args.num_classes) # 数据越多 meta_epoch 增长越快
args.note = """
1.system sample 从 D_L 取 meta_dataset
2.每个 meta_epoch 后，finetune 3 epochs；或者直接从 D_L 取更多数据；去掉 finetune 阶段
3.重回 unlabel_dataloader 全局 shuffle, global_total_uc 增加趋势差不多 16335/60
最终结果 91.13
"""
pprint(vars(args))
# cudnn.benchmark = True
# get datasets
random.seed(args.seed) # 保证初始划分 label/unlabel 一致
label_dataset, unlabel_dataset, _, test_dataset = get_imb_meta_test_datasets(
args.dataset, args.num_classes, 0, args.imb_factor, args.split,
args.ratio) # CIFAR10
random.seed() # 解除
# imb_train/valid_meta/test
kwargs = {'num_workers': 4, 'pin_memory': True}
# 将 label/unlabel loader 放入同一 batch
label_loader = DataLoader(label_dataset,
batch_size=args.batch_size,
shuffle=True,
**kwargs)
unlabel_loader = DataLoader(
unlabel_dataset, # DataLoader will cvt np to tensor
batch_size=args.uc_batchsize,
shuffle=True,
default='exp',
type=str,
help='experiment tag to create tensorboard, model save dir name')
params = [
'--dataset', 'cifar10', '--num_classes', '10', '--imb_factor', '1',
'--num_meta', '0', '--tag', 'base'
]
args = parser.parse_args(params)
pprint(vars(args))
cudnn.benchmark = True
# get datasets
random.seed(args.seed)
train_dataset, _, test_dataset = get_imb_meta_test_datasets(
args.dataset, args.num_classes, args.num_meta, args.imb_factor)
# imb_train/valid_meta/test
kwargs = {'num_workers': 4, 'pin_memory': True}
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
drop_last=False,
shuffle=True,
**kwargs)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
shuffle=False,
**kwargs)
"""
baseline
- directly train on total cifar10 dataset
