Segmentation fault when using torch together with decord

[xiaosu-zhu]

EDIT: The root cause of Segmentation fault is from torch, not cv2, please see the correction below.

I have installed opencv-python-headless and decord together. The version info is listed below.

opencv-python-headless 4.10.0.84
decord 0.6.0

I try to import two libs in a python script:

import cv2
import decord

OR 

import decord
import cv2

No matter which order I import two libs, both of them lead to Segmentation fault (core dumped) crash, without any other logs.

[JonathanLi19]

Same problem here

[xiaosu-zhu]

Correction

I have made a several test and found the core problem is from torch cuda. You could try several snippets below:

1. ❌

import torch
from decord import VideoReader

vr = VideoReader(video_path)
a = torch.tensor([3], device='cuda')

---
Segmentation fault (core dumped)

2. ❌

import torch
from decord import VideoReader

a = torch.tensor([3], device='cuda')
vr = VideoReader(video_path)

---
Segmentation fault (core dumped)

3. ✅

import torch
a = torch.tensor([3], device='cuda')

from decord import VideoReader

vr = VideoReader(video_path)

---
NORMAL

I found when creating a cuda tensor DIRECTLY AFTER torch import and then import the decord, all things will be fine.

If you import decord ahead and create any cuda tensor, a Segmentation fault will immediately throw out.