qengineering / pytorch-jetson-nano

PyTorch installation wheels for Jetson Nano

PyTorch wheels for Jetson Nano, TX2, (AGX) Xavier

Find your operating system and Pytorch version in the table below. Follow the instructions in the provided guide.
The Jetson Nano wheels supports CUDA 10.2, cuDNN 8.0 and NEON.

---

PyTorch 2.0

Pytorch 2.0 and above uses CUDA 11. The Jetson Nano has CUDA 10.2.
Due to low-level GPU incompatibility, installing CUDA 11 on your Nano is impossible.
Pytorch 2.0 can only be installed on Jetson family members using a JetPack 5.0 or higher, such as the Jetson Nano Orion.
Unfortunately, it does not appear that this version will also be available for the Jetson Nano soon.

---

PyTorch 1.13.0, 1.12.0 and 1.11.0

PyTorch version 1.11 and above requires Python 3.7.
Since JetPack 4.6 has Python 3.6, you cannot install PyTorch 1.11.0 on a Jetson Nano.

It looks like Nvidia has no plans to release JetPack 5.0 for the Jetson Nano for now. It's only available for the Xavier series.
You can use the current experimental version of Jetson Nano with Ubuntu 20.04. We supply the wheels for this version.

---

Wheel: the installation wheel torch-version-cpxx-cpxx-linux_aarch64.whl (xx is the used python version)
Vision: the accompanying torchvision.
LibTorch: the C++ API for those who like to program. (The aarch64 version of libtorch-cxx11-abi-shared-with-deps-1.10.1+cu102.zip)
Guide: link to the installation tutorial.

Roadmap.

Operating system	PyTorch 2.0.0	PyTorch 1.13.0	PyTorch 1.12.0	PyTorch 1.11.0	PyTorch 1.10.0	PyTorch 1.9.0	PyTorch 1.8.0	PyTorch 1.7.0
Jetson Nano Ubuntu 20.04 (Python 3.8)		Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch TorchText Guide TorchText gist
Jetson Nano JetPack 4.6 (Python 3.6)					Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision Guide	Wheel Vision Guide
Raspberry Pi 64-bit Bullseye (Python 3.9)	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide
Raspberry Pi 64-bit Buster (Python 3.7)		Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide
Raspberry Pi Ubuntu 18.04 (Python 3.6)							Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide
Raspberry Pi Ubuntu 20.04 (Python 3.8)					Wheel Vision LibTorch Guide	Wheel Vision LibTorch Guide

We compiled the Pytorch 1.x versions with the clang compiler to prevent issues with the ARM NEON registers and the GNU compiler. For instance #61110 and #65673.




Pytorch 2.0 was build with the GNU compiler.

❗ C++ programmers please note

You should also use the clang compiler (version 8) if you want to compile Pytorch 1.x versions C++ code yourself.
The GNU GCC compiler will give you 'no expression errors'.

# set clang compiler at the command line
$ export CC=clang
$ export CXX=clang++

Don't worry if you plan to use Python. It only applies to C++ users.

---

Find PyTorch and TorchVision with other frameworks and deep-learning examples on our SD-image

---

pytorch-jetson-nano's Issues

arm64

My platform is xavier(arm64).
envriment:
python:3.6
libtorch:1.8
cuda:10.2

I want to using sourse code to build torchvision.
I find that the current compilation is x86.
Can torchvision be compiled on arm64?

Torchtext for pytorch 1.12

Did you have try compiled torchtext wheel that worked on pytorch 1.12 for jetson nano?

PyTorch 1.11.0 Built with GPU Support

Hi, just wondering if the wheel for Jetson Nano PyTorch 1.11.0 supports GPU? After installing the wheel torch.cuda.is_available() returns False.

Thank you! :)

torchvision 0.11 for cp36 and torch 1.11?

