Build and Host Real-world Machine Learning Services from Scratch – PyCon Taiwan 2019
Day 1, R2 13:00–13:45
As Python has become the most popular programming language for research, it is quite easy to clone the source code from state of the art machine learning projects, and then run the demos in your well-prepared local environments.
However, is it easy to turn these demos into production-ready services? How can you make them sustainable? What actual challenges may you face?
In this talk, I will share a story about how we leverage the power of Python to build computer vision services for customers, and further improve these services through a self-designed machine learning pipeline.
Day 2, 15:10–15:40
- Are you sometimes frustrated by the performance of Python?
- Do you always look for open source library to speed up your process?
- Do you feel numpy cannot give you much edge to enhance the performance now?
If you answer yes in any of the above questions, probably writing C extension will be a great solution to you.
First we will go through the reasons to learn writing C extension. Then the modern libraries, e.g. Cython and pybind11, will be introduced to develop C extension and compared with their features and functionalities. Also, a few simple but practical examples are demonstrated that writing C extension can be straightforward, and improves greatly your software quality and performance.
Slides: https://gavincyi.github.io/pycon-why-should-you-learn-writing-c-extension
Speaker: Gavin Chan
Gavin Chan is a principal quantitative developer in AXA Investment Managers Chorus Ltd with 7+ years of experience in software development and finance industry.
...
https://www.youtube.com/watch?v=kBzS-SCN-XM
PyCon APAC 2022|一般演講 Talks|國泰金控 Cathay Financial Holdings / 美光科技 Micron 冠名贊助
✏️ 共筆 Note:https://hackmd.io/@pycontw/HyCRLpQyi
?? Slido:https://app.sli.do/event/fLCSJrx9LtUJq631UaNLuQ
? 語言 Language:英文 English
? 層級 Level:中階 Intermediate
? 分類 Category:自然語言處理 Natural Language Processing
? 摘要 Abstract ?
Math word problem (數學應用問題) is one of the Holy Grail issue in modern AI community, especially in NLP/NLU field. It takes a machine that understands the semantics of human languages and equipts with mathematic skills to solve the equations desribed with natural languages. As Loki NLU engine provides accurate semantic parsing result and SymPy is good at solving equations, I'd love to share my experience of using Loki NLU engine to convert math word problems into equations, then solve them with SymPy. The beauty of these two tools is that they are both Python-based and it only takes basic Python skills to build the math word problem solver. To begin with, I present a simple comparison of NLU systems and their joy and tears (mostly tears) while dealing with math word problems. Then, the hybrid NLU system, Loki, is used to convert math word problems into equations. Finally, I introduce SymPy and how to use it to solve the equations to get the answer of the problem.
? 說明 Description ?
Mainstring math word problem solving procedures take this issue as an "Alignment Problem" or a "Searching Problem." Both appraoches can only get approximately 60% (more or less) accuracy.
On the other hand, Loki (Linguistic Oriented Keyword Interface) is a hybrid NLU system that has both rule-based Structural Pattern Matching module and a ML-based synonym module to provide way much accurate NLP/NLU accuracy and thus a better conversion between natural language and math equations.
We have done this with Loki on Chinese math word problems and written a paper on this topic. The paper is accepted as poster in ROCLING 2019 (2019 台灣計算機語言學年會) as poster paper.
https://aclanthology.org/2020.rocling-1.21.pdf
In this talk, I intent to do the same thing again, but this time I am doing it with with SymPy and with English math word problems.
? 講者介紹 About Speaker - PeterWolf ?
強人工智慧倡議者、設計者與實作者。
Strong AI advocater, designer and practitioners.
#pycontw #pyconapac2022 #python #nlu #sympy #mathwordproblem
Follow “PyCon Taiwan”
⭐️ Official Website: https://tw.pycon.org
⭐️ Facebook: https://www.facebook.com/pycontw
⭐️ Instagram: https://www.instagram.com/pycontw
⭐️ Twitter: https://twitter.com/PyConTW
⭐️ LinkedIn: https://www.linkedin.com/company/pycontw
⭐️ Blogger: https://pycontw.blogspot.com
...
https://www.youtube.com/watch?v=u6SkdB6Oh1E
Day 1, 10:40-11:25
Abstract
DAFunctor aims to reduce the increasing gap between science and engineering. ND-Array is a crucial part of modern algorithm design. It reduces the mental burden for designers and bypasses Python's slow loop. However, using lower-level languages like C/C++ is sometimes unavoidable on productization. Some strict development guidelines like MISRA even forbid dynamic memory allocation, making engineering more difficult. Manual translation to C/C++ usually introduces human errors. Other function-to-function auto translators generate lots of boilerplate code and require many intermediate buffers. On the contrary symbolic translation generates only the essential logic. Imperative programming style makes the performance directly depends on the implementation; symbolic translation can also eliminate part of the inefficiency caused by the programmer, thus make benchmarking more normalized. The making of DAFunctor also involves fun low-level hacks on the Python interpreter.
Slides: https://www.slideshare.net/Buganini/dafunctor-250351849
HackMD: https://hackmd.io/@pycontw/2021/%2F%40pycontw%2FByS9zeYfK
Speaker: Buganini
Pythonista since 2008
...
https://www.youtube.com/watch?v=aaPkvbZj6WU