Editor’s note: Mybridge, which develops reading strategies for developers, has released a New Year’s review of the top 30 open source projects in machine learning in 2017. The projects were selected from more than 8,800 projects with an acceptance rate of just 0.3 percent, providing an efficient guide for data scientists who missed out on many interesting projects. The following is a compilation of nonji.

Over the past year, we compared nearly 8,800 open source machine learning projects and whittled them down to the top 30. This is a very competitive list of the best open source machine learning libraries, data sets, and applications released between January and December 2017, with the average number of Github stars awarded to each project being 3,558.

Open source projects are an important way to participate in practice and deepen learning. In this article, you can read the source code and try to build something on the project. Plus, you’ll have plenty of free time to do some fun things you might have missed out on in 2017.

Basic knowledge

1. Neural Networks

Understanding neural networks is the first step to advanced machine learning, which is relatively elementary and requires only high school math. For pure novelty, you can start by reading about learning from Zero: Understanding and coding neural networks from Python and R.

Deep Learning A-Z™: Hands-on Artificial Neural Networks Contains 23 hours of videos and 22 articles, has accumulated more than 68,000 comments and is rated 4.5/5 stars (Teaching English).

2.TensorFlow

Similarly, learning to use TensorFlow is a prerequisite for starting machine learning projects. The Complete Guide to TensorFlow for Deep Learning with Python

30 open source projects

1.FastText

Address: github.com/facebookresearch/fastText?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

FastText is a free, open source, lightweight library launched by Facebook in 2017 that allows developers to learn text categorization and train in supervising text classifiers. It has the following major advantages over tools like Word2vec:

  • Suitable for large data and efficient training;

  • Support multi-language expression;

  • Facebook has unveiled pre-trained Word Vectors in 90 languages;

  • It may be ported to mobile devices in the future.

2.Deep Photo Style Transfer

Address: github.com/luanfujun/deep-photo-styletransfer?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Deep Photo Style Transfer

Introduction: This is a deep convolutional neural network based image style conversion project, a collaboration between Cornell University and Adobe. Different from the general style transfer, it realizes the style transformation of real scenes, including weather, season and art editing style, while the output image also has high fidelity.

3.face_recognition

Address: github.com/ageitgey/facerecognition?utmsource=mybridge&utmmedium=blog&utmcampaign=read_more

Summary: This project integrates the world’s simplest face recognition library and provides a Python API that allows developers to recognize faces using a deep learning framework. It also provides a simple command-line tool called face_recognition.

4.Magenta

Address: github.com/tensorflow/magenta?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Magenta is a Google Brain project that allows developers to create compelling music and visual art using machine learning.

5.sonnet

Address: github.com/deepmind/sonnet?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Sonnet is an advanced framework for building complex neural networks, built on TensorFlow and opened in 2017 by the DeepMind team. It is based on TensorFlow, but much like Torch. Designed specifically for DeepMind, it adds a number of unique tools to the general neural network library, most notably the StarCraft II API.

6.deeplearn.js

Address: github.com/PAIR-code/deeplearnjs?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Deeplearn.js is a hardware-accelerated open source JavaScript library developed by Google Brain PAIR for machine learning. It provides high-performance machine learning building blocks that allow users to train neural networks in a browser or run pre-trained models in inference mode.

7.Fast Style Transfer

Address: github.com/lengstrom/fast-style-transfer?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Brief introduction: This project focused on image and video style transfer. It was provided by Logan Engstrom of MIT. It combined Gatys’ neural algorithm art style, Fei-fei Li et al. ‘s real-time style transfer and super-resolution perceptual loss, and Ulyanov’s Instance Normalization. Transfer famous painting styles to any image in fractions of a second.

8.PySC2

Address: github.com/deepmind/pysc2?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

PySC2 is a Python component of DeepMind’s open source StarCraft II Learning Environment (SC2LE). It can access the game environment through blizzard’s official Starcraft ii client API and use it as an intensive learning environment (Python). This component provides an interactive interface for the reinforcement learning agent to obtain observations and send action instructions.