NVIDIA provides only Torch 1.11.0 for Jetpack 4.6.1 (Python 3.6) for the Jetson Nano (https://developer.download.nvidia.com/compute/redist/jp/v461/pytorch/torch-1.11.0a0+17540c5+nv22.01-cp36-cp36m-linux_aarch64.whl). I haven't found the corresponding Torchvision.

If I install your Torch 1.10 it won't import (it complains of missing libmpi_cxx.so).

If I intend to install your Torchvision 0.11 (torchvision-0.11.0a0), it removes the nvidia torch 1.11....

I haven't found any working combination...

Would you build the Torchvision 0.11 for nvidia Torch 1.11.0 ?

Jetson Nano, Jetpack 4.6

Hi Team
Thank you for your efforts to build and make the torch files available!

I want to know two things

1. Are the torch files built using CUDA? I had installed the default torch provided by nvidia it doesnt use CUDA.
2. I am getting an error as below when i install any of the 4 torch files (1.7 to 1.10).

"import torch; print(torch.__version__)"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "venv/lib/python3.6/site-packages/torch/__init__.py", line 196, in <module>
    _load_global_deps()
  File "venv/lib/python3.6/site-packages/torch/__init__.py", line 149, in _load_global_deps
    ctypes.CDLL(lib_path, mode=ctypes.RTLD_GLOBAL)
  File "/usr/lib/python3.6/ctypes/__init__.py", line 348, in __init__
    self._handle = _dlopen(self._name, mode)
OSError: libcurand.so.10: cannot open shared object file: No such file or directory

Do you know what could be causing the issue?
Thank your for your help in advance!

Bugs about torchvision installation

My JETPACK version is 4.6.2, as the documentation says, the torch version I installed is 1.10.0, and when I try to install the included torchvision wheel, it throws an error
`jshzt@ubuntu:~/hzt/whl$ pip3 install torch-1.10.0a0+git36449ea-cp36-cp36m-linux_aarch64.whl
Processing ./torch-1.10.0a0+git36449ea-cp36-cp36m-linux_aarch64.whl
Collecting typing-extensions (from torch==1.10.0a0+git36449ea)
Using cached https://files.pythonhosted.org/packages/45/6b/44f7f8f1e110027cf88956b59f2fad776cca7e1704396d043f89effd3a0e/typing_extensions-4.1.1-py3-none-any.whl
Collecting dataclasses; python_version < "3.7" (from torch==1.10.0a0+git36449ea)
Using cached https://files.pythonhosted.org/packages/fe/ca/75fac5856ab5cfa51bbbcefa250182e50441074fdc3f803f6e76451fab43/dataclasses-0.8-py3-none-any.whl
Installing collected packages: typing-extensions, dataclasses, torch
Successfully installed dataclasses-0.8 torch-1.10.0a0+git36449ea typing-extensions-4.1.1

jshzt@ubuntu:~/hzt/whl$ pip3 install torchvision-0.11.0a0+fa347eb-cp36-cp36m-linux_aarch64.whl
Processing ./torchvision-0.11.0a0+fa347eb-cp36-cp36m-linux_aarch64.whl
Collecting numpy (from torchvision==0.11.0a0+fa347eb)
Using cached https://files.pythonhosted.org/packages/51/60/3f0fe5b7675a461d96b9d6729beecd3532565743278a9c3fe6dd09697fa7/numpy-1.19.5.zip
Collecting torch (from torchvision==0.11.0a0+fa347eb)
Could not find a version that satisfies the requirement torch (from torchvision==0.11.0a0+fa347eb) (from versions: )
No matching distribution found for torch (from torchvision==0.11.0a0+fa347eb)
`

google drive url for torch 1.10.0 + python 3.7

Hello, Thank you for your documents for using torch on Jetson target boards.
I need the torch==1.10.0 & python3.7 wheel on Jetson NX board, but I think there is only google drive url for torch==1.10.0 & python3.6.
My jetpack version is 4.4 and there is any way to use torch for my environment?