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

Last update: Dec 29, 2022

Related tags

Overview

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

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.

Installation

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: https://www.anaconda.com/download/

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 https://github.com/xinzheli1217/learning-to-self-train.git 
cd learning-to-self-train

Project Architecture

.
├── data_generator              # dataset generator 
|   └── meta_data_generator.py  # data genertor for meta-train phase
├── models                      # tensorflow model files 
|   ├── models.py               # resnet12 CNN class
|   └── meta_model_LST.py       # semi-supervised meta-train model class
├── trainer                     # tensorflow trianer files  
|   └── meta_LST.py             # semi-supervised meta-train trainer class
├── utils                       # a series of tools used in this repo
|   └── misc.py                 # 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
|
├── exp_train.py                # the python file with main function and parameter settings for meta-training
└── exp_test.py                 # 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 exp_train.py --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 exp_test.py --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 exp_train.py and the exp_test.py 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 exp_train.py and exp_test.py.

Performance

(%)	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`

Citation

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

@inproceedings{li2019lst,
  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},
  booktitle={NeurIPS},
  year={2019}
}

Acknowledgements

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

Comments

[help] How to use a custom dataset with this model?

Dear author, When I run on dataset FC100, I get the following error: FileNotFoundError: [Errno 2] No such file or directory: '/home/xxx/program/learning-to-self-trainmaster/code_for_LTTL/pretrain_weights_dir/FC100/pretrain_weights/fc_weights_1shot.npy' How to get to fc_weights_1shot.npy and how to get unlabeled data from other datasets?

opened by Ping-Wang1 0
Pre-trained model links are not available

Hi! First, thank you very much for this implementation :D I'm trying download the pre-trained models but the link is without access and I need request to download

opened by izmcm 0
np.load error './pretrain_weights_dir/miniImagenet/pretrain_weights/weights_14000.npy'

edit: just noticed it was pickled in py 2.7, used load with latin encoding and all is good Hi, I can't seem to load this file, np.load('./pretrain_weights_dir/miniImagenet/pretrain_weights/weights_14000.npy') I am getting an unpickling error. it's in line 81 inside trainer.meta_LST.py

DId anyone else encounter this?

Thanks, Amit

opened by alfassy 0
code for MAML based LST

Hi Xinzhe,

I am doing some baselines based on 4-layer CNN, could you please share your MAML based implementation for LST? so that I can add your approach to our baselines.

Thank you!

opened by Hugo101 0
code for the CVIU paper

Hello Xinzhe,

Thanks for your great work! May I request the code for the extended version of this paper which has been published in the CVIU journal? That would be great if we can have the implementation and compare different approaches.

Thanks! Best, Changbin

opened by Hugo101 0
pretrained swn_weights_xxx.npy are not provided in the link of pretrained models.

Hi! First, thank you very much for this project. I try to run the meta-test phase separately, without meta-training, just based on the pre-trained models provided online. But I found that the files "swn_weights_xxx.npy" are not provided.

Would it be possible to provide pretrained "swn_weights_xxx.npy"? Or, I must firstly run meta-train phase based on the pre-trained models, as well as save the parameters. Only after this step, I can run meta-test phase based on the saved weights?

opened by wqhIris 0
Clarification regarding total number of meta-train,meta-test tasks and shot number.

Hello, first great work!

I am trying to re-implement this paper in Pytorch! And I know this codebase is based on the official meta-transfer learning implementation (https://github.com/yaoyao-liu/meta-transfer-learning) since it's basically the same group of authors.

I hope to ask for some few clarifications.

The total number of tasks during meta-training is equal to the number of meta-train iterations right? Which is 15000? Also the total number of tasks during meta-testing AND meta-validation is still 600? Both the same as in the MTL implementation?

Moreover, the number of shots of the support and query during meta-train in one-shot setting is 1 (or 5 for 5-shot) and 15 respectively right? While during meta-validation and meta-testing.,the number of instances per class (shot) for the support set and query is uniformly 1 (or 5 for 5-shot)?

Thank you for your time.

opened by JKDomoguen 1
Meta SSL learned ckpts

Hi,

Thank you for putting this repo together.
I'm able to get through just about everything, and I'm very grateful for that.
For convenience, would you be able to share a SSL meta-learned ckpt as well (preferably the one as reported in the paper!) I'm able to run the SSL meta-learning, but looks like it'll take a few days. Since I'm only interested in evaluating the model, it'd be great to have your ckpt.

Thanks! Jackson

opened by wangkua1 0

Owner

GitHub

Code for T-Few from "Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning"

T-Few This repository contains the official code for the paper: "Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learni

220 Dec 31, 2022

Few-NERD: Not Only a Few-shot NER Dataset

Few-NERD: Not Only a Few-shot NER Dataset This is the source code of the ACL-IJCNLP 2021 paper: Few-NERD: A Few-shot Named Entity Recognition Dataset.

