.train)
dict_f, _ = extract_features(model_1_ema, cluster_loader, print_freq=50)
cf_1 = torch.stack(list(dict_f.values()))
# DBSCAN cluster
if args.no_source:
rerank_dist = compute_jaccard_dist(cf_1, lambda_value=0, source_features=None,
use_gpu=False).numpy()
else:
cluster_loader_source = get_test_loader(dataset_source, args.height, args.width, args.batch_size,
args.workers, testset=dataset_source.train)
dict_f_source, _ = extract_features(model_1_ema, cluster_loader_source, print_freq=50)
cf_1_source = torch.stack(list(dict_f_source.values()))
rerank_dist = compute_jaccard_dist(cf_1, lambda_value=args.lambda_value, source_features=cf_1_source,
use_gpu=False).numpy()
del cf_1_source
tri_mat = np.triu(rerank_dist, 1) # tri_mat.dim=2
tri_mat = tri_mat[np.nonzero(tri_mat)] # tri_mat.dim=1
tri_mat = np.sort(tri_mat, axis=None)
top_num = np.round(args.rho * tri_mat.size).astype(int)
eps = tri_mat[:top_num].mean()
print('eps in cluster: {:.3f}'.format(eps))
print('Clustering and labeling...')
cluster = DBSCAN(eps=eps, min_samples=4, metric='precomputed', n_jobs=-1)
labels = cluster.fit_predict(rerank_dist)
num_ids = len(set(labels)) -1
print('Epoch {} have {} training ids'.format(nc, num_ids))
# generate new dataset
labeled_ind, unlabeled_ind = [], []
for ind, label in enumerate(labels):
if label == -1:
unlabeled_ind.append(ind)
else:
labeled_ind.append(ind)
# print('Epoch {} have {} labeled samples and {} unlabeled samples'.format(nc + 1, len(labeled_ind), len(unlabeled_ind)))
cf_1 = cf_1.numpy()
centers = []
for id in range(num_ids):
centers.append(np.mean(cf_1[labels == id], axis=0))
centers = np.stack(centers, axis=0)
del cf_1, rerank_dist
model_1.module.classifier = nn.Linear(2048, num_ids, bias=False).cuda()
model_1_ema.module.classifier = nn.Linear(2048, num_ids, bias=False).cuda()
model_1.module.classifier_max = nn.Linear(2048, num_ids, bias=False).cuda()
model_1_ema.module.classifier_max = nn.Linear(2048, num_ids, bias=False).cuda()
model_1.module.classifier.weight.data.copy_(
torch.from_numpy(normalize(centers[:, :2048], axis=1)).float().cuda())
model_1_ema.module.classifier.weight.data.copy_(
torch.from_numpy(normalize(centers[:, :2048], axis=1)).float().cuda())
model_1.module.classifier_max.weight.data.copy_(
torch.from_numpy(normalize(centers[:, 2048:], axis=1)).float().cuda())
model_1_ema.module.classifier_max.weight.data.copy_(
torch.from_numpy(normalize(centers[:, 2048:], axis=1)).float().cuda())
del centers
target_label = labels
for i in range(len(dataset_target.train)):
dataset_target.train[i] = list(dataset_target.train[i])
dataset_target.train[i][1] = int(target_label[i])
dataset_target.train[i] = tuple(dataset_target.train[i])
# Optimizer
params = []
for key, value in model_1.named_parameters():
if not value.requires_grad:
continue
params += [{"params": [value], "lr": args.lr, "weight_decay": args.weight_decay}]
optimizer = torch.optim.Adam(params)
# Trainer
trainer = ABMTTrainer(model_1, model_1_ema, num_cluster=num_ids, alpha=args.alpha)
epoch = nc
# # DBSCAN
dataset_target.train = [ori_train[i] for i in labeled_ind]
print(len(dataset_target.train), 'are labeled.')
