Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Learning to Self-Train for Semi-Supervised Few-Shot Classification

LICENSE Python TensorFlow

This repository contains the TensorFlow implementation for NeurIPS 2019 Paper "Learning to Self-Train for Semi-Supervised Few-Shot Classification".

Check the few-shot classification leaderboard.



In order to run this repository, we advise you to install python 2.7 or 3.5 and TensorFlow 1.3.0 with Anaconda.

You may download Anaconda and read the installation instruction on their official website:

Create a new environment and install tensorflow on it:

conda create --name lst-tf python=2.7
conda activate lst-tf
conda install tensorflow-gpu=1.3.0

Install other requirements:

pip install scipy tqdm opencv-python pillow matplotlib

Clone this repository:

git clone 
cd learning-to-self-train

Project Architecture

├── data_generator              # dataset generator 
|   └──  # data genertor for meta-train phase
├── models                      # tensorflow model files 
|   ├──               # resnet12 CNN class
|   └──       # semi-supervised meta-train model class
├── trainer                     # tensorflow trianer files  
|   └──             # semi-supervised meta-train trainer class
├── utils                       # a series of tools used in this repo
|   └──                 # miscellaneous tool functions
├── data                        # the folder containing datasets for experiments
├── pretrain_weights_dir        # the folder containing MTL pre-training weights
├── weights_saving_dir          # the folder containing meta-training weights
├── test_output_dir             # the folder containing meta-testing files
├── filenames_and_labels        # the folder containing image file paths and labels for experiments
├──                # the python file with main function and parameter settings for meta-training
└──                 # the python file with main function and parameter settings for meta-testing

Running Experiments

First, download our processed images: miniImagenet[Download Page] or tieredImagenet[Download Page], move the unziped folder to ./data. And then download the pre-trained models: miniImagenet[Download Page] or tieredImagenet[Download Page], move the unziped folder to ./pretrain_weights_dir.

Training from Pre-Trained Models

Run semi-supervised meta-train phase (e.g. 𝑚𝑖𝑛𝑖ImageNet, 1-shot) :

python --shot_num=1 --dataset='miniImagenet' --pretrain_class_num=64 --nb_ul_samples=10 --metatrain_iterations=15000 --exp_name='LST_mini_1_shot'

Run semi-supervised meta-test phase (e.g. 𝑚𝑖𝑛𝑖ImageNet, 1-shot) :

python --shot_num=1 --dataset='miniImagenet' --pretrain_class_num=64 --use_distractors=False --nb_ul_samples=100 --unfiles_num=10 --test_iter=15000 --recurrent_stage_nums=6 --nums_in_folders=30 --hard_selection=20 --exp_name='LST_mini_1_shot' 

Hyperparameters and Options

There are some main hyperparameters used in the experiments, you can edit them in the and the file for meta-train and meta-test phase respectively. There are two kinds of hyperparameters: (1) common hyperparameters that shared with meta-train and meta-test, (2) test-specific hyperparameters that used for recurrent self-training process in meta-test.

  • Common hyperparameters:

    • way_num number of classes
    • shot_num number of examples per class
    • dataset dataset used in the experiment (miniImagenet or tieredImagenet)
    • pretrain_class_num number of meta-train classes
    • exp_name name for the current experiment
    • meta_batch_size number of tasks sampled per meta-update in meta-train phase
    • base_lr step size alpha for inner gradient update
    • meta_lr the meta learning rate for SS and initial model parameters
    • min_meta_lr the min meta learning rate for all meta-parameters
    • swn_lr the meta learning rate for SWN
    • nb_ul_samples number of unlabeled examples per class
    • re_train_epoch_num number of re-training inner gradient updates
    • train_base_epoch_num number of total inner gradient updates during train (meta-train only)
    • test_base_epoch_num number of total inner gradient updates during test (meta-test only)
  • Test-specific hyperparameters:

    • use_distractors if using distractor classes during meta-test
    • num_dis number of distracting classes used for meta-testing
    • unfiles_num number of unlabeled sample files used in the experiment (There are 10 unlabeled samples per class in each file)
    • recurrent_stage_nums number of recurrent stages used during meta-test
    • local_update_num number of inner gradient updates used in each recurrent stage
    • nums_in_folders number of unlabeled samples (per class) used in each recurrent stage
    • hard_selection number of remaining samples (per class) after applying hard-selection

If you want to change other settings, please see the comments and descriptions in and


(%) 𝑚𝑖𝑛𝑖 𝒕𝒊𝒆𝒓𝒆𝒅 𝑚𝑖𝑛𝑖 (w/D) 𝒕𝒊𝒆𝒓𝒆𝒅 (w/D)
1-shot 70.1 ± 1.9 77.7 ± 1.6 64.1 ± 1.9 73.5 ± 1.6
5-shot 78.7 ± 0.8 85.2 ± 0.8 77.4 ± 1.8 83.4 ± 0.8


Please cite our paper if it is helpful to your work:

  title={Learning to Self-Train for Semi-Supervised Few-Shot Classification},
  author = {Li, Xinzhe and Sun, Qianru and Liu, Yaoyao and Zhou, Qin and Zheng, Shibao and Chua, Tat-Seng and Schiele, Bernt},


Our implementations use the source code from the following repositories and users:


Learning to Self-Train for Semi-Supervised Few-Shot




No releases published


No packages published