319 Dec 30, 2022

Code for 'Self-Guided and Cross-Guided Learning for Few-shot segmentation. (CVPR' 2021)'

SCL Introduction Code for 'Self-Guided and Cross-Guided Learning for Few-shot segmentation. (CVPR' 2021)' We evaluated our approach using two baseline

34 Oct 8, 2022

Implementation of the paper "Self-Promoted Prototype Refinement for Few-Shot Class-Incremental Learning"

Self-Promoted Prototype Refinement for Few-Shot Class-Incremental Learning This is the implementation of the paper "Self-Promoted Prototype Refinement

78 Dec 2, 2022

Project looking into use of autoencoder for semi-supervised learning and comparing data requirements compared to supervised learning.

2 Dec 17, 2021

Self-training for Few-shot Transfer Across Extreme Task Differences

Self-training for Few-shot Transfer Across Extreme Task Differences (STARTUP) Introduction This repo contains the official implementation of the follo

33 Oct 31, 2022

UniMoCo: Unsupervised, Semi-Supervised and Full-Supervised Visual Representation Learning

UniMoCo: Unsupervised, Semi-Supervised and Full-Supervised Visual Representation Learning This is the official PyTorch implementation for UniMoCo pape

49 Jan 2, 2023

ST++: Make Self-training Work Better for Semi-supervised Semantic Segmentation

ST++ This is the official PyTorch implementation of our paper: ST++: Make Self-training Work Better for Semi-supervised Semantic Segmentation. Lihe Ya

147 Jan 3, 2023

Hybrid CenterNet - Hybrid-supervised object detection / Weakly semi-supervised object detection

Hybrid-Supervised Object Detection System Object detection system trained by hybrid-supervision/weakly semi-supervision (HSOD/WSSOD): This project is

5 Dec 10, 2022

Unified Pre-training for Self-Supervised Learning and Supervised Learning for ASR

UniSpeech The family of UniSpeech: UniSpeech (ICML 2021): Unified Pre-training for Self-Supervised Learning and Supervised Learning for ASR UniSpeech-

282 Jan 9, 2023

Pytorch Implementation for CVPR2018 Paper: Learning to Compare: Relation Network for Few-Shot Learning

LearningToCompare Pytorch Implementation for Paper: Learning to Compare: Relation Network for Few-Shot Learning Howto download mini-imagenet and make

246 Dec 19, 2022

Ever felt tired after preprocessing the dataset, and not wanting to write any code further to train your model? Ever encountered a situation where you wanted to record the hyperparameters of the trained model and able to retrieve it afterward? Models Playground is here to help you do that. Models playground allows you to train your models right from the browser.

Models Playground ??️ Upload a Preprocessed Dataset ?? Choose whether to perform Classification or Regression ?? Enter the Dependent Variable ?

19 Dec 10, 2022

Few-shot Learning of GPT-3

Few-shot Learning With Language Models This is a codebase to perform few-shot "in-context" learning using language models similar to the GPT-3 paper.

224 Dec 28, 2022

Library of various Few-Shot Learning frameworks for text classification

FewShotText This repository contains code for the paper A Neural Few-Shot Text Classification Reality Check Environment setup # Create environment pyt

47 Jan 3, 2023

Few-Shot Graph Learning for Molecular Property Prediction

Few-shot Graph Learning for Molecular Property Prediction Introduction This is the source code and dataset for the following paper: Few-shot Graph Lea

94 Dec 12, 2022

Few-shot Relation Extraction via Bayesian Meta-learning on Relation Graphs

Few-shot Relation Extraction via Bayesian Meta-learning on Relation Graphs This is an implemetation of the paper Few-shot Relation Extraction via Baye

36 Nov 22, 2022

True Few-Shot Learning with Language Models

This codebase supports using language models (LMs) for true few-shot learning: learning to perform a task using a limited number of examples from a single task distribution.

124 Jan 4, 2023

Adaptive Prototype Learning and Allocation for Few-Shot Segmentation (CVPR 2021)

ASGNet The code is for the paper "Adaptive Prototype Learning and Allocation for Few-Shot Segmentation" (accepted to CVPR 2021) [arxiv] Overview data/

91 Dec 23, 2022

Spatial Contrastive Learning for Few-Shot Classification (SCL)

This repo contains the official implementation of Spatial Contrastive Learning for Few-Shot Classification (SCL), which presents of a novel contrastive learning method applied to few-shot image classification in order to learn more general purpose embeddings, and facilitate the test-time adaptation to novel visual categories.

34 Dec 25, 2022