Segmentation Fault When Using Decord and TorchVision Together

[johannukrow]

First of all, I'd like to express my admiration for this project! It’s been incredibly helpful in my work, and I truly appreciate all the effort that has gone into building and maintaining it. However, I’ve encountered an issue that seems to indicate an incompatibility between decord and torchvision. Here’s a minimal reproducible example:

from decord import VideoReader
from torchvision import models

models.segmentation.fcn_resnet50().cuda()

Error Encountered:

Segmentation fault (core dumped)

Environment Details:

torch: 2.3.1
torchmetrics: 1.6.0
torchvision: 0.18.1
decord: 0.6.0
Python: 3.11.10

[cryonox]

maybe try swiching the import order. I found sometimes it work. like this, instead of importing decord first.

from torchvision import models
from decord import VideoReader

[ziyannchen]

Replicated the same error just using simple codes here

import os
import decord
import torch

device = torch.device('cuda')
mean_tensor = torch.tensor([[[[104.]], [[117.]], [[123.]]]], device=device)
print(mean_tensor)

Running the script got sgementation fault (core dumped) error. Just comment import decord or switch import torch before import decord, the error will be gone.