9.AirSim

Address: github.com/Microsoft/AirSim?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

AirSim is a virtual training platform for unmanned aerial vehicles (UAVs) and unmanned aerial vehicles (UAVs). It contains tools such as real environment, weather environment, vehicle engine and sensors, providing realistic simulation experience from physical to visual.

10.Facets

Address: github.com/PAIR-code/facets?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Facets is a Google Open source visualization tool that includes Facets Overview and Facets Dive. It can help users understand and analyze machine learning datasets in a visual way, and can be embedded into Jupyter notebook or web for use.

11.Style2Paints

Address: github.com/lllyasviel/style2paints?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Video: www.bilibili.com/video/av14443094/

Description: this is an open source project provided by the domestic team, after entering a style image and line draft, the tool will automatically color the line draft. According to team member “Yisecond Yimeowu” (Zhihuaccounts), Style2Paints realize color cues based on semantic information migration, now available in version 2.0, which is the most accurate, natural and harmonious coloring tool available (Reddit account Q914847518).

12.Tensor2Tensor

Address: github.com/tensorflow/tensor2tensor?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Tensor2Tensor Tensor2Tensor (T2T) is a library of deep learning models and datasets used and maintained by researchers at the Google Brain team. It provides a model named MultiModel, which can be trained and learned in translation, grammar analysis, speech recognition, image recognition, object detection and other tasks. Despite All the hype and controversy it generated in Google’s paper One Model To Learn Them All, MultiModel doesn’t turn out To be that much of a technological advance.

13.pytorch-CycleGAN-and-pix2pix

Address: github.com/junyanz/pytorch-CycleGAN-and-pix2pix?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Summary: Image-to-image conversion in PyTorch.

14.Faiss

Address: github.com/facebookresearch/faiss?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Faiss is a library provided by Facebook Research for efficient similarity search and dense vector clustering. It contains algorithms that can search through any set of size vectors, as well as supporting code for evaluation and parameter debugging.

15.Fashion-mnist

Address: github.com/zalandoresearch/fashion-mnist?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Profile: Fashion-MnIST is a database similar to MNIST, containing a training set of 60,000 samples and a test set of 10,000 samples. This is a fashion database, where each sample is a 28×28 black and white image, divided into 10 categories.

16.ParlAI

Address: github.com/facebookresearch/ParlAI?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

ParlAI is a Conversational AI Research framework (Python) provided by Facebook Research that aims to provide developers with a unified framework for conversational models, centralized conversational data sets, and seamless integration with Amazon Mechanical Turk.

17.Fairseq

Address: github.com/facebookresearch/fairseq?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Brief introduction: Fairseq is a machine translation kit provided by Facebook Research, which realizes using CNN to make time series prediction analysis. Its encoder and decoder are both CNN, and inference is faster. Provide pre-training models for English to French, English to German, English to Romanian translation. But it’s only available in PyTorch.

18.Pyro

Address: github.com/uber/pyro?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Pyro is an extensible deep probabilistic programming library based on PyTorch developed by Uber AI Lab and Stanford University research team. Pyro is designed according to these principles:

  • General: Pyro is a general PPL — it can represent any computable probability distribution;

  • Scalability: Pyro scales to large data sets at a lower cost than handwritten code;

  • Lightweight: Pyro is flexible and maintainable, implemented by powerful, composable abstractions;

  • Flexibility: Pyro aims to automate and control only when needed, expressing generation and reasoning models through high-level abstractions.

19.iGAN

Address: github.com/junyanz/iGAN?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

IGAN is an open source project developed by uc Berkeley and Adobe CTL to generate interactive images based on generative adversarial networks (gans).

