PaddleFSL

This repository contains PaddleFSL, which is a python library for few-shot learning (FSL) building on top of PaddlePaddle2.0.

PaddleFSL provides low-level utilities and unified interface to create new algorithms for FSL, together with high-quality implementations of existing FSL algorithms and standardized benchmarks.

Overview

• examples: using examples of the package, especially high level algorithms including configurations of each task.
• paddlefsl: the package code.
  • paddlefsl.backbones: backbones such as MLP, convolutional networks and ResNet.
  • paddlefsl.datasets: class of datasets commenly used in meta-learning.
  • paddlefsl.model_zoo: high-level algorithms in meta-learning.
  • paddlefsl.task_sampler: taskset for splitting specific task of N ways, K shots and Q query points.
  • paddlefsl.utils: Environment and utilities for meta-reinforcement learning.
• raw_data: the directory to place the raw data files.
• test: unit testing files of the subpackages, functions and classes.

Installation

To use our package, users should first install paddlepaddle-v2.0.0 or later versions, see https://www.paddlepaddle.org.cn/install .

First, install our required packages.

# Clone our repository.
git clone https://github.com/tata1661/FSL-Mate.git
# Install requirements.
cd FSL-Mate/PaddleFSL
pip install -r requirements.txt

Second, update environment variables.

# Please edit env.sh to set the correct path of FSL-paddletoolkit directory.
# Then do:
source env.sh
# If you want to use our package frequently, you can add environment variables into .bashrc
cat env.sh >> ~/.bashrc
source ~/.bashrc

Finally, check whether the installation is successful.

# Start a python interpreter
python
>>> import paddlefsl
>>> paddlefsl.__version__
'1.0.0'

Datasets

We provide several few shot learning datasets and custom dataset interface. Dataset usages please see raw_data/README.md

Few Shot Tasks and High level algorithms

We provide implementations of several high level algorithms in different few shot tasks. Please see examples/README.md

Contributors

PaddleFSL is maintained by Yaqing Wang, Zijing Zhao and Xin Tian.

Please feel free to open an issue or contact us for feedbacks or advices. Thanks!