Augmented Meta-Transfer Learning(A-MTL)

for few shot image classification

Datasets

Directly download processed images: [Download Page]

𝒎𝒊𝒏𝒊ImageNet

The 𝑚𝑖𝑛𝑖ImageNet dataset was proposed by Vinyals et al. for few-shot learning evaluation. Its complexity is high due to the use of ImageNet images but requires fewer resources and infrastructure than running on the full ImageNet dataset. In total, there are 100 classes with 600 samples of 84×84 color images per class. These 100 classes are divided into 64, 16, and 20 classes respectively for sampling tasks for meta-training, meta-validation, and meta-test. To generate this dataset from ImageNet, you may use the repository 𝑚𝑖𝑛𝑖ImageNet tools.

Fewshot-CIFAR100

Fewshot-CIFAR100 (FC100) is based on the popular object classification dataset CIFAR100. The splits were proposed by TADAM. It offers a more challenging scenario with lower image resolution and more challenging meta-training/test splits that are separated according to object super-classes. It contains 100 object classes and each class has 600 samples of 32 × 32 color images. The 100 classes belong to 20 super-classes. Meta-training data are from 60 classes belonging to 12 super-classes. Meta-validation and meta-test sets contain 20 classes belonging to 4 super-classes, respectively.

Performance

MTL performance

backbone	Aug method	mini 1-shot	mini 5-shot	FC100 1-shot	FC100 5-shot
ResNet12	None	0.5328 + 0.0080	0.6857 + 0.0070	0.3922 + 0.0071	0.5153 + 0.0070
	Transport	0.5342 + 0.0085	0.6903 + 0.0070	0.4007+0.0079	0.5119 + 0.0071
ResNet18	None	0.4632 + 0.0077	0.6618 + 0.0069	0.3749 + 0.0078	0.5224 + 0.0071
	Transport	0.4734 + 0.0074	0.6597 + 0.0069	0.3804 + 0.0069	0.5256+0.0069

MAML

see README of maml

Meta-Transfer Learning

To reproduce the result of experiment, do the following command

cd /meta-transfer-learning

pre_train command

• pre_train(ResNet12, MiniImageNet)

python main.py --phase pre_train --model_type ResNet --dataset MiniImageNet --dataset_dir ../data/mini-imagenet --pre_batch_size 100 --pre_lr 0.1 --gpu xxx

• pre_train(ResNet18, MiniImageNet)

python main.py --phase pre_train --model_type ResNet18 --dataset MiniImageNet --dataset_dir ../data/mini-imagenet --pre_batch_size 64 --pre_lr 0.01 --gpu xxx

meta_train command

• meta_train(ResNet12, MiniImageNet, 1 shot)

python main.py --phase meta_train --dataset MiniImageNet --dataset_dir ../data/mini-imagenet/ --shot 1 --model_type ResNet --pre_batch_size 128 --pre_lr 0.1 --gpu xxx

• meta_train(ResNet12, FC100, 1 shot)

python main.py --phase meta_train --dataset FC100 --dataset_dir ../data/fc100/ --shot 1 --model_type ResNet --pre_batch_size 128 --pre_lr 0.1 --gpu xxx

meta_eval command

• meta_eval(ResNet12, MiniImageNet, 5 shot)

python main.py --phase meta_train --dataset MiniImageNet --dataset_dir ../data/mini-imagenet/ --shot 5 --model_type ResNet --pre_batch_size 128 --pre_lr 0.1 --gpu xxx

• meta_eval(ResNet12, FC100, 5 shot)

python main.py --phase meta_train --dataset FC100 --dataset_dir ../data/fc100/ --shot 5 --model_type ResNet --pre_batch_size 128 --pre_lr 0.1 --gpu xxx

• 需要注意的地方：
  1. --gpu需要根据自己显卡的情况来设置，一般设置为0
  2. 如果要跑Resnet18只需将--model_type ResNet改为--model_type ResNet18
  3. 如果要训练FC100数据集只需设置--dataset FC100 dataset_dir ../data/fc100
  4. --dataset_dir 需要根据实际情况选择数据集的路径
  5. 一个模型的pre_train 大概需要4-5h, meta_train 大概需要1-2小时
  6. 对于我们的A-MTL方法，只需在meta_train 和meta_eval 阶段加入--augment 1即可