labeled_loader_target = get_train_loader(dataset_target, args.height, args.width,
args.batch_size, args.workers, args.num_instances, iters, mutual=True)
labeled_loader_target.new_epoch()
trainer.train(epoch, labeled_loader_target, optimizer,
print_freq=args.print_freq, train_iters=len(labeled_loader_target))
def save_model(model_ema, is_best, best_mAP, mid, num_ids):
save_checkpoint({
'state_dict': model_ema.state_dict(),
'epoch': epoch + 1,
'best_mAP': best_mAP,
'num_ids': num_ids
}, is_best, fpath=osp.join(args.logs_dir, 'model'+str(mid)+'_checkpoint.pth.tar'))
if ((epoch+1)%args.eval_step==0 or (epoch==args.epochs-1)):
print('Evaluating teacher net:')
cmc, mAP_1 = evaluator_1_ema.evaluate(test_loader_target, dataset_target.query, dataset_target.gallery, cmc_flag=True)
is_best = (mAP_1>best_mAP)
best_mAP = max(mAP_1, best_mAP)
save_model(model_1_ema, is_best, best_mAP, 1, num_ids)
dataset_target.train = ori_train
print ('Test on the best model.')
checkpoint = load_checkpoint(osp.join(args.logs_dir, 'model_best.pth.tar'))
model_best = models.create(args.arch, num_features=args.features, dropout=args.dropout, num_classes=checkpoint['num_ids'])
model_best.cuda()
model_best = nn.DataParallel(model_best)
evaluator_best = Evaluator(model_best)
model_best.load_state_dict(checkpoint['state_dict'])
evaluator_best.evaluate(test_loader_target, dataset_target.query, dataset_target.gallery, cmc_flag=True)
def main_worker(args):
global start_epoch, best_mAP
cudnn.benchmark = True
if not args.evaluate:
sys.stdout = Logger(osp.join(args.logs_dir, 'log.txt'))
else:
log_dir = osp.dirname(args.resume)
sys.stdout = Logger(osp.join(log_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
# Create data loaders
iters = args.iters if (args.iters>0) else None
dataset_source, num_classes, train_loader_source, test_loader_source = \
get_data(args.dataset_source, args.data_dir, args.height,
args.width, args.batch_size, args.workers, args.num_instances, iters)
dataset_target, _, train_loader_target, test_loader_target = \
get_data(args.dataset_target, args.data_dir, args.height,
args.width, args.batch_size, args.workers, 0, iters)
# Create model
model = models.create(args.arch, num_features=args.features, dropout=args.dropout, num_classes=num_classes)
model.cuda()
model = nn.DataParallel(model)
# Load from checkpoint
if args.resume:
checkpoint = load_checkpoint(args.resume)
copy_state_dict(checkpoint['state_dict'], model)
start_epoch = checkpoint['epoch']
best_mAP = checkpoint['best_mAP']
print("=> Start epoch {} best mAP {:.1%}"
.format(start_epoch, best_mAP))
# Evaluator
evaluator = Evaluator(model)
if args.evaluate:
print("Test on source domain:")
evaluator.evaluate(test_loader_source, dataset_source.query, dataset_source.gallery, cmc_flag=True, rerank=args.rerank)
print("Test on target domain:")
evaluator.evaluate(test_loader_target, dataset_target.query, dataset_target.gallery, cmc_flag=True, rerank=args.rerank)
return
params = []
for key, value in model.named_parameters():
if not value.requires_grad:
continue
params += [{"params": [value], "lr": args.lr, "weight_decay": args.weight_decay}]
optimizer = torch.optim.Adam(params)
lr_scheduler = WarmupMultiStepLR(optimizer, args.milestones, gamma=0.1, warmup_factor=0.01, warmup_iters=args.warmup_step)
# Trainer
trainer = ABMTPreTrainer(model, num_classes, margin=args.margin)
# Start training
for epoch in range(start_epoch, args.epochs):
train_loader_source.new_epoch()
train_loader_target.new_epoch()
trainer.train(epoch, train_loader_source, train_loader_target, optimizer,
train_iters=len(train_loader_source), print_freq=args.print_freq)
lr_scheduler.step()
if ((epoch+1)%args.eval_step==0 or (epoch==args.epochs-1)):
_, mAP = evaluator.evaluate(test_loader_source, dataset_source.query, dataset_source.gallery, cmc_flag=True)
is_best = mAP > best_mAP
best_mAP = max(mAP, best_mAP)
save_checkpoint({
'state_dict': model.state_dict(),
'epoch': epoch + 1,
'best_mAP': best_mAP,
}, is_best, fpath=osp.join(args.logs_dir, 'checkpoint.pth.tar'))
print('\n * Finished epoch {:3d} source mAP: {:5.1%} best: {:5.1%}{}\n'.
format(epoch, mAP, best_mAP, ' *' if is_best else ''))
print("Test on target domain:")
evaluator.evaluate(test_loader_target, dataset_target.query, dataset_target.gallery, cmc_flag=True, rerank=args.rerank)