Meta-Learning Implementation

Implementation of MAML and Prototypical Network using torchmeta (additional methods will be added, soon.)

Dependency

First, download the torchmeta package:

pip install torchmeta

or

git clone https://github.com/tristandeleu/pytorch-meta.git
cd pytorch-meta
python setup.py install

Datasets: miniImageNet, tieredImageNet, CIFAR-FS, CUB

Next, download the datasets (reference from https://github.com/renmengye/few-shot-ssl-public):

miniImageNet: [Google Drive Link]

tieredImageNet: [Google Drive Link]

CIFAR_FS: [Google Drive Link] (or download available via code $\rightarrow$ highly recommend)

CUB : [Google Drive Link]

Check datasets.py:

• In the datasets.py file located within the utils folder, you will come across the import statements for datasets such as miniImagenet and tieredImagenet. When you navigate to the imported file, it's necessary to modify torchmeta.datasets.utils $\Rightarrow$ torchvision.datasets.utils because of check integretity.
• Make sure to verify that the download is set to True in the datasets.py (CIFAR_FS is available) $\Rightarrow$ As a result of modifications to Google Drive's policy, downloading will now need to be done manually. Move the datasets as indicated in the provided datasets path below.
• DATA_PATH : Your own datasets folder path

DATA_PATH
└─cub
  | CUB_200_2011.tgz
└─cifar100
  | cifar-fs
  | data.hdf5
  | fine_names.json
└─miniimagenet
  | mini-imagenet.tar.gz
└─tieredimagenet
  | tiered-imagenet.tar

Training / Evaluation / Test

Training

MAML

python train_maml.py --datasets [DATASETS] --epoch 60000 --num_shots 5 --batch_size 2 

Prototypical Networks

python train_proto.py --datasets [DATASETS] --num_ways_proto 20 --num_shots 5 --epoch 200 --batch_size 100 

Evaluation / Test

python eval_meta.py --[OPTIONS]

option arguments:  
--epoch:              epoch number (default: 60000)  
--num_ways:           N-way (default: 5)  
--num_ways_proto:     N-way for Proto-Net (default: 30)  
--num_shots:          k shots for support set (default: 5)  
--num_shots_test:     number of query set (default: 15) 
--imgc:               RGB(image channel) (default: 3)  
--filter_size:        size of convolution filters (default: 64)  
--batch_size:         meta-batch size (default: 2)  
--max_test_task:      number of tasks for evaluation (default: 1000)  
--meta_lr:            outer-loop learning rate (default: 1e-3)  
--update_lr:          inner-loop learning rate (default: 1e-2)  
--update_step:        number of inner-loop update steps while training (default: 5)  
--update_test_step:   number of inner-loop update steps while evaluating (default: 10) 
--update:             update method: MAML, ANIL, BOIL (default: MAML)
--scale_factor:       Scaling factor for the cosine classifier (default: 10)
--dropout:            dropout probability (default: 0.2)
--gpu_id:             gpu device number (default: 0)
--model:              model architecture: Conv-4, ResNet12 (default: conv4)
--datasets:           datasets: miniimagenet, tieredimagenet, cifar-fs, CUB (default: miniimagenet)
--version:            file version (default: 0)  

Result

MAML

Datasets		5 ways - 1 shot	5 ways - 5 shot
mini-ImageNet (MAML)	Original	48.70 $\pm$ 1.84	63.11 $\pm$ 0.92
	Ours	48.79 $\pm$ 0.16	62.43 $\pm$ 0.89
tiered-ImageNet (TPN)	Original	52.54 $\pm$ 0.35	70.97 $\pm$ 0.51
	Ours	50.01 $\pm$ 0.28	65.58 $\pm$ 0.15
CIFAR_FS (R2-D2)	Original	58.90 $\pm$ 1.90	71.50 $\pm$ 1.00
	Ours	57.36 $\pm$ 0.37	72.41 $\pm$ 0.78
CUB (FEAT)	Original	55.92 $\pm$ 0.95	72.09 $\pm$ 0.76
	Ours	56.98 $\pm$ 0.28	73.64 $\pm$ 0.23

Prototypical Network (Higher-Way)

Euclidean

Datasets		5 ways - 1 shot	5 ways - 5 shot
mini-ImageNet (ProtoNet)	Original	49.42 $\pm$ 0.78	68.20 $\pm$ 0.66
	Ours	49.45 $\pm$ 0.23	66.17 $\pm$ 0.15
tiered-ImageNet (TPN)	Original	53.31 $\pm$ 0.89	72.69 $\pm$ 0.74
	Ours	52.54 $\pm$ 0.35	71.97 $\pm$ 0.51
CIFAR_FS (R2-D2)	Original	55.50 $\pm$ 0.70	72.00 $\pm$ 0.60
	Ours	54.33 $\pm$ 0.20	73.60 $\pm$ 0.19
CUB (FEAT)	Original	51.31 $\pm$ 0.91	70.77 $\pm$ 0.69
	Ours	51.13 $\pm$ 0.77	70.23 $\pm$ 0.81

Reference

• torchmeta
• SiMT
• BOIL