Diversity with Cooperation: Ensemble Methods for Few-Shot Classification

Ensembles of deep neural networks combined with nearest centroid classifiers to solve few-shot classification.

Introduction

This repository contains original implementation of the paper 'Diversity with Cooperation: Ensemble Methods for Few-Shot Classification' by Nikita Dvornik, Cordelia Schmid and Julien Mairal. The paper is available at https://arxiv.org/abs/1903.11341

Cite

@inproceedings{dvornik2019diversity,
  title = {Diversity with Cooperation: Ensemble Methods for Few-Shot Classification},
  author = {Dvornik, Nikita and Schmid, Cordelia and Mairal, Julien},
  booktitle = {{IEEE International Conference on Computer Vision (ICCV)}},
  year = {2019}
}

Requirements

• Python 3.5
• PyTorch >= 1.2
• TorchVision >= 0.2
• tqdm

Preparation

Clone this git repository

git clone https://github.com/dvornikita/fewshot_ensemble.git

and install all the requirements listed above. In the file path.py, specify the path to your datasets. For mini-imagenet and tiered-imagenet, you can specify the path to the full ImageNet dataset. Make sure that your ImageNet version contains classes with names in splits/. You are all set.

Training and testing a single feature extractor

For example, you want to train a WideResNet28 on mini-ImageNet

python singles/train.py --model.model_name=wideresnet --data.dataset=mini_imagenet --model.backbone=wide 

This command will train a network on a classification task, using the training set of mini-ImageNet. Here, --model.model_name is a required argument that specifies the name of the experiment (needed to restore the weights later in testing and to visualize statistics in Tensorboard), --data.dataset specifies the name of the dataset you want to train on. The argument --model.backbone specifies the feature extractor you use ("deep" is ResNet18 and "wide" stands for WideResNet).

To test this model, you should run

python singles/test.py --model.model_name=wideresnet --data.dataset=mini_imagenet --model.backbone=wide --data.test_shot=5

This command will generate 1000 5-shot problems, test previously learned feature extractor "wideresnet" on them and report mean accuracy with confidence intervals. To test 1-shot performance, run the command with --data.test_shot=1.

Training and testing ensembles with relations

For example, if you want to train a Robust ensemble of 10 ResNets on mini-ImageNet

python ensembles/train.py --model.model_name=deep_robust10 --data.dataset=mini_imagenet --model.backbone=deep --ens.num_heads=10 --ens.relation_type=robust

This code will train 10 ResNet18 networks with Robust relations. Here, ens.num_heads sets the ensemble size, and --ens.relation_type specifies the type of relation between the network you want for training the ensemble (the value is in ['independent', 'cooperation', 'diversity', 'robust']).

To test the model, run:

python ensembles/test.py --model.model_name=deep_robust10 --data.dataset=mini_imagenet --model.backbone=deep --ens.num_heads=10 --data.test_shot=5

Distilling your ensemble and testing it

1. To distill the ensemble of 10 ResNet18, trained in the previously, run:

python ensembles/distill.py --model.model_name=robust10_distilled --data.dataset=mini_imagenet --model.backbone=deep --ens.num_heads=10 --distil.name=deep_robust10

Importantly, the flag --distil.name specifies the name of the full ensemble you want to distill into a single network with name "robust10_distilled". You can change parameters of distillation in config.py file. To test this model, run:

python singles/test.py --model.model_name=robust10_distilled --data.dataset=mini_imagenet --model.backbone=wide --data.test_shot=5

NOTE: When distilling the model, it is being automatically placed to singles/ directory. This is because it if just a single neural network.