FSL-Mate: A collection of resources for few-shot learning (FSL).

Description

try to finish the plan: https://github.com/tata1661/FSL-Mate/issues/24

I don't modify the content of source code but the project structure which is added into some more useful tools to make it more extensive, eg:

• Makefile: workflow scripts
• requirements-dev.txt: dependency packages only used in developemtn
• pyproject.toml: store the project metadata which can be used in many opensource toos, eg: pytest, mypy, pylint.

Another work in this pr: move the paddlefsl to the root dir to make it more clear about the source code.

To get the details, you can read the project tree structure in files of pr.

Description

Try to refactor MAML meta learning algriothm to make it more reusable in paddle-based applications.

Consideration

In MAML module, there are three things which is wired against the normal model application code：

• clone module which retrain the computation graph
• accumulate the gradient on cloned model
• backward the gradient based on query set data

Design

class BaseLearner(ABC):
    """Abstract Base Learner Class"""
    def __init__(self, module: Layer, optimizer: Optimizer) -> None:
        """The constructor of BaseLearner

        Args:
            module (Layer): the model to be trained
        """
        super().__init__()
        self._source_module = module
        self.cloned_module = None
        self.optimizer = optimizer

    def new_cloned_model(self,) -> Layer:
        """get the cloned model and keep the computation gragh

        Returns:
            Layer: the cloned model
        """
        self.cloned_module = clone_model(self._source_module)
        return self.cloned_module

    @abstractmethod
    def adapt(self, train_loss: Tensor) -> None:
        """Adapt the model to the current training loss

        Args:
            train_loss (Tensor): the current training loss
        """
        raise NotImplementedError


    @abstractmethod
    def step(self) -> None:
        """Perform a step of training

        Args:
            loss (float): _description_

        Raises:
            NotImplementedError: _description_
        """
        raise NotImplementedError

    def clear_grad(self):
        """clear the gradient in the computation graph
        """
        self.optimizer.clear_grad()

• adapt: accumulate the gradient on cloned model
• new_cloned_model: clone and save the model
• step: run step on optimizer in parameters of source model
• clear_grad: clear the gradient

Introduction

In this PR, it has done the followting things:

• [ ] do unit test using github action when pushed into repo.
• [ ] automaticlly deploy the package to the pypi server when the master branch changed

The document code base I will do it in next PR so that we can focus different things in different pr.

Run the sample code and report an error. The error information and code information are as follows.

ValueError: (InvalidArgument) Tensor holds the wrong type, it holds int, but desires to be int64_t. [Hint: Expected valid == true, but received valid:0 != true:1.] (at C:\home\workspace\Paddle_release\paddle/fluid/framework/tensor_impl.h:33) [operator < softmax_with_cross_entropy > error]

from paddlefsl.datasets import MiniImageNet import paddle import paddlefsl from paddlefsl.model_zoo import maml

TRAIN_DATASET = paddlefsl.datasets.MiniImageNet(mode='train') VALID_DATASET = paddlefsl.datasets.MiniImageNet(mode='valid') TEST_DATASET = paddlefsl.datasets.MiniImageNet(mode='test') MODEL = paddlefsl.backbones.Conv(input_size=(3, 84, 84), output_size=5)

def main(): train_dir = maml.meta_training(train_dataset=TRAIN_DATASET, valid_dataset=VALID_DATASET, ways=5, shots=1, model=MODEL, meta_lr=0.002, inner_lr=0.03, iterations=10000, meta_batch_size=32, inner_adapt_steps=5, report_iter=10)

if name == 'main': main()

Hi!

Thanks for creating this awesome repo!

Would this paper by chance fit in this repo:

Multi-level Semantic Feature Augmentation for One-shot Learning [paper] [code]

In rc_position_embedding.py line 69:

head_position_vector = paddle.to_tensor(rc_vector[:, :, -1], dtype='int64')

But in rc_init_vector.py line 208, the head_position_idx is -2 in rc_embedding.