20.Deep-image-prior

Address: github.com/DmitryUlyanov/deep-image-prior?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Thesis: sites. Skoltech. Ru/app/data/uploads/sites / 25/2017 / / 12 deepimageprior. PDF

Brief introduction: Deep-image-prior is an open source project based on the results of this paper. Dmitry Ulyanov, Andrea Vedaldi and Victor Lempitsky used convolution filtering to achieve denoising and image repair in the case of only seeing damaged images. The restored images are natural, smooth and clear.

21.face_classification

Address: github.com/oarriaga/faceclassification?utmsource=mybridge&utmmedium=blog&utmcampaign=read_more

Face_classification is a real-time face detection and emotion/gender classification project using keras CNN model and openCV’s FER 2013/IMDB dataset, supporting both still images and dynamic video.

22.Speech-to-Text-WaveNet

Address: github.com/buriburisuri/speech-to-text-wavenet?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Speech-to-text-wavenet is an open source Speech recognition system based on DeepMind’s WaveNet and TensorFlow end-to-end Speech recognition.

23.StarGAN

Address: github.com/yunjey/StarGAN?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Brief introduction: StarGAN is a 2017 collaboration between Korea University, Clova AI Research, The University of New Jersey, and the Hong Kong University of Science and Technology. It uses a single GAN to transform image to image in multiple domains. When used in human face generation, it can not only adjust the hair color, texture, skin color, gender, Can also synthesize a variety of vivid and interesting expressions.

The latest research of HkUST and other universities shows that this GAN can not only be facelifted, but also support online sex change (attached code implementation).

24.Ml-agents

Address: github.com/Unity-Technologies/ml-agents?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Overview: ML-Agents allows developers to create games and simulations using the Unity editor, where the editor can train agents by loading them into a game environment via the Python API, where developers can train their agents using machine learning methods such as reinforcement learning.

25.DeepVideoAnalytics

Address: github.com/AKSHAYUBHAT/DeepVideoAnalytics/?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Deep Video Analytics is a distributed visual search and data visualization platform provided by Dr. Akshay Bhat of Cornell University.

26.OpenNMT

Address: github.com/OpenNMT/OpenNMT?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

OpenNMT is a fully functional open source neural machine translation system. It is simple in design, easy to expand, and provides efficient, advanced, and accurate translation performance.

  • Speed and memory optimization for high-performance GPU training;

  • Simple general-purpose interface, requiring only source and target data files;

  • C++ translator, easy to achieve;

  • Allows other sequence generation tasks (such as image captions) to be extended.

27.Pix2pixHD

Address: github.com/NVIDIA/pix2pixHD?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Thesis: arxiv.org/pdf/1711.11585.pdf

Pix2pixHD is a collaboration between NVIDIA and the University of California, Berkeley. It is capable of conditional GAN synthesis and processing of 2048×1024 images. Recently, researchers have published a paper on how conditional GAN can generate images, which can be seen in detail: synthesizing high-resolution endless roads with GAN through collaborative rendering.

28.Horovod

Address: github.com/uber/horovod?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Horovod is a distributed TensorFlow training framework provided by Uber Engineering. It is easy to use and improves training speed significantly. 90% expansion efficiency can be achieved for Inception V3 and ResNET-101, and 79% for VGG-16.

29.AI-Blocks

Address: github.com/MrNothing/AI-Blocks?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Summary: AI-Blocks is a WYSIWYG (what you see is what you get) machine learning model building interface that can be used to create a simple scene by dragging objects. It can be run directly in the editor or exported to Tensorflow to run a standalone script.

30.deep-voice-conversion

Address: github.com/andabi/deep-voice-conversion?utmsource=mybridge&utmmedium=blog&utmcampaign=readmore

Deep-voice-conversion is a many-to-one voice conversion system that converts your voice to the voice of the target without the need for the target’s voice, text or conversation data, relying only on the spectrogram (wave frequency).

Medium. Mybridge. co/30-amazing- Machine-learning-projects-for-the-past-year-V-2018-b853B8621AC7