Introduction: This issue is the AI Briefing for 20201127, bringing you 8 pieces of news, hoping to be helpful to you~

Over 2000 words in total, reading the entire article will take 5-7 minutes.

1. NanoDet: Lightweight (1.8MB), Ultra-Fast (97fps on Mobile) Object Detection Project

Embedded AI Weekly Report: NanoDet Object Detection Model and NCNN Ported to RISC-V

Github:

https://github.com/RangiLyu/nanodet

Recently, a project named nanodet appeared on GitHub, which open-sources a real-time Anchor-free detection model for mobile devices, aiming to provide performance comparable to the YOLO series while being easy to train and port. The project had over 200 stars within just two days.

NanoDet is a super-fast and lightweight Anchor-free object detection model for mobile devices. This model has the following advantages:

Ultra-lightweight: Model file size is only 1.8MB;
Super-fast: Achieves 97fps (10.23ms) on mobile ARM CPUs;
Training-friendly: GPU memory cost is much lower than other models. It can run with a batch size of 80 on a GTX1060 6G;
Easy to deploy: Provides a C++ implementation based on the ncnn inference framework and an Android demo.

2. Under 1000 lines of code, 2000 stars on GitHub, genius hacker open-sources deep learning framework tinygrad

Github:

https://github.com/geohot/tinygrad

Video link:

https://www.youtube.com/channel/UCwgKmJM4ZJQRJ-U5NjvR2dg

In the era of deep learning, tech giants like Google, Facebook, and Baidu have open-sourced several frameworks to help developers learn, build, and train different types of neural networks more easily. These large companies have also invested significant effort in maintaining massive deep learning frameworks like TensorFlow and PyTorch.

Recently, genius hacker George Hotz has open-sourced a small deep learning framework called tinygrad, which combines functionalities of PyTorch and micrograd. The code count of tinygrad is under 1000 lines, and the project has received 2000 stars on GitHub.

“tinygrad may not be the best deep learning framework, but it is indeed a deep learning framework.”

George guarantees that the code count of tinygrad will always be less than 1000 lines.

3. TF Object Detection finally supports TF2!

For a long time, the most commonly used object detection libraries are mmdetection from Chinese University of Hong Kong and detectron2 from Facebook. However, both libraries are based on PyTorch, and the model deployment capabilities of PyTorch are still slightly inferior to TensorFlow. If you want to use TensorFlow’s object detection models, the best choice is still the official TF Object Detection library from Google.

With the arrival of TensorFlow 2.x, the TF Object Detection library now also supports TF2 and importantly, is compatible with TensorFlow 1.x, which is very nice. However, the official still recommends using the latest TF2 to train models.

In addition, TF2 has also added more models such as CenterNet and EfficientDet.

For details, please refer to:

https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2.md

4. Million-dollar AI open-source projects! Covering OCR, object detection, NLP, speech synthesis, and multiple directions

Github:

https://github.com/PaddlePaddle/PaddleHub

Today, I would like to recommend an impressive project,

roughly estimated, this project is worth over a million

Currently with 1.9k stars, but I boldly predict,

this project is sure to become popular, and the future star count should reach 10k or even 20k!

First of all, “no deep learning background required, no data or training process needed,” “sharing the dividends of the AI era,” and “all models are open-source and can run offline.”

It includes text recognition, face detection, image editing, object detection, and more. I can only say,

this repo is a blessing for the lazy!

5. New popular PyTorch tutorial, 900 stars on GitHub in 3 months, ranked 24 on the trend list

Original link:

https://zhuanlan.zhihu.com/p/242235063

GitHub:

https://github.com/lyhue1991/eat_pytorch_in_20_days

If you have previously learned about the popular project “30 Days to Get TensorFlow 2.0,” you will be familiar with this tutorial, which is its sister project “20 Days to Get PyTorch.”

Since the last project “30 Days to Get TensorFlow 2.0” became a hit, the author has been working hard to write “20 Days to Get PyTorch,” which has been completed in 4 months and has already gained over 1800 stars on GitHub, even reaching 24th place on the GitHub trend list.

Both projects use the same formula and have the same familiar taste~

6. How to level up to level 6 on Bilibili without lifting a finger? A senior student’s invention of a number-raising tool, you can too

GitHub: https://github.com/JunzhouLiu/BILIBILI-HELPER

How can you have a little assistant to help you check in every day?

Perhaps, now you can achieve this using GitHub Action scheduled tasks.

This little assistant, lurking on Bilibili, can check in, watch videos, and vote regularly every day…

After completing daily tasks, it will also kindly send you its “task report” for your reference.

Recently, a senior student shared a Bilibili helper tool on GitHub – BILIBILI-HELPER.

He uses GitHub Action scheduled tasks to automatically vote, like, share videos, and check in on live streams every day on Bilibili, helping you easily earn 65 experience points.

The author also encourages everyone to join him in becoming a level 6 expert. (manual dog head)

7. I have ported ncnn to RISC-V!

Original link:

https://zhuanlan.zhihu.com/p/160249065

The author has ported ncnn to RISC-V!

RISC-V, I like to abbreviate it as riscv, saves me a shift and a minus, is an open-source instruction set architecture (ISA) based on the principles of reduced instruction set computing (RISC). As a fully open-source instruction set, it naturally carries the halo of open-source culture, and despite the current market being almost entirely dominated by x86 and ARM, it continues to thrive and develop.

If you ask me why I want to run ncnn on riscv, it is the power of the open-source cultural gene, the magic of the English word meme.

Actually, I encountered some pitfalls during the porting process, and I would like to thank the experts from the Intelligent Software Research Center of the Chinese Academy of Sciences for their enthusiastic answers to my questions.

8. 5G experts discuss at Zijin Mountain, the 2020 China International Embedded Conference held in Nanjing

Original address:

https://36kr.com/p/985999118097030

On November 26, 2020, the China International Embedded Conference (Embedded China) opened grandly in Nanjing. This conference was jointly hosted by the Zijinshan Laboratory, the National Trusted Embedded Software Engineering Technology Research Center, Southeast University, Shanghai Academy of Sciences, Shanghai Industrial Technology Research Institute, and the Yangtze River Delta Embedded System and Software Industry Alliance. More than 300 representatives from industry, academia, and research, as well as well-known academicians and experts such as Academician He Jifeng and Academician Liu Yunjie, will discuss the theme of “5G, new infrastructure creating new ecology” at the conference.

The conference lasts for three days (25th-27th), featuring a main forum on creating a new ecology with 5G new infrastructure and four parallel forums on “Integration of 5G with Artificial Intelligence and Big Data,” “2020 Yangtze River Delta Embedded System Innovation Development,” “Chip and Machine Interaction Special Topic,” “2020 Third National College Student Embedded Chip and System Design Competition and the Fifth Intelligent Interconnection Innovation Competition Awards Ceremony,” and “Education and Professional Talent Cultivation in the Yangtze River Delta.”

Embedded AI Weekly Report: NanoDet Object Detection Model and NCNN Ported to RISC-V

You can add WeChat 17775982065 as a friend, indicating: company + name, to join the RT-Thread official WeChat group!

RT-Thread

Making development of IoT terminals simple and fast, maximizing the value of chips. Apache 2.0 license, free to use in commercial products without disclosing source code, no potential commercial risk.

Long press the QR code to follow us

Click to read the original text to register for the developer conference

Every “look” you click is taken seriously as a like