# Another choice for Office-home:
# width = 1024
# srcweight = 3
# is_cen = True
else:
width = -1
model_instance = MDD(base_net='ResNet50',
width=width,
use_gpu=True,
class_num=class_num,
srcweight=srcweight)
train_source_loader = load_images(source_file,
batch_size=32,
is_cen=is_cen)
train_target_loader = load_images(target_file,
batch_size=32,
is_cen=is_cen)
test_target_loader = load_images(target_file,
batch_size=32,
is_train=False)
param_groups = model_instance.get_parameter_list()
group_ratios = [group['lr'] for group in param_groups]
assert cfg.optim.type == 'sgd', 'Optimizer type not supported!'
optimizer = torch.optim.SGD(param_groups, **cfg.optim.params)
"Source label distribution: {}".format(source_label_distribution))
print(
"Target label distribution: {}".format(target_label_distribution))
print("True weights : {}".format(true_weights[:, 0].cpu().numpy()))
else:
true_weights = None
source_label_distribution = None
target_label_distribution = None
print(
"Source label distribution: {}".format(source_label_distribution))
print(
"Target label distribution: {}".format(target_label_distribution))
print("True weights : {}".format(true_weights))
train_source_loader = load_images(source_file, batch_size=args.batch_size, resize_size=resize_size, crop_size=crop_size, \
is_cen=is_cen, root_folder=args.root_folder)
if use_ssda:
train_target_loader = load_images(target_file, batch_size=args.batch_size-args.labeled_tgt_batch_size, \
resize_size=resize_size, crop_size=crop_size, is_cen=is_cen, \
root_folder=args.root_folder)
train_labeled_target_loader = load_images(labeled_target_file, batch_size=args.labeled_tgt_batch_size, \
resize_size=resize_size, crop_size=crop_size, is_cen=is_cen, \
root_folder=args.root_folder)
else:
train_target_loader = load_images(target_file, batch_size=args.batch_size, resize_size=resize_size, \
crop_size=crop_size, is_cen=is_cen, root_folder=args.root_folder)
train_labeled_target_loader = None
test_source_loader = load_images(source_test_file, batch_size=16, resize_size=resize_size, crop_size=crop_size, \
is_train=False, is_cen=is_cen, root_folder=args.root_folder)
# Another choice for Office-home:
# width = 1024
# srcweight = 3
# is_cen = True
else:
width = -1
model_instance = DANN(base_net='ResNet50',
width=width,
use_gpu=True,
class_num=class_num,
srcweight=srcweight)
train_source_clean_loader = load_images(source_file,
batch_size=32,
is_cen=is_cen,
split_noisy=False)
train_source_noisy_loader = train_source_clean_loader
train_target_loader = load_images(target_file,
batch_size=32,
is_cen=is_cen)
test_target_loader = load_images(target_file,
batch_size=32,
is_train=False)
param_groups = model_instance.get_parameter_list()
group_ratios = [group['lr'] for group in param_groups]
assert cfg.optim.type == 'sgd', 'Optimizer type not supported!'
optimizer = torch.optim.SGD(param_groups, **cfg.optim.params)
def train(config):
train_source_loader = load_images(config['source_path'],
batch_size=config['batch_size'],
resize_size = config['resize_size'],
is_cen = config['is_cen'],
)
train_target_loader = load_images(config['target_path'],
batch_size=config['batch_size'],
resize_size=config['resize_size'],
is_cen=config['is_cen'],)
test_target_loader = load_images(config['target_path'],
batch_size=config['batch_size'],
resize_size=config['resize_size'],
is_cen=config['is_cen'],
is_train=False)
model_instance = MDD(base_net=config['base_net'],
width = config['width'],
use_gpu = True,
class_num = config['class_num'],
srcweight = config['srcweight'])
param_groups = model_instance.get_parameter_list()
group_ratios = [group['lr'] for group in param_groups]
assert config['optim']['type'] == 'sgd', 'Optimizer type not supported!'
optimizer = torch.optim.SGD(param_groups, **config['optim']['params'])
assert config['lr_scheduler']['type'] == 'inv', 'Scheduler type not supported!'
lr_scheduler = INVScheduler(gamma = config['lr_scheduler']['gamma'],
decay_rate = config['lr_scheduler']['decay_rate'],
init_lr = config['init_lr'])
model_instance.set_train(True)
print("start train...")
iter_num = 0
epoch = 0
total_progress_bar = tqdm.tqdm(desc='Train iter', total=config['max_iter'])
while True:
for (datas, datat) in tqdm.tqdm(
zip(train_source_loader, train_target_loader),
total=min(len(train_source_loader),
len(train_target_loader)),
desc='Train epoch = {}'.format(epoch),
ncols=80, leave=False):
inputs_source, labels_source = datas
inputs_target, labels_target = datat
optimizer = lr_scheduler.next_optimizer(group_ratios, optimizer, iter_num / 5)
optimizer.zero_grad()
if model_instance.use_gpu:
inputs_source, inputs_target, labels_source = Variable(inputs_source).cuda(), Variable(
inputs_target).cuda(), Variable(labels_source).cuda()
else:
inputs_source, inputs_target, labels_source = Variable(inputs_source), Variable(
inputs_target), Variable(labels_source)
train_batch(model_instance, inputs_source, labels_source, inputs_target, optimizer)
# val
if iter_num % config['eval_iter'] == 0 and iter_num !=0:
eval_result = evaluate(model_instance,
test_target_loader,)
print(eval_result)
acc = round(eval_result['accuracy'], 2)
print(acc)
torch.save(model_instance.c_net.state_dict(), os.path.join(config["output_path"], str(acc*100)+"_model.pth.tar"))
iter_num += 1
total_progress_bar.update(1)
epoch += 1
if iter_num >= config['max_iter']:
break
print('finish train')
# Another choice for Office-home:
# width = 1024
# srcweight = 3
# is_cen = True
else:
width = -1
model_instance = MDD(base_net='ResNet50',
width=width,
use_gpu=True,
class_num=class_num,
srcweight=srcweight)
train_source_loader = load_images(source_file,
batch_size=32,
is_cen=is_cen,
drop_last=True)
test_source_loader = load_images(source_file,
batch_size=32,
is_train=False)
param_groups = model_instance.get_parameter_list()
group_ratios = [group['lr'] for group in param_groups]
assert cfg.optim.type == 'sgd', 'Optimizer type not supported!'
optimizer = torch.optim.SGD(param_groups, **cfg.optim.params)
assert cfg.lr_scheduler.type == 'inv', 'Scheduler type not supported!'
lr_scheduler = INVScheduler(gamma=cfg.lr_scheduler.gamma,
decay_rate=cfg.lr_scheduler.decay_rate,