Home

Awesome

few-shot-ssl-public

Code for paper Meta-Learning for Semi-Supervised Few-Shot Classification. [arxiv]

Dependencies

Our code is tested on Ubuntu 14.04 and 16.04.

Setup

First, designate a folder to be your data root:

export DATA_ROOT={DATA_ROOT}

Then, set up the datasets following the instructions in the subsections.

Omniglot

[Google Drive] (9.3 MB)

# Download and place "omniglot.tar.gz" in "$DATA_ROOT/omniglot".
mkdir -p $DATA_ROOT/omniglot
cd $DATA_ROOT/omniglot
mv ~/Downloads/omniglot.tar.gz .
tar -xzvf omniglot.tar.gz
rm -f omniglot.tar.gz

miniImageNet

[Google Drive] (1.1 GB)

Update: Python 2 and 3 compatible version: [train] [val] [test]

# Download and place "mini-imagenet.tar.gz" in "$DATA_ROOT/mini-imagenet".
mkdir -p $DATA_ROOT/mini-imagenet
cd $DATA_ROOT/mini-imagenet
mv ~/Downloads/mini-imagenet.tar.gz .
tar -xzvf mini-imagenet.tar.gz
rm -f mini-imagenet.tar.gz

tieredImageNet

[Google Drive] (12.9 GB)

# Download and place "tiered-imagenet.tar" in "$DATA_ROOT/tiered-imagenet".
mkdir -p $DATA_ROOT/tiered-imagenet
cd $DATA_ROOT/tiered-imagenet
mv ~/Downloads/tiered-imagenet.tar .
tar -xvf tiered-imagenet.tar
rm -f tiered-imagenet.tar

Note: Please make sure that the following hardware requirements are met before running tieredImageNet experiments.

Core Experiments

Please run the following scripts to reproduce the core experiments.

# Clone the repository.
git clone https://github.com/renmengye/few-shot-ssl-public.git
cd few-shot-ssl-public

# To train a model.
python run_exp.py --data_root $DATA_ROOT             \
                  --dataset {DATASET}                \
                  --label_ratio {LABEL_RATIO}        \
                  --model {MODEL}                    \
                  --results {SAVE_CKPT_FOLDER}       \
                  [--disable_distractor]

# To test a model.
python run_exp.py --data_root $DATA_ROOT             \
                  --dataset {DATASET}                \
                  --label_ratio {LABEL_RATIO}        \
                  --model {MODEL}                    \
                  --results {SAVE_CKPT_FOLDER}       \
                  --eval --pretrain {MODEL_ID}       \
                  [--num_unlabel {NUM_UNLABEL}]      \
                  [--num_test {NUM_TEST}]            \
                  [--disable_distractor]             \
                  [--use_test]

Simple Baselines for Few-Shot Classification

Please run the following script to reproduce a suite of baseline results.

python run_baseline_exp.py --data_root $DATA_ROOT    \
                           --dataset {DATASET}

Run over Multiple Random Splits

Please run the following script to reproduce results over 10 random label/unlabel splits, and test the model with different number of unlabeled items per episode. The default seeds are 0, 1001, ..., 9009.

python run_multi_exp.py --data_root $DATA_ROOT       \
                        --dataset {DATASET}          \
                        --label_ratio {LABEL_RATIO}  \
                        --model {MODEL}              \
                        [--disable_distractor]       \
                        [--use_test]

Citation

If you use our code, please consider cite the following:

@inproceedings{ren18fewshotssl,
  author   = {Mengye Ren and 
              Eleni Triantafillou and 
              Sachin Ravi and 
              Jake Snell and 
              Kevin Swersky and 
              Joshua B. Tenenbaum and 
              Hugo Larochelle and 
              Richard S. Zemel},
  title    = {Meta-Learning for Semi-Supervised Few-Shot Classification},
  booktitle= {Proceedings of 6th International Conference on Learning Representations {ICLR}},
  year     = {2018},
}