So I think it should be:

head_position_vector = paddle.to_tensor(rc_vector[:, :, -2], dtype='int64')

Introduction

After https://github.com/tata1661/FSL-Mate/pull/23 merged, we can convert all of examples, under examples/image_classification, to meta learner based examples.

I have an online meeting with @tata1661 before, and we believe that it will be more concise if the model-zoo examples be replaced with meta learner (optim-based) . Let's get it on the way.

Bumps pillow from 8.2.0 to 8.3.2.

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

在examples/image_classification/添加 gnn_image_classification.py 用于MiniImageNet测试 在examples/relation_classification/ 添加 gnn_relation_classification.py 用于FewRel测试 在backbones/ 添加 gnn_iclr.py 在model_zoo/ 添加 gnn.py 在test/backbones/ 添加 gnn_iclr_test.py

（此 ISSUE 为 PaddlePaddle Hackathon 第二期活动的任务 ISSUE，更多详见 【PaddlePaddle Hackathon 第二期】任务总览）

【任务说明】

• 任务标题：在PaddleFSL将模型库中的MAML和ANIL转为优化算法api
• 模型技术标签：PaddlePaddle, few-shot learning, meta learning
• 任务难度：困难
• 详细描述：在PaddleFSL中，小样本学习经典算法MAML和ANIL以模型方式提供。要求将它们转换成任何模型可调用的算法api。需要重新封装MAML和ANIL，设计其使用方式。可以参考learn2learn和higher的设计。

【提交内容】

• 设计文档，并提 PR 至 PaddlePaddle/community 的 rfcs/FSL-Mate 目录
• 任务 PR 到 PaddleFSL
• 任务单测文件
• 调用路径：paddlefsl.metaopt.maml和paddlefsl.metaopt.anil

【合入标准】

在Omniglot和miniImageNet数据集的5-way 1-shot任务和5-way 5-shot任务进行测试。使用算法api版的MAML, ANIL获得与原汇报结果一致或更高的效果。

【技术要求】

熟练掌握 Python

【答疑交流】

• 如果在开发中对于上述任务有任何问题，欢迎在本 ISSUE 下留言交流。
• 对于开发中的共性问题，在活动过程中，会定期组织答疑，请大家关注官网&QQ群的通知，及时参与。

There are some unused files in the project, eg:.DS_Store file and __pycache__ directory.

You have set the configuration to ignore .DS_Store file, but it only ignore the file under the top directory of project which is disable for another .DS_Store files. So as the __pycache__ directory.

运行gnn的例子时，提示报错Exception has occurred: ModuleNotFoundError，No module named 'paddle.nn' 我的环境是 1080TI Ubuntu20.04 nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2021 NVIDIA Corporation Built on Fri_Dec_17_18:16:03_PST_2021 Cuda compilation tools, release 11.6, V11.6.55 Build cuda_11.6.r11.6/compiler.30794723_0 NVIDIA-SMI 510.39.01 Driver Version: 510.39.01 CUDA Version: 11.6 在conda中安装了paddlepaddle-gpu版本2.3.1 paddle-bfloat 0.1.7 pypi_0 pypi paddlehelix 1.0.1 pypi_0 pypi paddlepaddle-gpu 2.3.1.post116 pypi_0 pypi pandas 1.3.5 pypi_0 pypi pgl 2.2.3.post0 pypi_0 pypi

Install problem 1.1.0

File "/home/fangsifan/Few_Shot/FSL-Mate/PaddleFSL/paddlefsl/backbones/gin.py", line 17, in from pahelix.model_zoo.pretrain_gnns_model import PretrainGNNModel

ModuleNotFoundError: No module named 'pahelix'

Description

In the suggested structuring python project, there are some conventional structures & files & naming-style to make the project more readable and scalable. And what's more, there are some github Action tools to run CI to lint & test code, run CD to publish the project.

