Jittor是基于JIT编译和元运算符的高性能深度学习框架

Jittor是基于JIT编译和元运算符的高性能深度学习框架

Jittor: a Just-in-time(JIT) deep learning framework

Quickstart | Install | Tutorial | Chinese

Jittor is a high-performance deep learning framework based on JIT compiling and meta-operators. The whole framework and meta-operators are compiled just-in-time. A powerful op compiler and tuner are integrated into Jittor. It allowed us to generate high-performance code with specialized for your model.

The front-end language is python. Module Design is used in the front-end, which is the most popular design for deeplearning framework interface. The back-end is implemented by high performance language, such as CUDA,C++.

The following example shows how to model a two-layer neural network step by step and train from scratch In a few lines of Python code.

import jittor as jtfrom jittor import Modulefrom jittor import nnimport numpy as npclass Model(Module): def __init__(self): self.layer1 = nn.Linear(1, 10) self.relu = nn.Relu() self.layer2 = nn.Linear(10, 1) def execute (self,x) : x = self.layer1(x) x = self.relu(x) x = self.layer2(x) return xdef get_data(n): # generate random data for training test. for i in range(n): x = np.random.rand(batch_size, 1) y = x*x yield jt.float32(x), jt.float32(y)learning_rate = 0.1batch_size = 50n = 1000model = Model()optim = nn.SGD(model.parameters(), learning_rate)for i,(x,y) in enumerate(get_data(n)): pred_y = model(x) dy = pred_y - y loss = dy * dy loss_mean = loss.mean() optim.step(loss_mean) print(f"step {i}, loss = {loss_mean.data.sum()}")

Contents

QuickstartInstallTutorialContributingThe TeamLicense

Quickstart

We provide some jupyter notebooks to help you quick start with Jittor.

Example: Model definition and trainingBasics: Op, VarMeta-operator: Implement your own convolution with Meta-operator

Install

We provide a Docker installation method to save you from configuring the environment. The Docker installation method is as follows:

# CPU onlydocker run -it --network host jittor/jittor# CPU and CUDAdocker run -it --network host jittor/jittor-cuda

Jittor is written in Python and C++. It requires a compiler for JIT compilation, Currently, we support four compilers:

CPU compiler (require at least one of the following) g++ (>=5.4.0)clang (>=8.0) GPU compiler (optional) nvcc (>=10.0 for g++ or >=10.2 for clang)

Jittor environment requirements:

System: Ubuntu >= 16.04 (or Windows Subsystem of Linux)Python version >= 3.7C++ compiler(g++ or clang)

Note: Currently Jittor runs on the Windows operating system through WSL. For the installation method of WSL, please refer to Microsoft official website. WSL does not yet support CUDA.

Jittor offers three ways to install: pip, script or manual.

Pip install

sudo apt install python3.7-dev libomp-devsudo python3.7 -m pip install git+https://github.com/Jittor/jittor.gitpython3.7 -m jittor.test.test_example

single line script install

We provide single line command for quick installation the latest version of Jittor(Ubuntu>=16.04):

# install with clang and cudawget -O - https://raw.githubusercontent.com/Jittor/jittor/master/script/install.sh | with_clang=1 with_cuda=1 bash# install with clangwget -O - https://raw.githubusercontent.com/Jittor/jittor/master/script/install.sh | with_clang=1 bash# install with g++ and cudawget -O - https://raw.githubusercontent.com/Jittor/jittor/master/script/install.sh | with_gcc=1 with_cuda=1 bash# install with g++wget -O - https://raw.githubusercontent.com/Jittor/jittor/master/script/install.sh | with_gcc=1 bash

After execution, the script will show some environment variables you need to export.

If you use Jittor for CPU computing, we strongly recommend clang(>=8.0) as the back-end compiler of Jittor. Because some customized optimizations will be enabled.

manual install

We will show how to install Jittor in Ubuntu 16.04 step by step, Other Linux distributions may have similar commands.

