Xavier入门教程软件篇-编译安装PyTorchv1.3.0(jetpack.4.2.3)

Xavier入门教程软件篇-编译安装PyTorch(Caffe2)v1.3.0(jetpack.4.2.3)

说明：

• 介绍如何在Xavier安装PyTorch v1.3.0
• PyTorch已经包含caffe2
• 环境:Jeppack 4.2.3

步骤：

• 新建目录

mkdir -p ~/dl/PyTorch
cd ~/dl/PyTorch

• 最大性能

$ sudo nvpmodel -m 0
$ sudo ~/jetson_clocks.sh

• 下载源码

$ git clone --recursive --branch v1.3.0 http://github.com/pytorch/pytorch
$ cd pytorch
# if you are updating an existing checkout
$ git submodule sync
$ git submodule update --init --recursive

• 配置环境变量

$ export USE_NCCL=0
$ export USE_DISTRIBUTED=0
$ export TORCH_CUDA_ARCH_LIST="5.3;6.2;7.2"
$ export PYTORCH_BUILD_NUMBER=1
$ export PYTORCH_BUILD_VERSION=1.3.0

• 构建wheel 针对python2

$ sudo apt-get install python-pip cmake
$ pip install -U pip

$ sudo pip install -U setuptools
$ sudo pip install -r requirements.txt

$ pip install scikit-build --user

$ python setup.py bdist_wheel

• 构建wheel 针对python3

$ sudo apt-get install python3-pip cmake

$ sudo pip3 install -U setuptools
$ sudo pip3 install -r requirements.txt

$ sudo pip3 install scikit-build 

$ python3 setup.py bdist_wheel

• 编译完成后的whl文件位置

$ cd ~/dl/PyTorch/pytorch/dist
$ ls
torch-1.3.0a0+54a63e0-cp36-cp36m-linux_aarch64.whl

• 安装whell

# Python 2.7 (download pip wheel from above)
$ pip install torch-1.3.0a0+8554416-cp27-cp27mu-linux_aarch64.whl --user

# Python 3.6 (download pip wheel from above)
$ pip3 install torch-1.3.0a0+54a63e0-cp36-cp36m-linux_aarch64.whl --user

• PyTorch v1.3 对应 torchvision v0.5.0
• 源码安装torchvision

$ git clone https://github.com/pytorch/vision
$ cd vision
$ git checkout -b v0.5.0
$ sudo python3 setup.py install

测试：

• 测试cuda和torch

ubuntu@AiROS:~$ python3
Python 3.5.2 (default, Jul 10 2019, 11:58:48) 
[GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> print(torch.__version__)
1.2.0a0+54a63e0
>>> print(torch.cuda.is_available())
True
>>> a = torch.cuda.FloatTensor(2)
>>> print(a)
tensor([0., 0.], device='cuda:0')
>>> b = torch.randn(2).cuda()
>>> print(b)
tensor([ 1.3568, -1.6426], device='cuda:0')
>>> c = a + b
>>> print(c)
tensor([ 1.3568, -1.6426], device='cuda:0')
>>> 

• 测试vision

import torchvision

参考：

• https://pytorch.org/
• https://github.com/pytorch/pytorch
• https://elinux.org/Jetson_Zoo#PyTorch_.28Caffe2.29
• https://devtalk.nvidia.com/default/topic/1049071/jetson-nano/pytorch-for-jetson-nano/