Shulou(Shulou.com)06/02 Report--

2018 is coming to an end. Let's review what interesting things have happened in the field of machine learning in the past year.

Let's first take a look at the top open source projects in Mybridge AI, then talk about the developments in machine learning this year, and finally explore what we can expect in the new year.

Top open source projects

BERT

BERT, whose full name is Bidirectional Encoder Representations from Transformers, is a new method to solve natural language processing based on TensorFlow, and its performance is better. We can use the pre-training model in BERT to solve the problem, which has great advantages in performance, such as being able to identify the context in a sentence. It is very popular in Github, with 8848 stars. For complete academic papers, please visit here.

Https://github.com/google-research/bert

Https://arxiv.org/abs/1810.04805

DeepCreamPy

DeepCreamPy is a deep learning tool that can reconstruct deleted areas of an image like Photoshop. We use image editing tools (such as PS) to fill the deleted area with green, and the neural network can restore it. The project has 6365 stars in Github.

Https://github.com/deeppomf/DeepCreamPy

TRFL

The TRFL project can be used to write reinforcement learning agents in TensorFlow. The specific documentation is here.

Https://github.com/deepmind

Https://github.com/deepmind/trfl/blob/master/docs/index.md

Horizon

Horizon is a reinforcement learning platform based on PyTorch and uses Caffe2 to provide services for the model. The main advantage of Horizon is that designers consider production use cases when designing this platform. For more details, check out the FacebookResearch official documentation. In addition, if you want to use Horizon, you can view the documentation.

Https://github.com/facebookresearch/Horizon?

Https://heartbeat.fritz.ai/introduction-to-pytorch-for-deep-learning-5b437cea90ac

Https://github.com/facebookresearch/Horizon/blob/master/docs/usage.md

DeOldify

DeOldify is a deep learning library for shading and restoring old images. Developers have combined several different methods to achieve this goal, including: generation confrontation Network with self-attention Mechanism (Self-Attention GenerativeAdversarial Networks), Progressive Growing of GANs, and TTUR (TwoTime-Scale Update Rule).

Https://github.com/jantic/DeOldify

Https://arxiv.org/abs/1805.08318

Https://arxiv.org/abs/1710.10196

Https://arxiv.org/abs/1706.08500

AdaNet

AdaNet is a TensorFlow-based library that automatically learns models without the involvement of many technicians. The project is based on the AdaNet algorithm. To access the official documentation of AdaNet, please click here.

Https://github.com/tensorflow/adanet

Http://proceedings.mlr.press/v70/cortes17a.html

Https://adanet.readthedocs.io/

Graph Nets

Graph Nets is the DeepMind library used to build Sonnet and TensorFlow. The Graph network imports a graph, and the output is also a graph.

Https://github.com/deepmind/graph_nets

Maskrcnn-benchmark

The Maskrcnn-benchmark project can help us build object detection and segmentation tools in Pytorch. This library has the advantages of high speed, high memory efficiency, multiple GPU training and inference, and provides CPU support for inference.

Https://github.com/facebookresearch/maskrcnn-benchmark

PocketFlow

The PocketFlow project is a framework for accelerating and compressing deep learning models. It solves the computational cost problem of most deep learning models. The project was originally developed by researchers at Tencent AI Lab. For more information on its implementation and official documentation, please click here.

Https://github.com/Tencent/PocketFlow

Https://pocketflow.github.io/

MAMEToolkit

MAMEToolKit is a library that trains reinforcement learning algorithms for arcade games. Using this tool, you can track the state of the game and receive game frame data.

Https://github.com/M-J-Murray/MAMEToolkit

The main Development of Machine Learning Framework

PyTorch 1.0

During the PyTorch conference in October, Facebook released a preview of PyTorch 1.0. PyTorch 1.0 solves the following problems: time-consuming training, networking problems, slow extensibility, and some inflexibility brought about by the Python programming language.

PyTorch 1.0 introduces a set of compilation tools, Torch.jit, which will bridge the gap between production and research. Torch.jit includes the Torch Script language in Python, and in PyTorch 1.0, we can build models using graphical patterns, which are useful in developing high-performance and low-latency applications.

Auto-Keras

You may have heard of automated machine learning (automated machine learning), which automates the search for the best parameters of a machine learning model. In addition to Auto-Keras, there are other automated machine learning models, such as Google's AutoML. Auto-Keras is written based on Keras and ENAS, where ENAS is the latest version of neural network structure search.

Https://cloud.google.com/automl/

Https://autokeras.com/

Https://keras.io/

Https://arxiv.org/abs/1802.03268

Https://en.wikipedia.org/wiki/Neural_architecture_search

TensorFlow Serving

Using the TensorFlow Serving system, we can deploy the TensorFlow model to the production environment more easily. Although TensorFlow Serving was released in 2017, this year it pays more attention to applying the model to the production environment.

Https://www.tensorflow.org/serving/

Machine Learning Javascript

There are already Javascript frameworks, such as TensorFlow.js and Keras.js, that allow developers to run models on browsers. The model is implemented and used in a way that is very similar to conventional frameworks such as Keras or TensorFlow.

Https://js.tensorflow.org/

Https://github.com/transcranial/keras-js

look into the future

2019 is just around the corner, and with the development of automation tools such as Auto-Keras, the job of developers is expected to become easier. In addition, we also have advanced research and excellent communities, and the performance of all kinds of machine learning frameworks will be climbed to another high-rise.

Ali Yunyunqi community organization translator. The original title of the article "2018 Year-in-Review: Machine Learning Open Source Projects & Frameworks" translator: Mags, revision: yuan Hu.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.