Thought

I also maintain the project: python-wechaty and paddle-prompt. So I'm familiar with the related opensource toolkits in python project. I can create an another pr to complete this If you agree with it ? How do you think about it ? @tata1661

大家好，非常高兴地告诉大家，第二期 PaddlePaddle Hackathon 开始了。PaddlePaddle Hackathon 是面向全球开发者的深度学习领域编程活动，鼓励开发者了解与参与 PaddlePaddle 开源社区。本次共有四大专区：PaddlePaddle、Paddle Family、Paddle Friends、Paddle Eval，共计100+个任务供大家认领。详细信息可以参考 PaddlePaddle Hackathon 说明。大家是否已经迫不及待了呢~

本 ISSUE 是 Paddle Friends 专区 FSL-Mate 方向任务合集。具体任务列表如下：

| 序号 | 难度 | 任务 ISSUE | | ---- | ---- | --- | | 91 | ⭐️ | 新增random search超参数搜索api | | 92 | ⭐️⭐️ | 新增TPE超参数搜索api | | 93 | ⭐️⭐️⭐️ | 在PaddleFSL将模型库中的MAML和ANIL转为优化算法api |

若想要认领本次活动任务，请至 PaddlePaddle Hackathon 2 Pinned ISSUE 完成任务 ISSUE 认领。

活动官网：PaddlePaddle Hackathon 2

（此 ISSUE 为 PaddlePaddle Hackathon 第二期活动的任务 ISSUE，更多详见 【PaddlePaddle Hackathon 第二期】任务总览）

【任务说明】

• 任务标题：新增random search超参数搜索api
• 模型技术标签：PaddlePaddle，hyperparameter optimization
• 任务难度：简单
• 详细描述：随机搜索即在搜索空间随机的搜索超参数。它是一种不需要优化问题梯度的数值优化方法，也是常用的基线超参数搜索算法。可参考NNI或HyperOpt，实现random search的超参数搜索api。

【提交内容】

• 设计文档，并提 PR 至 PaddlePaddle/community 的 rfcs/FSL-Mate 目录
• 任务 PR 到 PaddleFSL
• 任务单测文件
• 调用路径：paddlefsl.hpo.rand

【合入标准】

在Omniglot和miniImageNet数据集的5-way 1-shot任务和5-way 5-shot任务进行测试。使MAML, ANIL, ProtoNet and RelationNet使用random search 搜索出的超参数能达到比原汇报结果更高的效果。

【技术要求】

熟练掌握 Python

【答疑交流】

• 如果在开发中对于上述任务有任何问题，欢迎在本 ISSUE 下留言交流。
• 对于开发中的共性问题，在活动过程中，会定期组织答疑，请大家关注官网&QQ群的通知，及时参与。

（此 ISSUE 为 PaddlePaddle Hackathon 第二期活动的任务 ISSUE，更多详见 【PaddlePaddle Hackathon 第二期】任务总览）

【任务说明】任务标题：新增TPE超参数搜索api

• 模型技术标签：PaddlePaddle, hyperparameter optimization
• 任务难度：中等
• 详细描述：树状结构Parzen估计方法(treeparzen estimator method，TPE)是一种常用的基于树状结构 Parzen 密度估计的非标准贝叶斯优化算法，在高维空间表现很好。可参考scikit-optimize或NNI，实现TPE超参数搜索api。

【提交内容】

• 设计文档，并提 PR 至 PaddlePaddle/community 的 rfcs/FSL-Mate 目录
• 任务 PR 到 PaddleFSL
• 任务单测文件
• 调用路径：paddlefsl.hpo.tpe

【合入标准】

在Omniglot和miniImageNet数据集的5-way 1-shot任务和5-way 5-shot任务进行测试。使MAML, ANIL, ProtoNet and RelationNet使用Bayesian optimization搜索出的超参数能达到比原汇报结果更高的效果。

【技术要求】

熟练掌握 Python，理解贝叶斯优化

【答疑交流】

• 如果在开发中对于上述任务有任何问题，欢迎在本 ISSUE 下留言交流。
• 对于开发中的共性问题，在活动过程中，会定期组织答疑，请大家关注官网&QQ群的通知，及时参与。