EASY - Ensemble Augmented-Shot Y-shaped Learning: State-Of-The-Art Few-Shot Classification with Simple Ingredients.

This repository is the official implementation of EASY - Ensemble Augmented-Shot Y-shaped Learning: State-Of-The-Art Few-Shot Classification with Simple Ingredients.

EASY proposes a simple methodology, that reaches or even beats state of the art performance on multiple standardized benchmarks of the field, while adding almost no hyperparameters or parameters to those used for training the initial deep learning models on the generic dataset.

Downloads

Please click the Google Drive link for downloading the features, backbones and datasets.

Each of the files (backbones and features) have the following prefixes depending on the backbone:

Each of the features file is named as follow :

• if not AS : " features .pt11"
• if AS : " featuresAS .pt11"

Testing scripts for EASY

Run scripts to evaluate the features on FSL tasks for Y and ASY. For EY and EASY use the corresponding features.

Inductive setup using NCM

Test features on miniimagenet using Y (Resnet12)

$ python main.py --dataset-path "
     
      " --dataset miniimagenet --model resnet12 --test-features '
      
       /minifeatures1.pt11' --preprocessing ME

      
     

Test features on miniimagenet using ASY (Resnet12)

$ python main.py --dataset-path "
     
      " --dataset miniimagenet --model resnet12 --test-features '
      
       /minifeaturesAS1.pt11' --preprocessing ME

      
     

Test features on miniimagenet using EY (3xResNet12)

$ python main.py --dataset-path "
       
        " --dataset miniimagenet --model resnet12 --test-features "[
        
         /minifeatures1.pt11, 
         
          /minifeatures2.pt11, 
          
           /minifeatures3.pt11]" --preprocessing ME

          
         
        
       

Test features on miniimagenet using EASY (3xResNet12)

$ python main.py --dataset-path "
       
        " --dataset miniimagenet --model resnet12 --test-features "[
        
         /minifeaturesAS1.pt11, 
         
          /minifeaturesAS2.pt11, 
          
           /minifeaturesAS3.pt11]" --preprocessing ME 

          
         
        
       

Transductive setup using Soft k-means

Test features on miniimagenet using Y (ResNet12)

$ python main.py --dataset-path "
     
      " --dataset miniimagenet --model resnet12 --test-features '
      
       /minifeatures1.pt11'--postprocessing ME --transductive --transductive-softkmeans --transductive-temperature-softkmeans 20

      
     

Test features on miniimagenet using ASY (ResNet12)

$ python main.py --dataset-path "
     
      " --dataset miniimagenet --model resnet12 --test-features '
      
       /minifeaturesAS1.pt11' --postprocessing ME --transductive --transductive-softkmeans --transductive-temperature-softkmeans 20

      
     

Test features on miniimagenet using EY (3xResNet12)

$ python main.py --dataset-path "
       
        " --dataset miniimagenet --model resnet12 --test-features "[
        
         /minifeatures1.pt11, 
         
          /minifeatures2.pt11, 
          
           /minifeatures3.pt11]" --postrocessing ME  --transductive --transductive-softkmeans --transductive-temperature-softkmeans 20

          
         
        
       

Test features on miniimagenet using EASY (3xResNet12)

$ python main.py --dataset-path "
       
        " --dataset miniimagenet --model resnet12 --test-features "[
        
         /minifeaturesAS1.pt11, 
         
          /minifeaturesAS2.pt11, 
          
           /minifeaturesAS3.pt11]" --postrocessing ME  --transductive --transductive-softkmeans --transductive-temperature-softkmeans 20

          
         
        
       

Training scripts for Y

Train a model on miniimagenet using manifold mixup, self-supervision and cosine scheduler

$ python main.py --dataset-path "
    
     " --dataset miniimagenet --model resnet12 --epochs 0 --manifold-mixup 500 --rotations --cosine --gamma 0.9 --milestones 100 --batch-size 128 --preprocessing ME 

    

Important Arguments

Some important arguments for our code.

Training arguments

• dataset: choices=['miniimagenet', 'cubfs','tieredimagenet', 'fc100', 'cifarfs']
• model: choices=['resnet12', 'resnet18', 'resnet20', 'wideresnet', 's2m2r']
• dataset-path: path of the datasets folder which contains folders of all the datasets.

Few-shot Classification

• preprocessing: preprocessing sequence for few shot given as a string, can contain R:relu P:sqrt E:sphering and M:centering using the base data.
• postprocessing: postprocessing sequence for few shot given as a string, can contain R:relu P:sqrt E:sphering and M:centering on the few-shot data, used for transductive setting.

Few-shot classification Results

Experimental results on few-shot learning datasets with ResNet-12 backbone. We report our average results with 10000 randomly sampled episodes for both 1-shot and 5-shot evaluations.

MiniImageNet Dataset (inductive)

TieredImageNet Dataset (inductive)

CUBFS Dataset (inductive)

CIFAR-FS Dataset (inductive)

FC-100 Dataset (inductive)

Minimagenet (transductive)

CUB-FS (transductive)

CIFAR-FS (transductive)

FC-100 (transductive)

Tiered Imagenet (transducive)