matlab资源 用于流数据的机器学习和反馈循环的 Python 库 仅供学习参考用代码.zip

<div align="center"> <img src="https://github.com/eserie/wax-ml/blob/main/docs/_static/wax_logo.png" alt="logo" width="40%"></img> </div> # WAX-ML: A Python library for machine-learning and feedback loops on streaming data  [](https://wax-ml.readthedocs.io/en/latest/) [](https://badge.fury.io/py/wax-ml) [](https://codecov.io/gh/eserie/wax-ml) [](https://github.com/ambv/black) [**Quickstart**](#quickstart-colab-in-the-cloud) | [**Install guide**](#installation) | [**Change logs**](https://wax-ml.readthedocs.io/en/latest/changelog.html) | [**Reference docs**](https://wax-ml.readthedocs.io/en/latest/) ## Introduction 🌊 Wax is what you put on a surfboard to avoid slipping. It is an essential tool to go surfing ... 🌊 WAX-ML is a research-oriented [Python](https://www.python.org/) library providing tools to design powerful machine learning algorithms and feedback loops working on streaming data. It strives to complement [JAX](https://jax.readthedocs.io/en/latest/) with tools dedicated to time series. WAX-ML makes JAX-based programs easy to use for end-users working with [pandas](https://pandas.pydata.org/) and [xarray](http://xarray.pydata.org/en/stable/) for data manipulation. WAX-ML provides a simple mechanism for implementing feedback loops, allows the implementation of reinforcement learning algorithms with functions, and makes them easy to integrate by end-users working with the object-oriented reinforcement learning framework from the [Gym](https://gym.openai.com/) library. To learn more, you can read our [article on ArXiv](http://arxiv.org/abs/2106.06524) or simply access the code in this repository. ## WAX-ML Goal WAX-ML's goal is to expose "traditional" algorithms that are often difficult to find in standard Python ecosystem and are related to time-series and more generally to streaming data. It aims to make it easy to work with algorithms from very various computational domains such as machine learning, online learning, reinforcement learning, optimal control, time-series analysis, optimization, statistical modeling. For now, WAX-ML focuses on **time-series** algorithms as this is one of the areas of machine learning that lacks the most dedicated tools. Working with time series is notoriously known to be difficult and often requires very specific algorithms (statistical modeling, filtering, optimal control). Even though some of the modern machine learning methods such as RNN, LSTM, or reinforcement learning can do an excellent job on some specific time-series problems, most of the problems require using more traditional algorithms such as linear and non-linear filters, FFT, the eigendecomposition of matrices (e.g. [[7]](#references)), principal component analysis (PCA) (e.g. [[8]](#references)), Riccati solvers for optimal control and filtering, ... By adopting a functional approach, inherited from JAX, WAX-ML aims to be an efficient tool to combine modern machine learning approaches with more traditional ones. Some work has been done in this direction in [[2] in References](#references) where transformer encoder architectures are massively accelerated, with limited accuracy costs, by replacing the self-attention sublayers with a standard, non-parameterized Fast Fourier Transform (FFT). WAX-ML may also be useful for developing research ideas in areas such as online machine learning (see [[1] in References](#references)) and development of control, reinforcement learning, and online optimization methods. ## What does WAX-ML do? Well, building WAX-ML, we have some pretty ambitious design and implementation goals. To do things right, we decided to start small and in an open-source design from the beginning. For now, WAX-ML contains: - transformation tools that we call "unroll" transformations allowing us to apply any transformation, possibly stateful, on sequential data. It generalizes the RNN architecture to any stateful transformation allowing the implementation of any kind of "filter". - a "stream" module, described in [🌊 Streaming Data 🌊](#-streaming-data-), permitting us to synchronize data streams with different time resolutions. - some general pandas and xarray "accessors" permitting the application of any JAX-functions on pandas and xarray data containers: `DataFrame`, `Series`, `Dataset`, and `DataArray`. - ready-to-use exponential moving average filter that we exposed with two APIs: - one for JAX users: as Haiku modules (`EWMA`, ... see the complete list in our [API documentation](https://wax-ml.readthedocs.io/en/latest/wax.modules.html) ). - a second one for pandas and xarray users: with drop-in replacement of pandas `ewm` accessor. - a simple module `OnlineSupervisedLearner` to implement online learning algorithms for supervised machine learning problems. - building blocks for designing feedback loops in reinforcement learning, and have provided a module called `GymFeedback` allowing the implementation of feedback loop as the introduced in the library [Gym](https://gym.openai.com/), and illustrated this figure: <div align="center"> <img src="docs/tikz/gymfeedback.png" alt="logo" width="60%"></img> </div> ### What is JAX? JAX is a research-oriented computational system implemented in Python that leverages the XLA optimization framework for machine learning computations. It makes XLA usable with the NumPy API and some functional primitives for just-in-time compilation, differentiation, vectorization, and parallelization. It allows building higher-level transformations or "programs" in a functional programming approach. See [JAX's page](https://github.com/google/jax) for more details. ## Why to use WAX-ML? If you deal with time-series and are a pandas or xarray user, b ut you want to use the impressive tools of the JAX ecosystem, then WAX-ML might be the right tool for you, as it implements pandas and xarray accessors to apply JAX functions. If you are a user of JAX, you may be interested in adding WAX-ML to your toolbox to address time-series problems. ## Design ### Research oriented WAX-ML is a research-oriented library. It relies on [JAX](https://jax.readthedocs.io/en/latest/notebooks/quickstart.html) and [Haiku](https://github.com/deepmind/dm-haiku) functional programming paradigm to ease the development of research ideas. WAX-ML is a bit like [Flux](https://fluxml.ai/Flux.jl/stable/) in [Julia](https://julialang.org/) programming language. ### Functional programming In WAX-ML, we pursue a functional programming approach inherited from JAX. In this sense, WAX-ML is not a framework, as most object-oriented libraries offer. Instead, we implement "functions" that must be pure to exploit the JAX ecosystem. ### Haiku modules We use the "module" mechanism proposed by the Haiku library to easily generate pure function pairs, called `init` and `apply` in Haiku, to implement programs that require the management of parameters and/or state variables. You can see [the Haiku module API](https://dm-haiku.readthedocs.io/en/latest/api.html#modules-parameters-and-state) and [Haiku transformation functions](https://dm-haiku.readthedocs.io/en/latest/api.html#haiku-transforms) for more details. In this way, we can recover all the advantages of object-oriented programming but exposed in the functional programming approach. It permits to ease the development of robust and reusable features and to develop "mini-languages" tailored to specific scientific domains. ### WAX-ML works with other libraries We want existing machine learning libraries to work well together while try