Step 1: Choose your back-end compiler

# g++sudo apt install g++ build-essential libomp-dev# OR clang++-8wget -O - https://raw.githubusercontent.com/Jittor/jittor/master/script/install_llvm.sh > /tmp/llvm.shbash /tmp/llvm.sh 8

Step 2: Install Python and python-dev

Jittor need python version >= 3.7.

sudo apt install python3.7 python3.7-dev

Step 3: Run Jittor

The whole framework is compiled Just-in-time. Let's install jittor via pip

git clone https://github.com/Jittor/jittor.gitsudo pip3.7 install ./jittorexport cc_path="clang++-8"# if other compiler is used, change cc_path# export cc_path="g++"# export cc_path="icc"# run a simple testpython3.7 -m jittor.test.test_example

if the test is passed, your Jittor is ready.

Optional Step 4: Enable CUDA

Using CUDA in Jittor is very simple, Just setup environment value nvcc_path

# replace this var with your nvcc location export nvcc_path="/usr/local/cuda/bin/nvcc" # run a simple cuda testpython3.7 -m jittor.test.test_cuda

if the test is passed, your can use Jittor with CUDA by setting use_cuda flag.

import jittor as jtjt.flags.use_cuda = 1

Optional Step 5: Run full tests

To check the integrity of Jittor, you can run full tests.

python3.7 -m jittor.test -v

if those tests are failed, please report bugs for us, and feel free to contribute ^_^

Tutorial

In the tutorial section, we will briefly explain the basic concept of Jittor.

To train your model with Jittor, there are only three main concepts you need to know:

Var: basic data type of jittorOperations: Jittor'op is simular with numpy

Var

First, let's get started with Var. Var is the basic data type of jittor. Computation process in Jittor is asynchronous for optimization. If you want to access the data, Var.data can be used for synchronous data accessing.

import jittor as jta = jt.float32([1,2,3])print (a)print (a.data)# Output: float32[3,]# Output: [ 1. 2. 3.]

And we can give the variable a name.

c.name('c')print(c.name())# Output: c

Operations

Jittor'op is simular with numpy. Let's try some operations. We create Var a and b via operation jt.float32, and add them. Printing those variables shows they have the same shape and dtype.

import jittor as jta = jt.float32([1,2,3])b = jt.float32([4,5,6])c = a*bprint(a,b,c)print(type(a), type(b), type(c))# Output: float32[3,] float32[3,] float32[3,]# Output:

Beside that, All the operators we used jt.xxx(Var, ...) have alias Var.xxx(...). For example:

c.max() # alias of jt.max(c)c.add(a) # alias of jt.add(c, a)c.min(keepdims=True) # alias of jt.min(c, keepdims=True)

if you want to know all the operation which Jittor supports. try help(jt.ops). All the operation you found in jt.ops.xxx, can be used via alias jt.xxx.

help(jt.ops)# Output:# abs(x: core.Var) -> core.Var# add(x: core.Var, y: core.Var) -> core.Var# array(data: array) -> core.Var# binary(x: core.Var, y: core.Var, op: str) -> core.Var# ......

More

If you want to know more about Jittor, please check out the notebooks below:

Quickstart Example: Model definition and trainingBasics: Op, VarMeta-operator: Implement your own convolution with Meta-operator Advanced Custom Op: write your operator with C++ and CUDA and JIT compile itProfiler: Profiling your modelJtune: Tool for performance tuning

Those notebooks can be started in your own computer by python3.7 -m jittor.notebook

Contributing

Jittor is still young. It may contain bugs and issues. Please report them in our bug track system. Contributions are welcome. Besides, if you have any ideas about Jittor, please let us know.

You can help Jittor in the following ways:

Citing Jittor in your paperrecommend Jittor to your friendsContributing codeContributed tutorials and documentationFile an issueAnswer jittor related questionsLight up the starsKeep an eye on jittor......

Contact Us

Website: http://cg.cs.tsinghua.edu-/jittor/

Email: jittor@qq.com

File an issue: https://github.com/Jittor/jittor/issues

The Team

Jittor is currently maintained by Dun Liang, Guo-Ye Yang, Guo-Wei Yang and Wen-Yang Zhou etc. from the Tsinghua CSCG Group. If you are also interested in Jittor and want to improve it, Please join us!

License

Jittor is Apache 2.0 licensed, as found in the LICENSE.txt file.

版权声明：本文内容由网络用户投稿，版权归原作者所有，本站不拥有其著作权，亦不承担相应法律责任。如果您发现本站中有涉嫌抄袭或描述失实的内容，请联系我们jiasou666@gmail.com 处理，核实后本网站将在24小时内删除侵权内容。