【PaddlePaddle Hackathon 2】89、增加 Taichi 和 PaddlePaddle 高效结合的案例

[TCChenlong]

（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：增加 Taichi 和 PaddlePaddle 高效结合的案例。

• 技术标签：深度学习框架，高性能计算。

• 任务难度：简单。

• 详细描述：选取一个 PaddlePaddle 中暂不支持的 op，使用 Taichi 编写该 op 的并行实现，并在一个 PaddlePaddle 和Taichi的交互案例中展示效果。

  • 注意本题目为开放性题目，需要在题目一的基础上完成。

• 相关实现：如需使用可微分op，请参考 Taichi 中与 torch autograd 交互的测试样例

【提交内容】

• 设计文档，并提 PR 至 PaddlePaddle/community 的 rfcs/Taichi 目录

• PR 请提交到用户自己的公开 repo，提交 repo 的链接即可；

• Repo 中需要包含详细的案例使用步骤，以及必要的代码讲解和背景知识。

【技术要求】

• 熟悉 Taichi 和 PaddlePaddle；

• 熟练掌握 C++ 、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。

[0xzhang]

The least squares algorithm is a classic algorithm and PaddlePaddle doesn't directly support for now. Before I finished the first task. I didn't have a proper idea. Now I'm going to implement a taichi-based lstsq example. Like in reference, add a benchmark would be better. Is this an appropriate idea?

https://jekel.me/2019/Compare-lstsq-performance-in-Python/

PyTorch, TensorFlow and NumPy all support linalg.lstsq.

https://www.netlib.org/lapack/lug/node27.html

[yuanming-hu]

I feel like Taichi may not be the optimal tool for lstsq, which is mainly dense regular matrix operations.

Perhaps it's a good idea to use Taichi and its AutoDiff to implement some weird operators that are not provided by PaddlePaddle? Good examples would be NCReLU (ref1, ref2)/MixedConv (ref). Without Taichi, users will need to implement these kernels using CUDA and bind them to PaddlePaddle through C++/Pybind11. With Taichi you can implement everything in pure "Python" with a small amount of code.

(cc @ailzhang who may also be interested in this topic :-))

[0xzhang]

Thanks for your reply!

It must have been influenced by my working background. I have experience implementing various mathematical operations on supercomputers. But I don't know enough knowledges about deep learning. Therefore in deep learning scope, I don't have a good idea about this task.

And I have also understood a key point following your suggestion. Basic mathmatical operator such as lstsq have already be implemented in a library like NumPy. Taichi doesn't need to build wheels. The real value that Taichi creates is, that developers can use Taichi to implement a variety of customized new algorithms more easier. Based on the Taichi framework, Python-like codes could be translated into high-performance, portable machine codes.

Salut, productivity and performance!

[yuanming-hu]

The real value that Taichi creates is, that developers can use Taichi to implement a variety of customized new algorithms.

Exactly! Taichi is very suitable for writing complex kernels that are hard to compose using basic linear algebra components :-)

[0xzhang]

Thanks! I see. However, due to lack of experience in deep learning, there may be someone more suitable for this task.