Torchvison compatibility

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin Nano
,
1

I was training a YOLOv8 object detection model on my jetson orin nano, and after the first epoch I got the following error message:

AttributeError: partially initialized module ‘torchvision’ has no attribute ‘extension’ (most likely due to a circular import)

I realized my torchvision version was 0.19.0 which is not correct for my torch version which is torch-2.3.0a0+ebedce2. I tried installing 0.18.0 through pip but that uninstalled my torch version and reinstalled torch 2.3.0 without CUDA enabled. Which torchvision version is compatible with torch-2.3.0a0+ebedce2 and how can I install it?

Additionally, I tried using the installation instructions for torchvision from this link (PyTorch for Jetson) but got the following error:

In file included from /home/anaygokhale/torchvision/torchvision/csrc/io/image/cpu/decode_jpeg.cpp:3:
/home/anaygokhale/torchvision/torchvision/csrc/io/image/cpu/exif.h:221:39: error: ‘png_structp’ was not declared in this scope
  221 | inline int fetch_png_exif_orientation(png_structp png_ptr, png_infop info_ptr) {
      |                                       ^~~~~~~~~~~
/home/anaygokhale/torchvision/torchvision/csrc/io/image/cpu/exif.h:221:60: error: ‘png_infop’ was not declared in this scope
  221 | inline int fetch_png_exif_orientation(png_structp png_ptr, png_infop info_ptr) {
      |                                                            ^~~~~~~~~
/home/anaygokhale/torchvision/torchvision/csrc/io/image/cpu/exif.h:221:78: error: expression list treated as compound expression in initializer [-fpermissive]
  221 | inline int fetch_png_exif_orientation(png_structp png_ptr, png_infop info_ptr) {
      |                                                                              ^
error: command '/usr/bin/aarch64-linux-gnu-gcc' failed with exit code 1

Please help me with this issue

2

Torch vision need to be built from source.
Check below link for more info.

3

@nagesh_accord But the “Build from Source” section only mentions torch, not torchvision. I tried the torchvision method that is listed under the “Installation” section and that gave me the error that I mentioned in my original post.

5

Hi,

The command for building TorchVision is at “Installation”.

Thanks.

6

@AastaLLL As I mentioned in both my previous messages, Those commands did not work for me. I have put the error message in my initial post. Please help me with that issue

7

Hi,

Which JetPack do you use?
For JetPack 6.0, please install the packages in the topics directly to ensure compatibility.

PyTorch v2.3.0

  • JetPack 6.0 (L4T R36.2 / R36.3) + CUDA 12.2
    ** torch 2.3 - torch-2.3.0-cp310-cp310-linux_aarch64.whl 1.5k
    ** torchaudio 2.3 - torchaudio-2.3.0+952ea74-cp310-cp310-linux_aarch64.whl 622
    ** torchvision 0.18 - torchvision-0.18.0a0+6043bc2-cp310-cp310-linux_aarch64.whl 939

Thanks.

8

Hello,
What would the correct command be for the installation of the torchvision and torchaudio using the above packages?
The torchvision installation instructions gave me an error as mentioned previously.

9

Hi

Please try below for the PyTorch:

And below for the TorchAudio and TorchVision:

Thanks.

11

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.